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.0004020 | The Effect of Deworming on Growth in One-Year-Old Children Living in a Soil-Transmitted Helminth-Endemic Area of Peru: A Randomized Controlled Trial | Appropriate health and nutrition interventions to prevent long-term adverse effects in children are necessary before two years of age. One such intervention may include population-based deworming, recommended as of 12 months of age by the World Health Organization in soil-transmitted helminth (STH)-endemic areas; however, the benefit of deworming has been understudied in early preschool-age children.
A randomized, double-blind, placebo-controlled trial was conducted to determine the effect of deworming (500 mg single-dose crushed mebendazole tablet) on growth in one-year-old children in Iquitos, Peru. Children were enrolled during their routine 12-month growth and development clinic visit and followed up at their 18 and 24-month visits. Children were randomly allocated to: Group 1: deworming at 12 months and placebo at 18 months; Group 2: placebo at 12 months and deworming at 18 months; Group 3: deworming at both 12 and 18 months; or Group 4: placebo at both 12 and 18 months (i.e. control group). The primary outcome was weight gain at the 24-month visit. An intention-to-treat approach was used. A total of 1760 children were enrolled between September 2011 and June 2012. Follow-up of 1563 children (88.8%) was completed by July 2013. STH infection was of low prevalence and predominantly light intensity in the study population. All groups gained between 1.93 and 2.05 kg on average over 12 months; the average difference in weight gain (kg) compared to placebo was: 0.05 (95% CI: -0.05, 0.17) in Group 1; -0.07 (95%CI: -0.17, 0.04) in Group 2; and 0.04 (95%CI: -0.06, 0.14) in Group 3. There was no statistically significant difference in weight gain in any of the deworming intervention groups compared to the control group.
Overall, with one year of follow-up, no effect of deworming on growth could be detected in this population of preschool-age children. Low baseline STH prevalence and intensity and/or access to deworming drugs outside of the trial may have diluted the potential effect of the intervention. Additional research is required to overcome these challenges and to contribute to strengthening the evidence base on deworming.
ClinicalTrials.gov (NCT01314937)
| The World Health Organization recommends starting population-based deworming interventions as of 12 months of age where intestinal worm infection is common; however, little is known about the benefits in early preschool-age children. We conducted a clinical trial to determine the effect of deworming on growth in one-year-old children in Peru. Participating children were randomly assigned to: 1) deworming at 12 months of age; 2) deworming at 18 months of age; 3) deworming at 12 and 18 months of age; or 4) no deworming (i.e. control group). A total of 1760 children were enrolled between September 2011 and June 2012, and followed up for one year. Overall, with one year of follow-up, no effect of deworming on growth could be detected in this population of preschool-age children. The potential benefit of the intervention may have been affected by low baseline infection prevalence and/or low compliance to the randomly assigned intervention. Additional research is required to overcome these challenges and to contribute to strengthening the evidence base on deworming.
| The soil-transmitted helminth (STH) disease cluster includes ascariasis, trichuriasis and hookworm disease. It is considered to be one of the most common Neglected Tropical Diseases (NTD), affecting an estimated 1.45 billion people worldwide [1]. STHs are transmitted in contaminated food, water and the environment in areas of poverty in low- and middle-income countries. These intestinal parasites have a direct and indirect adverse impact on nutritional status by disrupting normal nutrient intake, excretion and utilization in their hosts and by causing blood loss and loss of appetite [2,3].
WHO recommends large-scale preventive chemotherapy programs, using anthelminthic treatment (i.e. deworming), for the high-risk groups of women of reproductive age, especially pregnant women, school-age children (i.e. 5 to 14 years of age), and preschool-age children (i.e. 1 to 4 years of age) in STH-endemic areas [4,5]. Adverse effects from deworming are infrequent, and when reported, are mild and transitory, including gastrointestinal upset and diarrhea [6]. Deworming interventions are often school-based in order to reach school-age children. In preschool-age children, deworming is often piggybacked onto vaccination or supplementation programs, child health days, or programs for the elimination of lymphatic filariasis [7]. However, preschool-age children lag behind their school-age counterparts as scaling-up of school-based programs continues while that of preschool programs remains a challenge [7]. The global proportion of at-risk preschool-age children receiving deworming in 2012 was estimated to be on the order of 25% [7]. This coverage has decreased since previous reports [8].
Prior to 2002, children under two years of age had been excluded from deworming interventions as the burden of STH infection was perceived to be low in this age group and the safety profile of available anthelminthics was not well established. In 2002, WHO convened an informal consultation of experts, and subsequently recommended the inclusion of children between 12 and 24 months of age in deworming activities using single-dose albendazole (in a reduced dose of 200 mg) or mebendazole (in the usual dose of 500 mg) [9]. These recommendations were based on animal studies, toxicity data and other safety data [10]. Despite the WHO recommendations and increasing evidence of the occurrence of STH infection in early preschool-age children [10–15], many countries still exclude children under 24 months of age from their national deworming programs. Providing evidence on the potential benefits of deworming in the younger age group between one and two years of age is essential. A study reviewing data from 54 countries confirmed that preventive interventions must occur during the first two years of life to prevent growth deficits, such as stunting and underweight [16]. Interventions at this time are essential to prevent both short- and longer-term adverse health effects [17]. The evidence-base on including deworming as one of the essential early childhood interventions in this critical window is, however, limited. Randomized controlled trials conducted exclusively in school-age children or in both preschool-age and school-age children have provided mixed evidence on deworming benefits on growth and development [6,18,19]. Few studies have focused exclusively on the preschool-age population [12,20–22]. There is some evidence that adverse consequences of even low prevalence and intensity STH infection may be more pronounced in children during this critical time period [11].
Considering the unique nutritional demands and growth patterns of younger children, aggregated results from older children do not provide a clear indication of the potential benefit of deworming on growth and nutrition in younger age groups. To fill this research gap, we therefore conducted a randomized controlled trial on the effect, and optimal timing and frequency, of a deworming intervention incorporated into routine child health services at one year of age. Our objective was to determine whether deworming would improve growth by two years of age.
This study received ethics approval in Peru from the Comité Institucional de Ética of the Universidad Peruana Cayetano Heredia and the Instituto Nacional de Salud, in Lima, and the local Ministry of Health office (Dirección Regional de Salud (DIRESA) Loreto) in Iquitos (S1 Text). Ethics approval was obtained in Canada from the Research Ethics Board of the Research Institute of the McGill University Health Centre in Montréal, Québec (S1 Text). An independent Data Safety and Monitoring Committee (DSMC) was established with three members, from Canada, the U.S., and Peru, to review all adverse events and approve continuation of the trial at three time points. At baseline, eligibility was assessed, and an informed consent form was signed by both parents or guardians of the child. In the case of a single parent (e.g. due to death, separation or divorce), only one signature was required. The trial was registered with ClinicalTrials.gov (NCT01314937). The CONSORT checklist is described in S1 Checklist and the trial protocol is described in S2 Text.
We conducted a randomized, double-blind, parallel, placebo-controlled trial of a deworming intervention incorporated into routine growth and development (‘Crecimiento y Desarrollo” or CRED) visits in Iquitos, an STH-endemic area of the Peruvian Amazon. Details on baseline enrolment methodology and the study population have been described elsewhere [14]. Briefly, children were enrolled into the trial in their homes or participating health centres. Inclusion criteria were: 1) children attending any one of the 12 participating health centres for their 12-month CRED visit; and 2) children living in Belén, Iquitos, Punchana or San Juan districts. Exclusion criteria were: 1) children attending the health centre for suspected STH infection; 2) children who had received deworming treatment in the six months prior to the trial; 3) children whose families planned to move outside of the study area within the next 12 months; 4) children under 12 months of age or 14 months of age or older; and 5) children with any serious congenital or chronic medical condition. Any child who was excluded for medical reasons, and who was not already receiving regular health care, was referred to the health centre for follow-up by appropriate health personnel.
A baseline socio-demographic and epidemiological questionnaire (including family and child health and nutrition information) was administered in the home or health centre to the primary caregiver of the child. Baseline outcome measurements, including weight, length and the provision of a stool specimen, were ascertained in a subsequent visit in the health centre. All procedures were performed by dedicated, trained research assistants.
Following confirmation of eligibility, informed consent and all baseline outcome assessments in the health centres, children were randomized into one of four intervention groups:
Deworming consisted of a single-dose mebendazole tablet (500 mg) (manufactured by Janssen Pharmaceuticals Inc.; donated by INMED Peru). The placebo was identical to the deworming tablet in terms of size, colour and markings (manufactured and purchased from Laboratorios Hersil, Peru). Tablets were crushed and mixed with juice for ease of administration and safety [23]. The crushed tablet was administered by research assistants at the end of each visit after all outcome assessments had been completed. All children received deworming at the 24-month visit according to Peruvian Ministry of Health guidelines [24]. Children received usual care interventions and services from health centre personnel [24]. This included the administration of measles, mumps and rubella (MMR) vaccination at the 12-month visit, and diphtheria, pertussis and tetanus (DPT) vaccine booster at the 18-month visit.
Sample size calculations were based on detecting the smallest meaningful difference among intervention groups in mean weight gain over 12 months, and took into account potential effect dilution from treating infected and non-infected children. From previous research in the study area, STH prevalence was expected to be 25% at 12 months of age [13]. To estimate expected growth, longitudinal growth data was collected from health centre registries in the study area in 2011. Mean weight gain ± standard deviation between 12 and 24 months in 100 untreated children was calculated to be 2.0 kg ± 0.8 kg. The sample size was calculated a priori such that comparisons could be made between all four groups to look at the overall effect of deworming, as well as the effect of timing and frequency of deworming.
In order to have 80% power to detect a minimum difference of 0.20 kg in mean weight gain among intervention groups, assuming a common standard deviation of 0.8 and a significance level of 0.05, and using a one-way ANOVA which accounts for pair-wise multiple comparisons between all groups (i.e. 6 comparisons) using the Tukey correction, the estimated sample size per group was 366 children. The required sample size was increased to 440 children per group (1760 in total), to take into account potential loss-to-follow-up of 20% after 12 months (based on attrition rates from previous studies in the area by the research team [25,26]) (MC4G Software©, GP Brooks, Ohio University, 2008).
Computer-generated randomly ordered blocks of eight and twelve were used to randomly assign children to each intervention group in a 1:1:1:1 allocation ratio. Blocking ensured that the number of children assigned to each group would be balanced and reduced the potential for bias and confounding [27]. The random allocation sequence was generated by a biostatistician who was not otherwise involved in the trial. Research personnel not directly involved in the trial prepared small envelopes containing the randomly assigned intervention for each visit. These were numbered from 1 to 1760, with each number corresponding to one of the four intervention groups. Envelopes were stored in a temperature-regulated pharmacy at the research facility, and distributed by the Project Director (SAJ) or the local Study Coordinator (LP) in sequential order to research assistants until the sample size was achieved. Appropriate allocation concealment and randomly ordered block sizes ensured that the randomization sequence would not be predictable [27]. All health centre and research personnel, and parents of participants were blinded to intervention status.
Children were followed-up at their 18 and 24-month visit in the health centre, at which time all outcome ascertainments were repeated. At the 18-month visit the second randomly assigned intervention was administered. Each visit was scheduled six months after the previous visit. In the case that a participant did not attend their 18-month visit, children remained eligible for the 24-month visit, which was scheduled 12 months after initial enrolment. If participants were not located prior to the day of their anticipated follow-up visit, or a scheduled date was missed, a minimum of four additional attempts were made to locate them. The original end dates of the 18-month follow-up and 24-month follow-up (i.e. trial completion) were each extended by one month (i.e. seven months and 13 months after the end of enrolment, respectively) to maximize follow-up rates. A monetary reimbursement was provided to cover travel costs for each visit.
The pre-specified primary outcome measure was weight gain between the 12 and 24-month visit. Pre-specified secondary outcome measures were weight-for-age z-score, length gain, length-for-age z-score, change in STH infection prevalence and intensity, and change in development (i.e. cognitive, language and fine motor skills) between the 12 and 24-month visit. The development outcomes are reported separately.
Prior to commencing recruitment, in-depth practical training of the research assistants took place according to WHO guidelines [28,29] to ensure accurate outcome assessment and standardization. Inter and intra-rater reliability of over 95% was achieved for weight and length assessments, which are considered acceptable levels for anthropometric measurements [28,30].
Methods used for outcome measurements are described elsewhere [14]. Briefly, weight was measured using a portable electronic scale, accurate to the nearest 0.01 kg (Seca 334, Seca Corp., Baltimore, MD, USA). Length (i.e. the recommended measurement of height in children less than two years of age) was measured in duplicate as recumbent crown-heel length on a flat surface using a stadiometer (Seca 210, Seca Corp., Baltimore, MD, USA), accurate to the nearest millimetre. One stool specimen per child was collected to assess STH (e.g. Ascaris, Trichuris and hookworm) infection prevalence and intensity. For ethical reasons, only specimens from children receiving deworming treatment were immediately examined by trained laboratory technologists at the local research facility using the Kato-Katz method (single slide) for the presence and intensity (i.e. eggs per gram of feces) of STH infection [31]. At each time point, specimens from those children receiving placebo were stored at room temperature in 10% formalin and analyzed by the direct method for the presence of STH infection upon trial completion (Table 1).
This approach ensured that children found to be infected were treated. This approach also aimed to minimize effect dilution which would have occurred if treatment had been provided to those found to be STH positive, but randomized to receive placebo. The Kato-Katz method is the recommended technique for assessment of the prevalence and intensity of intestinal parasitic infection in fresh stool [31]. For a one-stool specimen, sensitivity and specificity are over 96% for Ascaris and over 91% for Trichuris [32]. There is lower sensitivity and specificity for hookworm; however, hookworm infection is generally uncommon in very young children in this study area [13]. Additional details on the collection of stool specimens, including the ethical rationale for using two methods of analysis and how blinding was maintained, are published elsewhere [14]. Lower sensitivity to detect STH infection from storage and later analysis of specimens by the direct method was also anticipated [14].
A socio-demographic and epidemiological questionnaire was administered at each visit. At the follow-up visits, this included a question on whether deworming had been received between study visits (i.e. outside of the trial).
Information on minor and severe adverse events was obtained through passive reporting at follow-up visits or in between visits. Severe adverse events were based on WHO definitions and included: 1) death; 2) life-threatening conditions; 3) in-patient hospitalization or prolongation of an existing hospitalization; 4) persistent or significant disability/incapacity; 5) cancer; or 6) overdose (accidental or intentional) [5]. All reported illnesses that did not meet the definition of a serious adverse event were considered to be minor adverse events. All adverse events were reported to ethics committees. Summary reports of adverse events were also provided to the DSMC.
Data collection activities during fieldwork were regularly supervised by the Project Director (SAJ) and local Project Coordinator (LP). The consistency of egg count assessments was evaluated among the laboratory technologists using standard quality control methods [31]. The laboratory supervisor read 10% of the slides of the laboratory technologists without prior knowledge of the result to ensure quality control.
Weight-for-age z scores (WAZ) and length-for-age z scores (LAZ) were calculated using WHO Anthro software (Version 3, 2011). WHO categories were used to classify STH intensity according to species-specific counts of eggs per gram of feces (epg) [33]. Both arithmetic and geometric mean epg were calculated.
The primary outcome of the trial was mean weight gain in kilograms (kg) between the baseline 12-month visit and the 24-month follow-up visit (i.e. after 12 months). Mean weight gain (kg) was compared between the four intervention groups using unadjusted one-way ANOVA procedure. Secondary analyses which were specified a priori were conducted to examine differences between intervention groups in terms of change in derived weight indices (i.e. mean WAZ change) and length and derived length indices (mean length gain and mean LAZ change). Multivariable linear regression was also conducted adjusting for age, sex, socioeconomic status (based on an asset-based proxy index) [34,35] and continued breastfeeding at 12 months of age.
All analyses were first expressed using an intention-to-treat (ITT) approach such that participants were analyzed according to their assigned intervention group. Multiple imputation, using a Markov Chain Monte Carlo (MCMC) model with five imputations, was used for those who did not attend the 24-month follow-up visit. Variables related to the outcome, and hypothesized to be related to missing the follow-up visit(s) were used to impute missing weight and length measurements. These variables were baseline weight, length, socioeconomic status, continued breastfeeding at 12 months, sex, and age. Imputation was done separately by randomly assigned treatment group. Additional analyses were specified a posteriori, including: 1) using a complete case approach on all participants who had attended the final follow-up visit, 2) using a per-protocol approach excluding those participants who did not attend all three visits and/or who reported having received deworming outside of the trial between baseline and the final follow-up visit and 3) restricted to children positive for STH infection at baseline. These analyses were conducted for the following reasons: 1) complete case analyses were conducted for comparison purposes with intention-to-treat analyses with imputed data; 2) per-protocol analyses were conducted to account for higher than anticipated non-compliance to the assigned intervention; and 3) subgroup analysis in STH-infected children were conducted to account for the lower than anticipated baseline STH infection prevalence.
The primary research question on the effect of deworming was determined by comparing growth outcomes between each intervention group and the control group. To explore the secondary research question on the effect of the timing of deworming (i.e. at the 12-month visit or at the 18-month visit), growth outcomes in Group 1 were compared to Group 2. To explore the secondary research question on the effect of the frequency of deworming (i.e. provided once or twice), growth outcomes in Group 1 and Group 2 were each compared to Group 3. All three research questions were specified a priori.
The effect of deworming on STH indicators at 24 months was also examined using a generalized linear model with a log link, a Poisson distribution, and a robust variance estimator to estimate the risk ratio for the dichotomous outcomes of any STH infection, Ascaris infection, Trichuris infection and hookworm infection, where no infection (i.e. no STH infection, no Ascaris infection, no Trichuris infection and no hookworm infection, respectively) comprised the reference group.
All statistical analyses were performed using the Statistical Analysis Systems statistical software package version 9.3 (SAS Institute, Cary, NC, USA).
Between September 2011 and June 2012, the parents of 2297 children were approached to participate in the trial. Five-hundred and thirty-seven children were excluded as they did not meet the inclusion criteria (n = 385), declined to participate (n = 126), or were approached but not enrolled once the sample size was reached (n = 26). A total of 1760 children were randomized to the four groups (Fig 1). All children received the assigned intervention at baseline. A total of 1606 children (91.2%) attended their first follow-up at the 18-month visit between March 2012 and January 2013. Due to parental refusal, three children did not receive their randomly allocated intervention. The average time between the baseline and first follow-up visit was 6.3 months (± 0.41) and between the first follow-up visit and the second follow-up visit was 6.3 months (± 0.47). The average time between the baseline and second follow-up visit was 12.6 months (± 0.67). Time between visits was equivalent among intervention groups. The second follow-up visit was completed between September 2012 and July 2013.
A total of 1517 children (86.2%) attended all three visits. Of those who did not attend all three visits, 108 (6.1%) attended the first visit only, 89 children (5.1%) attended the first and second visits and 46 children (2.6%) attended the first and last visits. The proportion of children reported to have received deworming outside of the trial during the study period was 25.7% in Group 1; 26.8% in Group 2; 26.3% in Group 3; and 30.3% in Group 4. These differences were not statistically significant (p = 0.49).
Baseline characteristics of the study population by intervention group are found in Table 2. Groups were similar in terms of baseline weight (kg) and length (cm), age (months), birth weight (kg) and length (cm), continued breastfeeding, up-to-date vaccinations and hospitalizations since birth. There were small differences in the proportion of girls in each group and vitamin A supplementation in the previous year. In terms of maternal and household characteristics, groups were similar in the proportion of mothers who were married or common-law, the level of maternal education, and access to potable water in the home. Small differences were found in maternal employment outside of the home and area of residence. Baseline characteristics were similar between children who attended the final follow-up visit and those who missed their final visit (S1 Table).
At baseline, the prevalence of any STH infection was 14.5% in the two groups whose specimens were analyzed by the Kato-Katz method (i.e. 13.6% in MBD/PBO and 15.2% in MBD/MBD) (Table 3). At the 18-month visit, any STH prevalence was 28.5% (i.e. 30.7% in PBO/MBD and 26.4% in MBD/MBD). As expected due to lower sensitivity, STH prevalence in children whose stool specimens were analyzed by the direct method at 12 and 18 months was moderately lower (i.e. 10.5% and 24.5%, respectively). Certain sensitivity analyses were therefore conducted in subgroups of children found to be STH-positive 1) by both the direct and Kato-Katz methods and 2) only by the Kato-Katz method. Despite potential misclassification of STH infection status in children whose specimens were analyzed by the direct method, this strategy allowed for maximum comparison among all groups. Infection was predominantly low intensity for Trichuris and hookworm infection at all three time points; however, moderate and heavy intensity Ascaris infection increased over the one-year follow-up period (Table 3).
At the 24-month visit, at which time all specimens were analyzed by the Kato-Katz method, the overall prevalence of any STH increased to 42.6%. Prevalence of Ascaris, Trichuris and any STH infection was moderately lower in the groups which received deworming at the 18-month visit. Hookworm infection remained negligible. No statistically significant difference in any STH prevalence or Ascaris or hookworm prevalence was observed in any of the deworming intervention groups compared to the control group; however, a statistically significantly lower prevalence of Trichuris infection was observed in Group 3, which received mebendazole at both the 12 and 18-month visits, compared to the control group (RR = 0.69; 95% CI: 0.52, 0.90) (S2 Table).
All groups gained between 1.93 and 2.05 kg in weight and between 9.61 and 9.84 cm in length, on average over 12 months. The greatest changes in all growth outcomes between the 12- and 24-month visits were seen in Group 1 (Table 4). The average difference in weight gain (kg) compared to placebo was: 0.05 (95% CI: -0.05, 0.17) in Group 1; -0.07 (95%CI: -0.17, 0.04) in Group 2; and 0.04 (95%CI: -0.06, 0.14) in Group 3. When comparing the outcomes in each of the deworming intervention groups to the control group, however, no statistically significant effect was detected in unadjusted or adjusted ITT analysis (Table 4). No statistically significant difference in any intervention group compared to the control group was seen in per-protocol analysis (S3 Table), complete case analysis (S4 Table) or in analysis restricted to only those children who were positive for STH infection at baseline (S5 Table).
In examining the effect of the timing at which deworming was administered, a statistically significant improvement was seen in Group 1 compared to Group 2, in terms of weight gain (unadjusted difference 0.12 kg; 95% CI: 0.01; 0.23), length gain (unadjusted difference 0.31 cm; 95% CI: 0.04, 0.58), WAZ change (unadjusted difference 0.13; 95% CI: 0.03, 0.23), and LAZ change (unadjusted difference: 0.12; 95% CI: 0.03, 0.21) between baseline and the final follow-up visit in unadjusted analyses (S6 Table). These results remained significant in adjusted analyses (S6 Table) per-protocol analysis (S7 Table), complete case analysis (S8 Table). In subgroup analyses restricted to children positive for STH infection at baseline, no significant differences were observed between groups (S9 Table).
In comparing the difference in anthropometric outcomes between Group 1, receiving deworming once yearly, and Group 3, receiving deworming twice yearly, no additional benefit on weight or length was apparent for twice-yearly deworming in unadjusted or adjusted analyses (S10 Table). Results remained consistent in per-protocol analysis (S11 Table), complete case analysis (S12 Table) and in restricted analyses to children infected with STH at baseline (S13 Table). A statistically significant benefit, however, was observed in Group 3 compared to Group 2, in terms of weight gain and WAZ change. These results remained significant for both weight gain and WAZ change when adjusting for baseline characteristics, in per-protocol and complete case analyses.
From baseline until the end of follow-up, 38 minor adverse events were reported and were similarly distributed among groups (i.e. Group 1: 7; Group 2: 10; Group 3: 12; and Group 4: 9). There were 18 serious adverse events reported: Group 1: 5 deaths and 2 hospitalizations; Group 2: 1 death and 0 hospitalizations; Group 3: 3 deaths and 2 hospitalizations; and Group 4: 2 deaths and 3 hospitalizations. Ten serious adverse events occurred after administration of mebendazole (i.e. 7 deaths and 3 hospitalizations) and eight serious adverse events occurred after administration of placebo (i.e. 4 deaths and 4 hospitalizations). The range of time between administration of the randomized intervention and occurrence of the serious adverse event was 6 days to 6 months for hospitalizations and 18 days and 7 months for deaths. None of these serious adverse events were deemed to be related to the deworming intervention by the DSMC, Research Ethics Committees in Canada and Peru, or the trial investigators.
This is the largest double-blind, randomized, placebo-controlled trial of deworming to our knowledge that has been conducted exclusively in children during the second year of life. This is the age at which WHO first recommends starting mass deworming programs, and it is also a time of rapid growth, development and STH acquisition. Our trial had several strengths. These include: 1) its randomized controlled design, which minimized confounding and the influence of external factors; 2) a large sample size, so that primary analyses were sufficiently powered; 3) a high follow-up rate, despite a highly mobile population and environmental challenges such as flooding which displaced many participants in the study area; 4) consistency of results in intention-to-treat, complete case and per-protocol analyses, demonstrating that results from children attending the final study visit are likely generalizable to the original trial population; and 5) community-based canvassing in the study area prior to recruitment to attempt to reach the entire study population [14].
We were not able to demonstrate an overall benefit on the primary research question of deworming on growth between any of the intervention groups compared to the control group after one year of follow-up in intention-to-treat analysis or in further sensitivity analyses. Our results are consistent with a recent cluster-randomized trial of albendazole (administered every six months to children from six months to six years of age) conducted in north India where light intensity STH infection was also predominant [21]. It is also consistent with a study in Uganda in children aged 15 months to 5 years who received quarterly albendazole [22]. Our findings do, however, contradict other trials in preschool-age children that found a positive effect of deworming on growth indicators [12,20].
The lack of benefit in our study compared to these other studies could reflect a true lack of effect of deworming on growth in the time period and/or study population. This population of children has a high level of malnutrition that may not be able to be treated solely by one or two doses of deworming in a one-year time period. In addition, we were not able to demonstrate a statistically significant reduction in any STH or species-specific prevalence with any of the deworming interventions, except for a reduction in Trichuris infection with twice-yearly deworming. The poor effect of the deworming intervention on STH prevalence measured after 6 and 12 months was almost certainly influenced by the dynamics of re-infection and new infection occurring between study time points. Future studies, which are adequately powered to detect changes in STH infection over time, are needed to confirm these findings.
If there were, however, a true effect that was not observed, the short follow-up time may have limited the potential to detect this benefit. It is likely that, in the one year period of our study, a steady state has not yet been achieved, in terms of either STH infection (e.g. as evidenced by the over threefold increase in STH prevalence from 12 to 24 months of age) or growth (e.g. as evidenced by a negative deviation of WAZ and LAZ compared to the international WHO growth standard over 12 months). Benefits of the deworming intervention may be apparent only with a longer follow-up time.
The low prevalence and intensity of infection, in particular, may have limited the impact of deworming, which reduces morbidity primarily through a reduction in moderate and heavy intensity infection. In deworming interventions, nutritional improvements are not a direct consequence of drug administration but a result of the elimination of parasites that are competing for nutrients. When the intervention is administered to a population with low prevalence and/or intensity, the short-term benefits could be difficult to measure. WHO recommends the periodic (once or twice-yearly) administration of antihelminthics as a means of controlling morbidity from STH infection. The nutritional benefits are a consequence of the maintenance of very a low level of these infections in childhood.
The baseline prevalence of STH infection in the study population was lower than had been anticipated based on a study conducted in the area just three to four years prior (i.e. in 2007 and 2008) [13]. The number of children who could have potentially benefited from deworming in the trial was therefore reduced, resulting in a reduction in power to detect an effect of the expected size. Results from a previous trial suggested that deworming could improve growth in young children, even with low intensity infection [12]; however, we did not observe this in our trial. Preventive chemotherapy programs include treatment of both infected and uninfected children; nonetheless, our results suggest that research studies should be conducted in areas of high STH prevalence to ensure as little effect dilution as possible. With increasing implementation of deworming programs, a rapid assessment in the age group and study area to determine baseline prevalence and intensity may be warranted before beginning any research study.
Our trial was unique in using a multiple group design to look additionally at the secondary research questions of differences in the timing and frequency among the groups that received deworming. Such considerations are important in operationalizing deworming interventions in this age group. Our results suggest that, if deworming is provided between one and two years of age, there is a significant benefit of providing it earlier rather than later. Our results also demonstrate that there was no added benefit from an additional dose provided at 18 months of age (over and above that at 12 months of age). These results were consistent in unadjusted and adjusted analysis, as well as in sensitivity analyses, for multiple growth indicators. A true benefit of earlier deworming compared to later deworming is biologically plausible, as suggested in nutritional research showing the importance of incorporating interventions as early as possible to prevent adverse health and nutritional consequences [16]. However, in light of the lack of benefit of any of the deworming interventions compared to the control group and the low STH prevalence and intensity at baseline, these results should be interpreted with caution. The difference in growth between Groups 1 and 2 may be due to a shorter follow-up from the time of intervention to the time of outcome measurement (i.e. 12 months in Group 1 vs. 6 months in Group 2), or a chance finding of lower average weight gain in Group 2 compared to all three other groups. Although the number of statistically significant findings was more than due to chance alone, and all comparisons, except for sensitivity analyses (i.e. complete case, per-protocol, and subgroup analyses in STH-infected individuals), were specified a priori, we cannot rule out the possibility that this difference could be a spurious finding due to chance alone.
One issue that arose after beginning the study was difficulty with compliance, as over 25% of children received deworming at least once outside of the assigned intervention group. A higher proportion of children in the control group had been reported to have received deworming outside of the trial, but this proportion was not statistically significantly different from the other groups. Even after taking non-compliance into account by conducting a per-protocol analysis, excluding those who took deworming outside of the trial and/or who did not attend all three study visits, no statistically significant difference in growth outcomes was observed. The ease of access to deworming was an unexpected result as deworming is not routinely provided to children under two years of age in the study area; however, this level of access to deworming outside of the study protocol has been observed in other trials in preschool-age children [22]. With the growing presence of deworming campaigns and availability of antihelminthics without a prescription in many countries, this finding of non-compliance will likely become increasingly common. Although for ethical and logistical purposes we could not restrict access to antihelminthics outside of the trial, it is imperative that compliance is measured in all deworming research studies and taken into account in the analysis of results.
For ethical and scientific reasons, we did not immediately analyze stool specimens from children randomized to placebo at the 12 or 18-month visits. This meant that accurate STH prevalence and intensity were not available for those receiving placebo (i.e. Group 2 (PBO/PBO) and Group 3 (MBD/MBD) at 12 months, and Group 1 (MBD/PBO) and Group 3 (MBD/MBD) at 18 months), and that results from the Kato-Katz and direct methods were not easily comparable (Table 1). Examining a single stool specimen with a single technique (i.e. Kato-Katz) may have also decreased the sensitivity to detect STH infection, particularly in those with low intensity infection [37]; however, considering the sample size and age group of children in the study, the collection of multiple specimens was not considered to be feasible. Our strategy had the advantage of providing accurate overall baseline STH prevalence of the study population (as all groups would be expected to have similar baseline prevalences due to randomization), and accurate final STH prevalences at the 24-month visit (at which time all groups were analyzed by the Kato-Katz method). As STH infection status was a secondary outcome (weight gain being the primary outcome), misclassification of infection status would not have affected the analyses on growth or on the effect of the deworming intervention on STH infection at the 24-month visit. Misclassification might have affected only the secondary subgroup analyses restricted to the STH-infected population. Despite the limitations to the strategy we employed, we consider ours to have methodological advantages to other strategies which have been used, such as: 1) not collecting stool specimens, which provides no information on baseline or follow-up STH infection[20], or; 2) analyzing all stool specimens immediately, and treating those found to be infected, regardless of allocated intervention group, which would dilute the effect size by providing treatment to those randomized to placebo if found to be STH-positive [22].
Overall, this is the first trial to provide evidence on the effect of deworming, including optimal timing and frequency, on growth exclusively in children in the critical time window between one and two years of age. This trial demonstrated the feasibility of incorporating deworming into routine growth and development health clinics along with other essential early childhood interventions. We were also able to demonstrate safety of the deworming intervention in this age group, with similar numbers of serious adverse events occurring after mebendazole and placebo administration. This is consistent with results from previous studies [22,23,38]. Continued observational follow-up of the trial cohort is currently taking place, and will provide evidence on the longer-term effects of deworming up to five years of age.
Future studies looking at the benefit of deworming on growth in this age group should include study areas of higher STH prevalence and/or intensity, higher potential compliance to the assigned intervention (i.e. lower availability of anthelminthics in the community) and longer follow-up time. Further studies should include other factors that are important to consider for scaling up deworming interventions in this age group. This includes: 1) cost-effectiveness of preventive chemotherapy vs. analyzing and treating only infected individuals; 2) feasibility and cost-effectiveness of integrating deworming with other health, nutritional and environmental interventions, particularly health education and micronutrient supplementation; 3) health and nutritional consequences of low intensity infection in younger age groups; and 4) inclusion of high-risk children living in more remote areas and/or those who do not regularly attend health services. This type of research is essential to contribute to strengthening the evidence base on deworming.
|
10.1371/journal.pntd.0004180 | Changes in the Proteome of Langat-Infected Ixodes scapularis ISE6 Cells: Metabolic Pathways Associated with Flavivirus Infection | Ticks (Family Ixodidae) transmit a variety of disease causing agents to humans and animals. The tick-borne flaviviruses (TBFs; family Flaviviridae) are a complex of viruses, many of which cause encephalitis and hemorrhagic fever, and represent global threats to human health and biosecurity. Pathogenesis has been well studied in human and animal disease models. Equivalent analyses of tick-flavivirus interactions are limited and represent an area of study that could reveal novel approaches for TBF control.
High resolution LC-MS/MS was used to analyze the proteome of Ixodes scapularis (Lyme disease tick) embryonic ISE6 cells following infection with Langat virus (LGTV) and identify proteins associated with viral infection and replication. Maximal LGTV infection of cells and determination of peak release of infectious virus, was observed at 36 hours post infection (hpi). Proteins were extracted from ISE6 cells treated with LGTV and non-infectious (UV inactivated) LGTV at 36 hpi and analyzed by mass spectrometry. The Omics Discovery Pipeline (ODP) identified thousands of MS peaks. Protein homology searches against the I. scapularis IscaW1 genome assembly identified a total of 486 proteins that were subsequently assigned to putative functional pathways using searches against the Kyoto Encyclopedia of Genes and Genomes (KEGG) database. 266 proteins were differentially expressed following LGTV infection relative to non-infected (mock) cells. Of these, 68 proteins exhibited increased expression and 198 proteins had decreased expression. The majority of the former were classified in the KEGG pathways: “translation”, “amino acid metabolism”, and “protein folding/sorting/degradation”. Finally, Trichostatin A and Oligomycin A increased and decreased LGTV replication in vitro in ISE6 cells, respectively.
Proteomic analyses revealed ISE6 proteins that were differentially expressed at the peak of LGTV replication. Proteins with increased expression following infection were associated with cellular metabolic pathways and glutaminolysis. In vitro assays using small molecules implicate malate dehydrogenase (MDH2), the citrate cycle, cellular acetylation, and electron transport chain processes in viral replication. Proteins were identified that may be required for TBF infection of ISE6 cells. These proteins are candidates for functional studies and targets for the development of transmission-blocking vaccines and drugs.
| High-throughput proteomics offers an approach to evaluate changes in cell protein levels following arboviral infection. Research to understand the molecular basis of human-flavivirus interactions has advanced significantly over the past decade, but comparatively little is known regarding interactions between ticks and tick-borne flaviviruses (TBFs). Here, we employed a proteomics approach using an I. scapularis ISE6 cell line infected with the TBF Langat virus (LGTV) to identify proteins and biochemical pathways affected by viral infection. An LC-MS/MS approach was used to identify proteins that were subsequently assigned to putative cellular pathways based on orthology to proteins in the KEGG database. Biochemical pathways common among arthropods in response to infection with flavivirus and possibly unique to tick-flavivirus interactions, were identified. In vitro cellular assays using small molecules suggest the involvement of the ISE6 proteins, malate dehydrogenase (MDH2), and mitochondria in viral replication. These analyses provide a basis for further studies to identify tick proteins associated with viral replication that could be targeted to disrupt TBF transmission.
| Tick-borne flaviviruses (TBFs; family Flaviviridae) are a complex of positive, single-stranded RNA viruses, many of which cause hemorrhagic fever and encephalitis in humans and are associated with high morbidity and mortality [1, 2]. Humans are incidental hosts for TBFs that are transmitted by an infected tick (subphylum Chelicerata, subclass Acari; superfamily Ixodida) during blood feeding. Tick-borne encephalitis virus (TBEV) is the most prevalent TBF worldwide and is responsible for over 10,000 confirmed cases of encephalitis globally per annum [3, 4]. Several TBFs associated with hemorrhagic disease are identified on the Centers for Disease Control and Prevention (CDC) “Select Biological Agents and Toxins” list (http://www.selectagents.gov/) due to their high virulence (biosafety level 3 and 4), anticipated ability to establish zoonotic transmission cycles, and their potential use in bioterrorism. Of these, Kyasanur Forest Disease virus (KFDV) is responsible for an estimated 400–500 human cases per year in India [5–7] while Omsk hemorrhagic fever virus (OHFV) is estimated to cause an average of 24 human cases per year (1946–2000) [8]. In the U.S., The increasing incidence of human cases of Powassan virus (POWV) and the corresponding genotype virus, Deer Tick virus (DTV) [9, 10] in the northeast and upper mid-west of the U.S., has refocused attention on TBFs in North America.
Langat virus (LGTV) was discovered in Southeast Asia in the 1950s [11]. LGTV exhibits low levels of virulence to humans, is classified as biosafety level 2 (BSL2) and employed routinely as a model for more virulent TBFs such as TBEV, KFDV, OHFV, and POWV/DTV. Other than for TBEV [5, 12], there are no vaccines or therapeutics available to prevent or treat infection with these virulent TBFs. Globally, there is an urgent need to identify novel prophylactics and therapeutics against TBFs.
The NIH-funded Ixodes scapularis (Lyme disease tick) Genome Project represents the first genome assembly for a tick and an important resource to understand the molecular processes in ticks [13]. The IscaW1.2 annotation comprises 20,450 gene models predicted via a combination of ab initio methods and manual curation. These models are a source of new targets [14] for the identification of novel chemistries [15] and vaccines [16–18] for control of ticks and tick-borne diseases.
Research has shown that proteins and metabolites produced by human [19, 20] and mosquito [21–24] cells (i.e., “host-cell factors”) may facilitate or play essential roles in flaviviral infection [25–28]. The mechanisms by which these molecules contribute to the pathogenesis of the Flaviviridae, are not well understood. Proteomics has been used to investigate interactions between ticks and bacterial pathogens [29–31]. Studies have also investigated global changes in the transcriptome of I. scapularis and in tick cells following LGTV infection [32], although there is little known about how these responses correlate to changes at the protein level. Tick proteins that facilitate viral infection and replication in the arthropod vector are logical targets for interventions aimed at disrupting transmission of TBF. Here we developed an in vitro assay using the I. scapularis ISE6 embryonic cell line [33–35] and LGTV (TP21 wildtype strain). We performed high-resolution LC-MS/MS analyses to evaluate global changes in the proteome of tick cells following flavivirus infection and identified proteins that displayed increased and decreased expression. We describe the cellular response to infection and employ small molecule functional assays to evaluate the involvement of several tick proteins in the infection and replication of LGTV in ISE6 cells.
Ixodes scapularis embryonic ISE6 cells (provided by T. Kurtti, University of Minnesota, Minneapolis, MN) were cultured at 34°C in L15B-300 medium in the absence of CO2 [36, 37]. Baby hamster kidney 15 (BHK15; ATCC cell provider) cells, used for plaque assay and immunofluorescent focus assay (IFA), were cultured at 37°C in Minimum Essential Medium (MEM) supplemented with L-glutamine, non-essential amino acids (NEAA), and 10% heat-inactivated fetal calf serum (FCS) with 5% CO2. Green African monkey kidney (Vero; ATCC cell provider) cells, used to create LGTV stock and for IFA to determine LGTV stock titer, were cultured at 37°C in MEM supplemented with L-glutamine, NEAA and 10% heat-inactivated FCS with 5% CO2. LGTV TP21 wildtype strain, passage 2 (obtained from A. Pletnev, NIH-NAID, Bethesda, MD [38]) stock was amplified in Vero cells (multiplicity of infection 0.01) [39] and grown as described above, except with 2.5% heat-inactivated FCS, up to passage 4 (p4) to provide a working stock for experimental infections. Serial IFAs were conducted in parallel as previously described [40] in 96-well cell culture plates to determine LGTV stock titers.
To create non-infectious LGTV (UV-LGTV), LGTV p4 stock medium was placed in 48 well cell culture plates and treated with UV radiation at a distance of 11 cm from a standard (12.4 watt) UV lamp in a biological safety cabinet (Nuaire Labgard ES, Plymouth, MN) for 30 second intervals over a five minute period. LGTV inactivation was confirmed by blind passage of UV-LGTV on ~2 x 107 ISE6 cells and ~80% confluent BHK15 cells, followed by immunofluorescent and plaque assay as described by Perera et al. [41] to demonstrate lack of infectivity.
IFAs were used to assess the level of LGTV infection in ISE6 cell populations. Detection of the LGTV non-structural protein 3 (NS3) was performed using YP-conjugated chicken anti-LGTV NS3 (provided by S. Best, NIH-NAID, Hamilton, MT) as primary antibody and IgG-conjugated goat anti-chicken, Alexa Fluor 488 (Invitrogen, Grand Island, NY; A11039) as secondary antibody. Cell nuclei were labeled with 4',6-diamidino-2-phenylindole (DAPI; Life Technologies, Grand Island, NY; D1306). Glass coverslips were used to culture and infect cells for the IFAs and were placed onto microscope slides, which were viewed on an Olympus model IX81F-3 microscope and images were collected using an Olympus U-CMAD3 camera. Fluorescence excitation was provided by the EXFO X-Cite Series 120PC and Olympus IX2-UCB. Image overlays were produced with Metamorph Basic v7.6.5.0 software.
To establish an MOI and time-point corresponding to optimal LGTV replication in ISE6 cells, three concentrations (MOIs of 7, 13, and 26) of LGTV were used to infect cells. For each, cells were fixed at 3, 9, 24, and 48 hpi with five technical replicates that were imaged under 20x magnification. On the basis of complete infection (>96%) of ISE6 cell populations between two MOIs (7 and 13) and time points (24 and 48), an MOI of 10 was selected for subsequent experiments for maximum infection. Separately, an assessment of the cumulative virus release was carried out in LGTV-infected ISE6 cells at a MOI of 10. Medium from these LGTV-infected ISE6 cells was harvested at 12 hour intervals for up to 120 hours, and subjected to plaque assays to measure replication.
Peptides with homology to I. scapularis, IscaW1.2 gene models were assigned to putative functional class by searching accession numbers against the KEGG orthology database (http://www.genome.jp/kegg/ko.html) and the KEGG pathway database (http://www.genome.jp/kegg/pathway.html). ISE6 proteins with orthology to KEGG entries were populated within KEGG pathways that also included mammalian and arthropod orthologs.
The concentration of cells in each sample (cells/ml) was estimated by counting cell number on a Scepter 2.0 Automated Cell Counter with 40 μM Scepter sensors (EMD Millipore; PHCC20040) in order to equalize cell numbers between biological replicates and between treatment groups prior to protein extraction. For cell population and growth analyses, initial cell counts (cells/mL) were determined manually using a hemocytometer and subsequently verified by sample analysis on the Scepter 2.0 Automated Cell Counter.
Trichostatin A (Sigma-Aldrich; T8552) and Oligomycin A (Sigma-Aldrich; 75351) were separately re-suspended in DMSO to a final concentration of 10 mM. 96-well plates, pre-treated with 0.01% Poly-L-Lysine (Sigma Aldrich; P4832), were separately seeded with ISE6 and Vero cells and incubated for 24 hours to final cell density of ~1 x105 cells/96 well. ISE6 and Vero cells were infected with LGTV (passage 4, MOI of 10) and (passage 4, MOI of 3), respectively. Following adsorption, compounds diluted in DMSO, were added to cells to a final concentration of 0.01, 0.1, 1, and 10 μM (1% of total overlay medium) and cells were incubated at 37°C. Culture supernatant was collected at 36 hpi and used to quantify LGTV replication by plaque assay. To assess cell viability, cells were treated with alamarBlue reagent (AbD Serotec; BUF012A) diluted 1:10 with fresh medium for 12 and 2 hours, respectively. Fluorescence (excitation at 560nm, emission at 590 nm) was measured at 48 and 38 hpi using a Molecular Devices SpectraMax M5 plate reader coupled with SoftMax Pro v4.8 software. Control was solvent only. Five technical replicates were performed for each concentration with biological replicates (n = 2).
Trypan blue cell exclusion assay was used to assess mortality of ISE6 cells following LGTV infection. Poly-L-Lysine-treated 96-well plates were seeded with ISE6 cells for 48 hours to a cell density of ~9 x 104 cells/well. Cells were treated with LGTV infection (MOI 10; p4 LGTV stock) or condition medium as described above. Cells were harvested at 12, 24, 36, and 48 hpi, centrifuged at 1,510 g for 5 min, medium was removed and the cell pellet was re-suspended in 1X PBS. Subsequently, a 1:1 0.4% trypan blue:cell suspension, was prepared, incubated for ~3 min at RT, the cells were immediately counted using a hemocytometer [54] and the percentage of stained ISE6 cells was determined for LGTV and mock treatments. Three technical replicates were collected per treatment with two biological replicates (n = 2).
IFA and plaque assays were used in time course experiments to assess levels of LGTV in ISE6 cells and to confirm UV inactivation of LGTV (Fig 1). Under the assay conditions described herein, IFA revealed that the maximum level of LGTV infection of the ISE6 cell population (>96%) corresponded to an MOI of 10 as determined by percentage of cells labeled with the LGTV NS3 protein (Fig 1A and 1B), and plaque assays revealed that the peak of LGTV release from ISE6 cells occurred at 36 hpi (Fig 1C). These conditions were selected for subsequent proteomic analyses. Plaque assays revealed that UV radiation for ≥120 sec was sufficient to achieve 100% inactivation of LGTV as determined by the lack of plaque formation (Fig 1D). The minimum time required for lack of plaque formation was 3.5 minutes. UV-LGTV used for proteomic analyses and subsequent assays was inactivated for five minutes.
ISE6 cell viability was reduced during the acute stage of infection with LGTV (i.e., ≤48 hpi) as measured based on presence of cellular reducing agents (FMNH2, FADH2, NADH, NADPH, and cytochromes). No change in cell growth or mortality was observed, as measured by counting cell population numbers and utilizing the trypan blue cell exclusion assay for LGTV-infected and mock-treated groups (S1 Fig).
Completion of the virus lifecycle as determined by release of infectious virus particles (Fig 1C) was observed in ISE6 cells infected with LGTV. In comparison, in cells treated with UV-inactivated virus (UV-LGTV) we observed no release of infectious virus particles (Fig 1D). Comparative proteomics analyses were used to identify proteins expressed throughout the process of cell infection (LGTV) versus those associated only with viral attachment and entry of the host cell (UV-LGTV). The sequence of proteomic analyses performed using the three treatments (LGTV, UV-LGTV, and mock) is shown in S2 Fig and S1 Table.
LC-MS data were compared for LGTV, UV-LGTV and mock samples (Fig 2). The expression pattern of LC-MS peaks for LGTV samples was more similar to that of UV-LGTV samples than to that of mock samples. The t-test and ANOVA (four separate statistical analyses) were used to identify proteins that exhibited differential expression (p < 0.05) between LGTV and UV-LGTV samples as compared to the mock samples (Fig 3A). In total, 486 ISE6 proteins (S2 Table) were identified based on homology to NCBI/VectorBase accessions. Of these, 266 and 248 proteins were identified as differentially expressed in the LGTV and UV-LGTV samples, respectively compared to mock samples. Sixty-eight proteins had increased expression, while 198 proteins showed decreased expression in the LGTV samples as compared to mock samples. Additionally, 82 and 166 proteins showed increased and decreased expression (Fig 3B), respectively in the UV-LGTV samples in comparison to mock samples. Overall, 243 proteins (50%) exhibited decreased expression while 120 (24.7%) showed increased expression in LGTV and UV-LGTV samples as compared to the mock treatment (Fig 4A).
Of the 486 ISE6 proteins identified in this study, 265 (54.5%) mapped to orthologous proteins in the KEGG database, while 221 proteins had no match (KEGG; genome.jp/keg/ko). Of the 265 proteins, 176 (36.2%) mapped to 66 KEGG pathways and 16 KEGG modules (S3A Fig and S3 Table). The KEGG pathways identified in the present study were categorized into five cellular functions: “metabolism”, “genetic information processing”, “environmental information processing”, “cellular processes”, and “organismal systems”. The majority of proteins (52%) were identified to the functional category “genetic information processing”, followed by “metabolic” (38.7%) and “cellular” (6.3%), “environmental information processing” (2%), and “organismal systems” (1%) (S3B Fig).
LGTV samples exhibited the highest number of proteins (53) identified to the KEGG pathway “genetic information processing” (Fig 4B–4D). Within this group, eight proteins exhibited increased expression and were classified in the pathway, “translation” (Fig 4B). For UV-LGTV samples, the majority of ISE6 proteins (57) were also classified in the pathway, “genetic information processing”. The majority of proteins exhibiting increased expression (17) were classified in the protein processing pathways of “folding, sorting, and degradation” (7 proteins; 41.2%), followed by “translation” (6 proteins; 35.3%) and “transcription” (4 proteins; 23.5%).
Proteins from the LGTV and UV-LGTV samples that lacked a match to KEGG database entries, also displayed differential expression. Of these, 30 proteins had increased expression and 91 had decreased expression in LGTV samples in comparison to mock samples (S4 Fig and S2 Table). Additionally, 38 and 85 proteins were identified with increased and decreased expression, respectively, in the UV-LGTV samples as compared to mock samples.
Proteins that showed an increase in expression in LGTV samples were mapped onto the KEGG functional categories of cell signaling (CYC, STK3, RPS6), proteolysis (UCHL3, PSMA, UBE2N), carbon-nitrogen hydrolase activity (DDAH, VNN), replication and mRNA processing (PARP, TRA2, CUTL, H2A, CSTF2), translation (RPS6, RPL17, AARS, NARS), glutamate metabolism/glutaminolysis (prostate-specific transglutaminase, putative ISCW011739; Fig 5 and S4 Table), pyruvate metabolism and energy association (MDH2; Fig 6). Proteins that exhibited decreased expression were associated with the functional categories of glycolysis (GAPDH; Fig 6), energy processes (ATP5H, ATP5A1), and mRNA surveillance (PABPC, PELO, MSI, THOC4).
Proteins exhibiting increased expression in UV-LGTV samples were mapped onto the KEGG functional categories of signaling (RHOGDI, RAB35, SIP, LAMC1), cytoskeletal components, (ACTN, TUBA), unfolded protein response and ER-associated degradation (HSPA1_8, RAB7A), lysosomal functions (PSAP), and phagosome functions (RAB7A). Proteins that exhibited a decrease in expression were associated with transport (BAP31), cell survival (BAP31, HYOU1, DERL1, GROEL), cell growth (SUMO, NOP10, MAD1L), translation (NOP10), and protein folding (GROEL).
Responses common to LGTV and UV-LGTV samples included proteins exhibiting increased expression and associated with signaling (ITGB, MO25), cytoskeletal structure perturbation (TLN), amino acid metabolism (ACAT, DP5CD, GLUD1, CARP, FAH), glutamate metabolism/glutaminolysis (DP5CD, GLUD1, membrane protein, putative ISCW001521; Fig 5 and S4 Table), RNA interference (AUB), and energy-production (ACAT). Proteins with decreased expression and common to both treatment groups were classified to KEGG functions of glycolysis (ALDOA/B/C, ALDH2/1B1/3A2; Fig 6), energy association (ATP5D, ATP5B), RNA interference (VIP), and structural manipulation (ACTB_G1, TUBB).
185 of the 265 ISE6 proteins with orthology to KEGG entries (70%) were also identified in a proteomics study of HCV infection of HUH7.5 cells [19] (S5 Fig). Sixteen ISE6 proteins (6%) matched orthologs identified in a study of West Nile virus (WNV) infection of Vero cells [55], 16 proteins matched orthologs in a yeast two-hybrid study of flavivirus-host interactions [56], and 15 proteins (5%) matched orthologs identified in Aedes aegypti infected with dengue virus (DENV) [28]. A subset of proteins that exhibited increased expression following LGTV infection and/or UV-LGTV treatment and matched proteins in the studies above, were associated with protein synthesis and proteolysis (Fig 7 and S5 Table). Of the remaining 66 proteins (24.9%), those that exhibited increased expression in LGTV samples were classified in the KEGG functional categories of proteolysis (PMSA, CARP), ATP association/interaction (PSMA, ANMK), cell and matrix adhesion (VNN, ITGB), and as well as oxidative stress and redox homeostasis (VNN and conserved hypothetical protein ISCW020127-PA). Additionally, the cellular function of hydrolase activity was suggested by increased expression of PSMA and VNN (S5 Table).
In order to manipulate metabolic functions and subsequent LGTV infection, small molecule assays were completed. In cellular assays, Trichostatin A (TSA), a compound known to inhibit histone deacetylase (HDAC) and to activate enzymes involved in intermediate metabolism, including MDH2, decreased viability of Vero cells (with and without LGTV infection) and LGTV replication (as measured by a decrease in release of infectious virus particles) at increasing concentrations (Fig 8A). Conversely, an increase in TSA concentration was associated with an increase in the viability of LGTV-infected ISE6 cells and an increase (~0.5 log pfu) in LGTV replication (Fig 8A). Oligomycin A (OligoA), a small molecule inhibitor of the mitochondrial H+ ATPase pump, known to inhibit terminal processes of the electron transport chain by reducing ATP production, was associated with a decrease in the viability of Vero cells (~20% reduction) and ISE6 cells (~60%) at increasing concentrations. Significant reduction of LGTV in the mammalian (~1.5 log reduction in pfu in Vero cells) and tick (~2 log reduction in pfu in ISE6 cells) system was observed with increasing concentrations of OligoA (Fig 8B).
We used an LC-MS/MS proteomics approach to analyze changes in the global protein expression profile of I. scapularis ISE6 cells following infection with LGTV and identified tick proteins tied to flavivirus infection and replication. The present study focused on proteins expressed during 36 hours post infection or the period of peak LGTV release from infected ISE6 cells (suggested 36–48 hpi in combination with published studies [57]). In total, 486 ISE6 proteins were identified, and of these, 66 exhibited increased expression and 198 proteins exhibited decreased expression following LGTV infection. Two hundred and sixty-five of the proteins identified (54.5%) had orthology to proteins of known function from a variety of eukaryotes (S1 and S3 Figs). Finally, we present in vitro small molecule data to demonstrate that metabolism in the mitochondria may be critical for tick-borne flavivirus infection.
Several proteins were identified were included in notch and mTOR signaling pathways. The putative histone deacetylase 1,2,3 (ISCW007830-PA) exhibited decreased expression in LGTV and UV-LGTV samples. Several studies [58, 59] suggest a link between herpesvirus infection and gene regulation through with the binding of viral proteins to histone deacetylases [59]. We hypothesize that LGTV infection may impact the regulation of ISE6 genes via effects on histone deacetylase. In other systems, it has been shown that histone deacetylase can act as a co-repressor in the notch signaling pathway. TSA traditionally binds and inhibits histone deacetylases and treatment of ISE6 cells with TSA during LGTV infection increased LGTV replication, suggesting that LGTV infection impacts gene regulation through histone deacetylases. The putative 40S ribosomal protein S6 (ISCW024315-PA) and Mo25 (ISCW004710-PA) exhibited increased expression in LGTV cells. These proteins are members of the mTOR signaling pathway which has been implicated in human cytomegalovirus (HCMV) infection of mammalian cells [60, 61] and DENV infection of A. aegypti mosquitoes [62]. Increased expression of Mo25 may reflect a cellular stress response while increased expression of S6 may reflect an increase in translation to maintain growth of the infected cell or facilitate LGTV replication. Manipulation of mTOR signaling has been noted with WNV infection in mammalian systems [63]. The putative calcyclin-binding protein CacyBP (ISCW013691-PA) known to function in the Wnt signaling pathway in other systems, had increased expression in UV-LGTV-treated cells and decreased expression in LGTV-infected ISE6 samples. Our observation suggests an increase in proteolysis following virus treatment since the Wnt pathway is associated with the Ca2+-dependent, ubiquitin-mediated proteolysis pathway. Future investigations regarding roles of post-translational modifications in regulating signaling pathways following tick-borne flavivirus infection is necessary.
Recently, the piwi-interacting RNA (piRNA) pathway has been implicated in the antiviral response of mosquitoes [64] and tick I. scapularis IDE8 cells [65]. Esther et al. 2014 identified three paralogs (ISCW015916, ISCW0021130, and ISCW011768) of the tick I. scapularis argonaute (aubergine) protein as antiviral factors to LGTV infection. The I. scapularis aubergine protein possesses the paz and piwi domains [66] associated with RNA binding. Homologs of these proteins were not identified in this study, although a homolog of argonaute (AUB; ISCW011373-PA) was identified that exhibited increased expression in both LGTV and UV-LGTV ISE6 cells and may play an antiviral role in LGTV infection.
Histone (H2A) is involved in DNA binding and chromatin packing of DNA, and therefore likely has a role in gene regulation and downstream host protein translation that may be important for homeostasis. The I. scapularis H2A (ISCW004478-PA) exhibited increased expression in LGTV-infected ISE6 cells. H2A also had increased expression during DENV infection of HUH7 liver cells and binds with the capsid protein to inhibit nucleosome formation in these human cells [67]. This protein has also been found to bind antisense RNA [68], also suggesting a possible anti-pathogen role as a result of changes in gene regulation.
The proteasome subunit alpha type protein (ISCW021572-PA) exhibited increased expression in LGTV samples and the 20S proteasome, regulatory subunit alpha type PSMA7/PRE6 (ISCW007139-PA) had increased expression in both LGTV and UV-LGTV samples. These proteins are subunits of the proteasome-associated 20S core particle and may exert antiviral roles through proteolysis and transcriptional regulation. Protein subunits of the proteasome have been shown to play a role in HCV internal ribosome entry site (IRES)-mediated translation [69] and may also interact with the HIV protein TAT and HBV protein HBX [70, 71].
Decreased expression of actin was observed in both LGTV-infected and UV-LGTV-treated samples. Cofilin (CFN; ISCW006326-PA), a putative actin-depolymerizing factor, exhibited decreased expression in these samples. CFN was also identified in a proteomic study of HCV-infected HUH7.5 cells [19]. Actin polymerization is involved with formation of actin stress fibers, a process that may facilitate vacuole formation [72] and mammalian neuronal cell entry of Japanese encephalitis virus [73]. UV-LGTV-treated cells exhibited increased expression of the signaling and structural proteins RHOGDI, and ACTN and TUBA, respectively. RHOGDI has been implicated in actin depolarization [74] and showed increased expression in HCV-infected HUH7.5 cells [19] at an early (12 hpi) infection time point. ACTN showed increased expression in HUH7.5 cells at early (24 hpi) and intermediate (48 hpi) time points post HCV infection and increased expression in UV-HCV-treated cells at a late (72 hr) time point. In the present study, we observed increased expression of ACTN in UV-LGTV samples at the 36 hpi time point. In addition to crosslinking actin fibers and facilitating filament assembly, ACTN been shown to bind the HCV nonstructural proteins NS3 and NS5 [56, 75]. We hypothesize that this protein may assist LGTV cell entry in ISE6 cells.
The proteins acetyl-CoA acetyltransferase (ACAT1; ISCW016117) and aldehyde dehydrogenase 4A1 (DP5CD; ISCW015982) exhibited increased expression in LGTV-infected cells. These enzymes operate upstream of the TCA cycle and are associated with the production of acetoacetyl-CoA and pyruvate, respectively during cellular metabolism (Fig 6). This result suggests an increase in acetyl-CoA production following viral infection. Interestingly, citrate synthase (CS; ISCW009586) showed decreased expression following LGTV infection and may reflect a reduction of TCA protein activity late in LGTV infection. We observed a decrease in expression of fumarate hydratase (FH; ISCW020593) that may also similarly reflect reduction of TCA protein activity late in LGTV infection. The increased expression of MDH2 (ISCW003528), a protein involved in the final steps of the TCA cycle, may produce an increase in oxaloacetate, S-malate, and NADH in ISE6 cells. Moreover, increased expression of fumarylacetoacetase (FAH; ISCW020196) may increase fumarate, also involved in the final steps of the TCA cycle. ACAT1, DP5CD, MDH2, and FAH may aid in maintaining the TCA cycle late in LGTV infection. In parallel, these observations suggest an impact of LGTV on the TCA cycle at 36 hours post infection that may be linked to successful replication of the virus.
Our observation of a decrease in expression of fructose-bisphosphate aldolase (ALDOA), glyceraldehyde-3-phosphate dehydrogenase (GADH), aldehyde dehydrogenase family 7 member A1 (ALDH3A2), and pyruvate kinase (PKLR) in LGTV-infected and UV-LGTV-treated cells, suggests an impact of LGTV on glycolytic processes. This finding is at odds with that of Patramool et al 2011, who observed that DENV-infected C6/36 A. albopictus cells [27] exhibit increased glycolysis. The in vivo study of Tchankouo-Nguetcheu et al 2010 highlighted an increased expression of glycolytic proteins in the midgut tissues of DENV-infected A. aegypti [28]. Diamond et al 2010 also identified members of the glycolysis pathway that exhibited increased expression at early to intermediate time points (i.e., prior to peak release of infectious virus) following HCV infection, but not at the late (during and following peak release of virus from the cell) time point [19]. Although ticks are exclusive blood feeders and mosquitoes regularly take sugar meals between blood meals, these data suggest the possible increase in glycolysis at early to intermediate time points post flaviviral infection, but a decrease in glycolysis at later time points.
Our in vitro studies have shown that the compounds TSA and OligoA can affect levels of LGTV replication, presumably through impacts on a variety of cellular metabolic processes. TSA is thought to inhibit histone deacetylases and stimulate the acetylation of histones and metabolic enzymes, while OligoA may inhibit oxidative phosphorylation and electron transport. OligoA may activate AMPK activity [76], inhibit ATP production, and affect cellular energy levels. Clearly, further studies are required to determine the mode of action of TSA and OligoA in the LGTV-ISE6 system.
Glutaminolysis can produce an alternative energy source for the cell by generating ATP during the conversion of glutamine to α-ketoglutarate. Although tick medium has relatively large amounts of glutamine, glutamic acid, and α-ketoglutarate, increased expression of proteins associated with glutaminolysis (Fig 5 and S4 Table) suggest that LGTV infection of ISE6 cells may stimulate glutaminolysis and the production of α-ketoglutarate, a key intermediate in the TCA cycle. Studies suggest that glutaminolysis is manipulated during infection of human cells by both HCMV [77, 78] and HCV [19]. Thus, glutaminolysis and α-ketoglutarate are likely critical not only for maintaining the TCA cycle, but also supporting oxidative phosphorylation and ATP production in the infected cell. Additionally, stimulation of α-ketoglutarate has been shown to increase mTOR activity [79, 80] which operates in parallel with glutaminolysis. In ongoing studies, we are assessing viral manipulation of glutamate dehydrogenase (GDH) activity using inhibitory compounds with the goal of disrupting flaviviral infection.
To contribute to an improved understanding of flavivirus-I.scapularis interactions, we developed an in vitro system to identify changes in ISE6 protein expression following infection with the TBF, LGTV. We present the first study to identify ISE6 proteins that are differentially-expressed following LGTV infection. In total, 486 proteins were identified with 66/198 showing increased/decreased expression following LGTV infection and 82/166 showing increased/decreased expression following UV-LGTV treatment. We identified proteins associated with the cellular functions of genetic information processing (GIP), metabolism, cellular processes, environmental information processing, and organismal systems. The majority of proteins populate GIP-specific pathways followed by metabolism-specific pathways.
The identifications of these proteins provide a critical resource to improve understanding of the I. scapularis proteome, improve gene annotations, and facilitate further studies in the tick cell culture system. Further understanding of protein function can also be achieved using approaches such as IFA, targeted mass spectrometry, small molecule in vitro assays, and RNAi. The present study is an important first step toward identifying tick proteins tied to LGTV replication as candidates for anti-tick vaccines and/or as targets for therapeutic screening to disrupt tick-borne flavivirus transmission.
|
10.1371/journal.pbio.2001402 | Digital Health: Tracking Physiomes and Activity Using Wearable Biosensors Reveals Useful Health-Related Information | A new wave of portable biosensors allows frequent measurement of health-related physiology. We investigated the use of these devices to monitor human physiological changes during various activities and their role in managing health and diagnosing and analyzing disease. By recording over 250,000 daily measurements for up to 43 individuals, we found personalized circadian differences in physiological parameters, replicating previous physiological findings. Interestingly, we found striking changes in particular environments, such as airline flights (decreased peripheral capillary oxygen saturation [SpO2] and increased radiation exposure). These events are associated with physiological macro-phenotypes such as fatigue, providing a strong association between reduced pressure/oxygen and fatigue on high-altitude flights. Importantly, we combined biosensor information with frequent medical measurements and made two important observations: First, wearable devices were useful in identification of early signs of Lyme disease and inflammatory responses; we used this information to develop a personalized, activity-based normalization framework to identify abnormal physiological signals from longitudinal data for facile disease detection. Second, wearables distinguish physiological differences between insulin-sensitive and -resistant individuals. Overall, these results indicate that portable biosensors provide useful information for monitoring personal activities and physiology and are likely to play an important role in managing health and enabling affordable health care access to groups traditionally limited by socioeconomic class or remote geography.
| A new wave of wearable sensors allows frequent and continuous measurements of body functions (physiology), including heart rate, skin temperature, blood oxygen levels, and physical activity. We investigated the ability of wearable sensors to follow physiological changes that occur over the course of a day, during illness and other activities. Data from these sensors revealed personalized differences in daily patterns of activities. Interestingly, we discovered striking changes in particular environments such as airline flights. Blood oxygen levels decreased during high-altitude flights, and this decrease was associated with fatigue. By combining sensor information with frequent medical measurements, we made two important health-related observations. First, wearable sensors were useful in identifying the onset of Lyme disease and inflammation. From this observation, we then developed a computational algorithm for personalized disease detection using such sensors. Second, we found that wearable sensors can reveal physiological differences between insulin-sensitive and insulin-resistant individuals, raising the possibility that these sensors could help detect risk for type 2 diabetes. Overall, these results indicate that the information provided by wearable sensors is physiologically meaningful and actionable. Wearable sensors are likely to play an important role in managing health.
| Physiological parameters such as heart rate (HR), blood pressure, and body temperature can provide critical information about the physical health status of a person. Elevation of any of these parameters can be of concern; elevated HR and blood pressure are associated with cardiovascular disease, and elevated body temperature occurs during pathogen infection and inflammation [1–4]. Peripheral capillary oxygen saturation (SpO2) is a measure of oxygen saturation of hemoglobin in the blood, and patients with chronic pulmonary disease often have lower resting SpO2 and are required to use supplementary oxygen to attain a more optimal SpO2 [5]. Skin temperature is associated with alertness levels and quality of sleep [6,7]. Although these different parameters are routinely measured in the physician’s office, they are not generally monitored outside of that context.
The infrequent collection of these measurements as currently practiced is problematic. First, changes in these parameters may not be identified until many months after an initial health condition has occurred. For instance, if a healthy person with reasonable health care access visits his or her physician every 2 y for a routine visit, then a condition may arise many months, or even longer, prior to a clinical symptom onset and thus go undetected for some time. Second, physiological parameters vary among individuals depending on their gender, life stage, and physical training, among other characteristics (e.g., [8,9]). These parameters also vary within the same person during their daily activities and with changes in the ambient environment. Because sparse clinical measurements of an individual are often compared to the average measurements of a population, the large variation within and among individuals results in a difficult medical assessment. Thus, infrequent short measurement periods or lack of adequate health care access makes it difficult to ascertain if a significant health change has occurred in a particular person. This information is particularly valuable for caregivers responsible for the health of others.
Emerging wearable biosensors (hereafter called “wearables”) are a low-cost technology that either continuously or frequently measures physiological parameters and provides a promising approach to routinely monitor personalized physiological measurements and potentially identify alterations in health conditions. Wearables are capable of passive and routine recording and immediate delivery of multiple types of measurements in real time to the wearer or physician with minimal attention or training required. In addition to physiological measurements such as HR and skin temperature, wearable technology has the potential to precisely capture the wearer’s daily physical activities, such as walking, biking, running, and other activities, often in conjunction with a GPS, which provides direct information about the location of the activity.
The popularity of wearable devices has substantially increased in recent years. As of July 2015, there are more than 500 different health care-related wearables present on the market and over 34.3 million devices sold. This is triple the number sold in 2013 [10].
Despite the revolution of wearable technology, studies to investigate their use in health care have been limited. One recent study using biosensors found no obvious benefit to users in health care costs or utilization [11]. In this work, we investigate the use of portable devices to (1) easily and accurately record physiological measurements in individuals in real time (or at high frequency), (2) quantify daily patterns and reveal interesting physiological responses to different circadian cycles and environmental conditions, (3) identify personalized baseline norms and differences among individuals, (4) detect differences in health states among individuals (e.g., people with diabetes versus people without diabetes), and (5) detect inflammatory responses and assist in medical diagnosis at the early phase of disease development, thereby potentially impacting medical care. In addition to a number of novel observations, through these analyses, we have gained considerable insight into the capabilities and value of these different devices in health and scientific research.
Our strategy was to intensely study one individual with many devices in order to determine the ease of collecting different types of data and to identify interesting patterns and then extend our analyses to a cohort of participants using a more limited number of devices (Fig 1A). We began by routinely measuring a 58-y-old male (Participant #1) over the course of 24 mo (Institutional Review Board [IRB] protocols IRB-23602, IRB-34907). From an extensive list of candidate devices, we selected seven that were easy to use, had reasonable accuracy, and had a direct interface for raw data access (see Material and Methods). The list of devices and their measurements is presented in Fig 1. These devices collectively measure (a) three physiological parameters, including HR, SpO2, and skin temperature, (b) six activity-related parameters, including sleep, steps, walking, biking, running, calories, and acceleration forces caused by movement, (c) weight, and (d) total gamma and X-ray radiation exposure. Collectively, these devices record more than 250,000 measurements each day (Fig 1). Many of the devices measure the same parameters, enabling cross-device comparison and assessment of measurement accuracy. During this period, Participant #1 also recorded the activities and travel in real time using web-based and smartphone-based software (see Material and Methods for details). Through numerous airline flights (S2A Table), an extensive analysis of the effects of air travel was assessed. Most information was collected through a smartphone, and all data were stored in a common database. Importantly, during the 2-y monitoring period, the participant was extensively monitored (73 visits) with standard medical tests, enabling a detailed comparison of medical data to the wearables information (S3 Table summarizes the medical tests performed in each examination).
In addition to a comprehensive analysis of a single individual, we also analyzed a larger group of participants to examine the consistency of our findings and explore differences and similarities across individuals (Fig 1A; IRB-23602, IRB-34907). Eighteen participants (ages 28 to 72; IRB-34907, whose enrollment criteria were all individuals ages 13 and older were eligible, see Materials and Methods) were analyzed for the effects of airline flight on SpO2 levels (IRB-34907; S1B Table). Our analysis of personal baseline and health also included physical monitoring of 43 individuals ages 35 to 70 y using a Basis device for up to 11 mo (average of 152 d; IRB-23602 and IRB-34907; individuals 18 or older were eligible, with a preference for those at risk for type 2 diabetes [T2D]; see Materials and Methods). The latter group is not T2D as defined by fasting plasma glucose <125 mg/dL and is free of chronic inflammatory conditions and major organ diseases. Four individuals, who self-reported as ill and had Basis Peak devices, were analyzed for the capacity of wearables to facilitate illness diagnosis and monitoring. Twenty individuals had quantification of systemic insulin resistance (IR) via the steady-state plasma glucose (SSPG) test as originally described and validated [12,13], and twelve of those individuals were classified as insulin resistant (S1A Table; SSPG greater than 140 mg/dL [14,15]). All together, over 1,788,538,186 measurements from 7,234 d were recorded.
Several of the devices, including the Basis device, used most frequently in our study had been validated for clinical-grade accuracy by the manufacturer (See Materials and Methods for details). Nonetheless, we performed extensive testing to assess the accuracy of the different devices against gold standard measurements and/or our instrument (Welch Allyn [WA] 6000 series), which is routinely used at the clinical laboratory services at Stanford University. We found that HR and SpO2 data collected using four devices (Scanadu Scout, iHealth-finger, Masimo, and Basis) were very close to that of the WA instrument over a wide range of values using the Bland–Altman method of comparison [16,17] and the Pearson correlation test (see S1 Fig). For example, HR measurements were within five beats per minute (BPM) and 10% of the WA instrument for all devices. SpO2 measurements were within 3% for all devices except for the Scanadu, which still yielded similar trends (see Material and Methods). Similarly we found that activity measurements were also close to standards for the conditions measured (e.g., MOVES App: steps: 0.79 +/- 0.16 standard deviation [SD] of the actual value; running: 0.96 +/- 0.05 SD of the actual value; details for all methods are presented in Material and Methods). Thus, we deemed the wearable biosensor measurements to be suitable for these studies.
In order to understand deviation from normal patterns, we first analyzed the collected data for systematic normal patterns, such as circadian rhythms, beginning with Participant #1. To reduce effects due to travel, our analyses focused on days lacking distance travel (defined as trips taken using airlines, assessed using GPS data from MOVES, and validated by comprehensive personal logs/calendars; see Material and Methods). Fig 2A–2D shows the circadian patterns of HR, skin temperature, and activity for 71 nontraveling d of Participant #1. As expected, we detected clear cyclical fluctuations over 24-h periods. For example, HR (measured using the Basis Peak) is generally lower at night (mean of 69.2 +/- 7.7 SD BPM from 10 p.m. to 6 a.m.) and higher during the day (mean of 84.5 +/- 11.3 SD BPM from 6 a.m. to 10 p.m.), with daily fluctuations (peak/trough or max/min) of 46.4 +/- 11.6 SD BPM (Fig 2B, S2A Fig), consistent with the sleep–wake cycle indicated by the Basis device (Fig 2A). Skin temperature measurements also generally followed a similar day-and-night pattern. Unlike that reported for core temperature [18], we found that skin temperature increases during sleep (a mean of 91.3 +/- 2.0°F for 10 p.m. to 6 a.m.; a mean of 86.6 +/- 3.2°F for 6 a.m. to 10 p.m., with daily fluctuations of 11.5 +/- 2.9 SD°F on average; Fig 2C, S2B Fig).
Comparison of physiological data with physical activity information revealed obvious activity-related physiological responses during specific time windows. Participant #1 often has an elevated HR during the 7 a.m.-to-8 a.m. and 6 p.m.-to-7 p.m. time windows, which included the typical time for bike commuting on weekdays (confirmed with daily calendar and consistent with MOVES information). On weekends, elevated HR was often evident in the 4 p.m.-to-6 p.m. window (S2 Fig), which is consistent with running activity measured using both the Basis device and MOVES. Overall, the correspondence of patterns detected with known activities indicates that the wearable devices can readily capture physiological information.
We also directly compared the physiological response in relation to different daily activities using data from the Basis device and MOVES apps (see Material and Methods). As shown in S3 Fig, our results replicate well-known patterns of physiological responses to events [19–22], including significantly faster HRs during exercise and significantly slower HRs during sleep compared to activity-free times, the mean of which is 78.4 ± 14.7 (SD) BPM; a mean of 67.6 ± 8.3 (SD) BPM, 101.1 ± 15.4 (SD) BPM, 114.1 ± 14.1 (SD) BPM, and 145.2 ± 18.1 (SD) BPM were observed during sleep, walking, cycling, and running, respectively (two-sided Wilcoxon rank sum p < 10−32). As expected, the measurements of HR, steps, calories, and skin temperature are very consistent for most of the activities, except the step measurement during cycling, which is not accurately detected using the Basis device (S3 Fig). Importantly, as described below, examination of recorded notes revealed a significant decrease in SpO2 levels measured by both the forehead and finger devices when Participant #1 reported fatigue (two-sided Wilcoxon rank-sum test p < 0.05; see below), and this finding was validated using systematic fatigue testing as described in the section on Airline Flights. Overall, these results indicate that our devices capture data as expected and also serve as a useful baseline to detect outlying measurements, as described below.
We further examined physiological parameters and activity patterns for 43 participants, including Participant #1, who wore a Basis device for between 1 and 24 mo. For the overall cohort, the resting HR was 72.10 ± 6.75 (SD) BPM and the resting skin temperature was 89.19 ± 1.88 (SD)°F. We found a significant difference in resting HR between men and women (Fig 3C): the resting HR of women was 73.70 BPM versus 68.80 BPM for men (p = 0.02248, Welch two-sample t test with 95% CI). These values are very similar to those reported by the NHANES study (74 BPM for women; 71 BPM for men) [23]. Women in our cohort also have a slightly higher average skin temperature (89.6°F) than men (88.7°F), but the value did not reach significance (p = 0.1724, Welch two-sample t test with 95% CI). In general, and as expected, we find that HR increases (S4A and S4B Fig) and skin temperature consistently decreases (S4C Fig) with increasing activity [21,22]. Furthermore, we see that the relationship between HR and skin temperature varies considerably among individuals (S4 Fig).
We analyzed the changes in HR and skin temperature between day and night for each of the 43 additional participants, although one individual (Participant #36) did not wear the Basis device while sleeping and was excluded from this analysis (Fig 2F). Average daytime HR (79.48 ± 6.96 BPM) was significantly higher than nighttime HR (66.99 ± 8.04 BPM; p = 4.836e-13, paired t test with 95% CI). The average difference between daytime and nighttime HR was 12.50 ± 7.80 BPM. Average daytime skin temperature (88.02 ± 2.02°F) was significantly lower than nighttime skin temperature (91.49 ± 1.77°F; p = 4.87e-13, paired t-test with 95% CI). The average difference between daytime and nighttime skin temperatures was 3.47 ± 2.17°F. This is consistent with findings that skin temperature increases by 7.2°F (winter) and 5.4°F (summer) during sleep [24].
As shown in Fig 2E, the differences in resting HR and skin temperature values differ widely among individuals. For resting HR, the values vary from 59.09 ± 6.59 to 84.97 ± 11.29 BPM. The range of values for skin temperature is smaller than that of HR, with values of 84.44 ± 3.85 to 93.65 ± 2.05°F, indicating tighter regulation of this physiological parameter. Although the Basis device has two sensors for detecting skin and ambient temperature, it is still possible that differences in skin temperature across individuals could be due to technical considerations in how the device is worn and/or exposed (S1C Table); however, the considerable differences in diurnal and nocturnal resting HR are likely to be due to personal differences between individuals because any measurement bias caused by how the device was worn is removed through differencing of daytime and nighttime values. Overall, we did not detect an obvious correlation between HR and skin temperature; this might be due to complications from the diverse activities of the individuals.
We also examined the activity patterns of the 43 individuals using the Basis device data. The individuals fell into four major groups, including individuals with highest activity in the early morning (Morning Active; Fig 2G, upper left panel), sustained activity during the day (All Day Active; Fig 2G, upper right panel), or peaks in activity either two times (Commuter Active; Fig 2G, lower left panel) or three times (Mealtimes Active; Fig 2G, lower right panel) daily. The peaks for the latter two categories fall between mealtimes (i.e., mid-morning, mid-afternoon, and, for one group, in the evening). We used this finding to train functional clustering machine learning algorithms to classify individuals by activity group (S2 Fig).
Because increased activity is associated with overall fitness levels, we examined the relationship between activity, resting HR, and weight loss. We observed that a higher average number of steps per day is associated with lower resting HR (adjusted R2 = 0.12, p = 0.01462; Fig 3A), and, when following changes over the course of 1 y, the increased average steps per minute is associated with a decrease in body mass index (BMI; adjusted R2 = 0.36, p = 0.003058; Fig 3B). In general, individuals with overall higher activity levels have less of a change in HR between high and low activity periods (S4B Fig), indicating increased fitness levels. Overall, these result indicate that there are highly varied baseline physiological differences as well as activity differences among individuals that relate directly to clinically relevant parameters, suggesting that individuals have personal physiome and activity patterns that can be tracked using wearable sensors.
From our detailed analysis of Participant #1, a striking and interesting change in physiological measurements was observed during airline flights, and, consequently, we pursued an in-depth analysis of physiological parameters during air travel. Cabin pressure in an aircraft is normally maintained at a reduced level with a minimum value comparable to that of 8,000 feet altitude [25], although several modern aircraft have been advertised to maintain higher cabin pressure. For Participant #1, we measured SpO2 and HR for 96 flights (summarized in S2A Table). The length of the flights varied from 23 min to 829 min, with 40 short flights (<2 h), 39 median-length flights (2–7 h), and 17 long flights (>7 h). Thirteen different aircraft models were included. The SpO2 level of Participant #1 was monitored by a forehead device (Scanadu) and/or finger monitoring devices (iHealth-finger, Masimo) in a continuous (Masimo) or discontinuous manner (Scanadu, iHealth-finger). The FlightAware website (https://flightaware.com/) was used to track specific details about the plane routing, altitude, and speed for each flight in real time.
We observed a striking decrease in SpO2 levels during airplane flights (a typical flight is shown in Fig 4A and S5A Fig), and this decrease is strongly negatively correlated with altitude. To summarize all flights, we binned each flight into five stages: before takeoff, ascent, cruise, descent, and after-landing stages (see Material and Methods). The overall distribution is shown in Fig 4B and S5B Fig for Scanadu and iHealth-finger measurements, respectively. Notably, Scanadu-measured SpO2 levels were at 97%–100%, 91%–96%, and 90% or less for 31.2%, 54.0%, and 14.8% of measurements, respectively, in the cruising stage as compared to 64.1%, 31.7%, and 4.3% in the stage prior to takeoff and 73.1%, 24.0%, and 2.9% in the after-landing stage (Fig 4B); iHealth-finger-measured SpO2 levels were at 97%–100%, 91%–96%, and 90% or less for 29.5%, 65.4%, and 5.1% measurements, respectively, in the cruising stage as compared to 88.0%, 10.8%, and 1.2% in the before takeoff stage and 80.6%, 13.9%, and 5.6% in the after-landing stage (S5B Fig). For the first 20 Masimo-measured flights, 19 had a significant inverse correlation of SpO2 levels with altitude (p < 4e-47 for each flight; the remaining flight had technical issues, see Material and Methods for details; see also S6A and S6B Fig for an aggregate SpO2 versus altitude of all flights, p < 1e-307). Thus, regardless of the device and flight, SpO2 levels correlate inversely with altitude. Seating locations were not found to have an effect on SpO2 levels. Overall, we observed a drop to a SpO2 of 96% or lower in all flights, and in many cases the drop was quite low (less than 94%) for a significant portion of the flight.
We hypothesized that the SpO2 reduction is most likely due to the reduced air pressure, leading to reduced available oxygen at cabin altitudes. To evaluate whether the reduced pressure is the primary cause versus other flight-related factors, we also measured SpO2 levels on a ~2 h automobile trip that climbed 993 meters and decreased 924 meters (altitude was determined by personal logs and the DraftLogic website https://www.daftlogic.com/sandbox-google-maps-find-altitude.htm; Fig 4C). SpO2 levels tightly correlated with altitude, indicating a direct relationship between SpO2 level and air pressure/oxygen [26].
Interestingly, we observed that on long flights the SpO2 levels are higher toward the end of the flight than those at the beginning (a typical flight in shown in Fig 4G; data for a number of long flights are shown in S6 Fig). Among the 17 Masimo continuous recorded flights that have records for the last quarter of the flight, we observed that on long flights greater than 7 h, SpO2 levels in the last quarter are significantly higher than at least one of the other three quarters measured at the same altitude level (Wilcoxon ranksum test p < 1x10-29). and this observation was not detected on the short flights (Fig 4H). Furthermore, if we binned the Masimo-recorded, high-altitude (>35,000 feet) SpO2 levels into two categories, (1) measured after 7 h from departure time and (2) measured earlier than 2 h from departure time, we observed significantly higher SpO2 levels in the former category compared to the latter one. This increase is likely due to either adaptation or a physiological change after rest/inactivity. For the long westward flights, Basis-measured activity was relatively constant (i.e., primarily sitting >95% of the flight) and with little or no sleep (Basis-quantified and self-reported); an example is shown in Fig 4G, and the SpO2 increase at the end of flight was always observed (see Fig 4G–4I and S6 Fig; Scanadu measurements are in S5F Fig), indicating that this increase is most likely due to adaptation. This observation demonstrates that humans can adapt to low oxygen after a number of hours on an aircraft.
To determine how SpO2 measurements relate to macrophenotypes, we also examined physiological measurements during periods when the participant logged their alertness status as either “tired” or “alert” using a blind scoring system (its calibration has been quantified as described in Material and Methods). As shown in Fig 4E (bottom panel), the SpO2 level reported when “tired” on flights was significantly lower compared to that measured when “alert” (two sample Kolmogorov–Smirnov test p < 3x10-8), similar to that reported on nonflying days (Fig 4E upper panel, two-sample Kolmogorov–Smirnov test p < 5x10-6). This observation was evident by both finger (Fig 4E) and forehead devices (S5C Fig). To be more quantitative in reporting fatigue, the participant also performed a psychomotor vigilance test and quantified fatigue by measuring the speed with which participants respond to a visual stimulus (see Methods and Material for details). As shown in Fig 4F and S5D Fig, a longer response time was required when the participant logged their state as “tired” rather than “alert,” and the SpO2 level is strongly negatively correlated with the response time (Pearson correlation test R = -0.88 to -0.91, p < 6x10-5). This result not only validates the self-reported system of fatigue but also provides a quantitative summary of the relationship between fatigue and SpO2 level. Although reductions in SpO2 levels during flights have been reported previously [27–32], to our knowledge, this the first report of (a) adaptation on long flights and (b) fatigue levels on actual commercial aircraft with objective assessment of fatigue (see Discussion).
To further examine whether oxygen levels decreased in other individuals during airplane flights, we measured SpO2 levels in 17 other individuals from diverse ethnic backgrounds (European, Jewish, African American, Indian Asian, and East Asian) using the iHealth-finger device or Masimo device (details in Material and Methods, S1B and S2B Tables). In every case, decreased SpO2 levels were observed during cruising (difference between median SpO2 varies from 2% to 9%, two-sided Wilcoxon rank sum test p < 0.001). We also found that baseline SpO2 and decrease in SpO2 during cruising varies with different individuals (Fig 4D). A plot of altitude versus SpO2 reveals that, in general, SpO2 decreases are lower at lower-altitudes flights, and this is particularly evident for individuals with more than four flights (S6 Fig). Overall, these results indicate that reduced SpO2 levels during air travel are a general phenomenon and occur in all types of aircraft. This result is consistent with published observations [27–32] and further indicates that the decreased SpO2 during air travel is evident across different ethnic groups.
To investigate the capacity of wearables to facilitate disease diagnosis and monitoring, we examined the association between unusual physiological signals and disease status or disease markers. This was uniquely possible for our study because we had frequently sampled and performed a number of biomedical assays during the entire monitoring period (see S3 Table for the list of tests).
We began our analysis by focusing on Participant #1, who was measured continuously for HR and skin temperature and frequently for SpO2 levels for a period of 679 d (measurements were recorded for 603 d; see Material and Methods) and had a very large number of days (73) with extensive clinical testing during this period. As indicated above, physiological parameters change dynamically with daily activities (e.g., significantly faster HRs during exercise [S3 and S4A Figs] and significantly slower HRs during sleep [S3 Fig, Fig 2B and 2F]). Therefore, we compared each parameter according to the corresponding activity information (see Material and Methods for details) and chose those periods lacking physical activity to calculate the percentage-of-outliers (i.e., percentage of reads per day classified as outliers from the overall personal mean) to search for periods with significantly different measurements. During the 603 d of monitoring, we identified 8 d with abnormal HR and skin temperature pattern (Fig 5A; S7A Fig). Interestingly, most of these days fell into four periods that are of very high interest from a health perspective.
(1) The most significant period was a 5-d period (Days 470–474), during which we detected abnormally elevated HR (~14% to ~55% reads per day were defined as significant outliers compared to the corresponding baseline norm) and skin temperature (~5% to ~19% of the reads per day were defined as significant outliers compared to the corresponding baseline norm) during each of these days (Fig 5A and 5B). During this period, Participant #1 was suffering from Lyme disease (as diagnosed on Day 487 by a positive antibody test, S7D Fig). Lyme disease is a Borrelia bacterial infection primarily transmitted to humans through tick bites; 12 d prior to this period (on Day 458), the participant had been exposed to an area in rural Massachusetts where high levels of Lyme-infected ticks are present. (Note that a “bull’s eye” rash was not observed during the initial period of infection). Importantly, the participant first noticed a possible health concern at the onset of the elevated HR/skin temperature period (on Day 470) and by abnormally low SpO2 readings both during an airline flight and afterwards (Fig 5C). This elevated condition was followed over the next 4 d by a persistent elevated temperature reading with an oral thermometer (98.9–102°F; see Material and Methods for details). Importantly, the normalized HR and skin temperature from the wearable device were significantly elevated multiple times during this period, and the HR exhibited a strong correlation with measurements taken with the oral thermometer over this period (R = 0.81, p <0.05, S7B Fig). The participant visited a physician at Day 474 and received treatment with doxycycline; the symptoms and abnormal vital signs disappeared the following day (Fig 5B).
(2) Another interesting period of outlying HR and skin temperature was Day 518 (Fig 5A, S7A Fig). About 20% of the HR reads and about 5% of the skin temperature reads on that day were defined as outliers compared to the corresponding baseline norm. Importantly, an elevated high-sensitivity C-reactive protein (hs-CRP) level was detected that day by a blood test (hs-CRP: 24.8 mg/L; baseline is normally <0.2 mg/L), although clinical symptoms (i.e., fever) were not reported. The elevated CRP, HR, and skin temperature were identified during data analysis.
(3) Another interesting period is from Day 455 to Day 456 (Fig 5A, S7A Fig). About 16% of the HR reads on Day 455 and about 8% of the skin temperature reads on Day 456 were classified as outlier compared to the corresponding baseline norm. Importantly, Participant #1 was diagnosed with a Human rhinovirus infection during this period (GeneMark test) and had elevated hs-CRP levels and inflammatory cell counts (hs-CRP: 16.2 mg/L, white blood cell: 10.8 K/ul, Neutrophil: 8.4 K/ul, 77.9%).
(4) Another interesting period is from Day 665 to Day 669 (Fig 5A). About 26% of the HR reads on Day 667 and about 2% of skin temperature reads were defined as outliers compared to the corresponding baseline norm. Importantly, an elevated hs-CRP level was detected on Days 665 and 669 by a blood test (hs-CRP: 4.3 mg/L and 15.4 mg/L on Days 665 and 669, respectively). The participant reported congestion during this period.
A plot of outlying HR and skin temperature as a function of time during each of these periods further demarcates their deviation from baseline and illustrates that the Lyme disease period has the highest outlying measurements (S8A Fig).
In summary, each of the circumstances with both elevated outlying HR and skin temperature was associated with elevated hs-CRP, indicative of a high inflammatory response (Fig 5A, S7 Fig), and for three of the four periods, clinical symptoms were reported. The participant did not report any other illness during this period of monitoring. These data indicate that there is a strong correlation between inflammatory response and elevated HR and skin temperature, which can be detected by wearables (Pearson correlation coefficient for CRP and the fraction-of-outlying-heart-rate (R = 0.96, p < 10e-28); coefficient for CRP and fraction-of-outlying-skin-temperatures (R = 0.94, p < 10e-24). For the case of Lyme disease, the abnormal physiological measurements of SpO2 and HR were important in alerting the participant to the disease.
To determine whether disease-associated events might be detected in other individuals using wearables, we identified three other individuals in our cohort who self-reported as ill and had Basis devices (but not SpO2 measurement devices; Fig 5D and 5E, S7E–S7G Fig). One individual had been ill twice. In each of these four instances, high CRP levels and elevated HRs were evident relative to their personal backgrounds (between 2.02 and 4.66 SDs above background). Although one of the individuals also had elevated skin temperature during this period (S7E Fig), interestingly, for two of the individuals (three illnesses), we did not detect elevated skin temperature. This might relate to differences in how the device was worn, as our survey results demonstrate that the device was worn loosely for at least one of these two individuals. As with Participant #1, for these three individuals, we also searched for other periods of elevated resting HR. For Participant #37, the illness period was the strongest outlier by a very large amount (4.66 SDs above background; rank #1 out of 25 d of monitoring for fraction of HR outlying measurements, S7F and S7G Fig). For Participant #58, the two illness days were in the top 5% of elevated HR outliers (3.40 and 2.02 SDs above background; ranks #10 and #19 out of 568 d of monitoring, Fig 5D), but we do not have the corresponding CRP levels to the other dates with outlying HR to know if those dates represent periods of illness/inflammation. For Participant #59, elevated HR occurred between 48–72 h prior to reported symptoms (3.55 SD above background; rank #1 out of 138 days of monitoring, Fig 5E) and elevated skin temperature on the day of and 48 h prior to reported symptoms (2.15 and 2.45 SDs above background, ranks #4 and #2 of 138 d of monitoring, respectively; S7E Fig).
In summary, we observed elevated HR during each ill period for all four individuals (eight total events), which suggests that monitoring of HR (and sometimes skin temperature) using a wearable device can detect inflammatory periods.
To examine the resolution at which illness might be confidently identified, we developed a computational approach called “Change-of-Heart” or COH to identify periods with abnormal HR patterns. HR was chosen because, as described above, it reliably detected all periods with elevated CRP levels in each of the individuals. We were unable to reliably map elevated skin temperature at high resolution during these periods across all individuals, and thus this parameter was not pursued. Specifically, we focused on deviations in resting HRs relative to an inactive period and applied a peak-finding–based algorithm to the smoothed continuous HR signal to search for peaks different from a global and local distribution (see Material and Methods). This peak-finding method is optimal for identifying times of transition from healthy to ill states, and thus preferentially detects early periods of infection, which is most desirable.
As shown in Fig 5F, during the 679 d when Participant #1 was monitored, we identified 11 periods with elevated HR. These periods successfully tagged all of the four sick periods indicated above, sometimes with multiple peaks, and also revealed four other periods during which no illness was reported. Application of this approach to the other three individuals also revealed peaks during each of their ill periods. For all four individuals, we are able to identify all of the sick periods using this method with area under the receiver operating characteristic curves larger than 0.9 for each individual (S8B Fig). Importantly, each illness period is identified (100% sensitivity), and for most of the sick periods, significant signals were evident at the very beginning of the illness period. Overall, these results indicate that elevated HRs are present during illness and can be detected using wearable devices.
The availability of clinical measurements on our participants enabled us to investigate associations between information collected from wearables with clinically important data. We focused on diabetes-related measurements because many of our participants were at risk for T2D. Diabetes is a significant rising global health problem, and IR is highly correlated with progression to T2D [33]. Twenty individuals in our cohort underwent measurement of their SSPG, a direct measurement of resistance to insulin-mediated glucose uptake (See Material and Methods) [12,13].
We performed a stepwise modeling approach to examine the relationships between SSPG values and HR, activity, and BMI, beginning with a simple univariate model and then building to bi- and trivariate models. We first examined the associations between daytime, nighttime, and delta (daytime minus nighttime) HR and SSPG (Fig 6A, 6C and 6D) because of evidence that diabetes is associated with changes in diurnal variation of HR [34]. Both daytime HR (Fig 6C) and delta HR (Fig 6A) were positively correlated with SSPG (Daytime HR: β = 4.5, 95% CI 1.2–7.8), p = 0.0107; Delta HR: β = 4.1 (95% CI 1.1–7.1), p = 0.0098), but nighttime HR (Fig 6D) was not.
Because our previous results showed a relationship between overall activity and resting HR (Fig 3), we wanted to evaluate whether the relationship we discovered between daytime or delta HR and SSPG was due to differences in study participant activity. We first assessed whether there was a relationship between daily activity and SSPG (Fig 6B) and found that average daily steps had an inverse relationship (β = -0.012, 95% CI -0.022–-0.002, p = 0.0183) with SSPG. We also evaluated the relationship between average daily steps and HR and found that daily steps was not significantly associated with daytime HR (β = -0.0008, 95% CI -0.0021–0.0005, p = 0.1943) but did have a significant inverse relationship with nighttime HR (β = -0.0017, 95% CI -0.0030–-0.0004, p = 0.0115). Thus, the association of higher daytime HR with higher SSPG levels is unlikely to be due to differences in participant daily activity. Including overall activity in a multivariate regression model with delta HR to predict SSPG resulted in an improved adjusted R2 to 0.41 from 0.17 in the univariate model with delta HR as the only predictor. These results suggest that information from different wearable sensor data types in combination can improve the ability to detect important physiological changes as compared to information from a single sensor.
To assess whether BMI plays a role in the relationship between delta HR and SSPG, we further expanded our multivariate regression to include BMI. BMI is known to have a positive correlation with HR [35], and IR and is negatively correlated with overall activity levels (Kruger et al., 2016). In this model, delta HR remained a strong predictor of SSPG levels (β = 5.05, 95% CI 2.73–7.37, p = 0.0003) independent of daily activity (β = -0.010, 95% CI -0.021–0.000, p = 0.0509) and BMI (β = 7.58, 95% CI 1.83–13.33, p = 0.0130, adjusted R2 = 0.52). Thus, combining information from multiple wearable sensors and electronic medical records to capture the relevant underlying physiological parameters enables enhanced prediction of SSPG. Overall, these results indicate that individuals with different degrees of IR and insulin sensitivity have important physiological differences and that these differences can be measured using wearable devices.
Lastly, to examine the diversity of measurements that can be quantified, we also explored whether individuals encounter periods of radiation exposure using a personal radiation tracker (RadTarge II D700) to monitor the local environmental radiation level over a 6-mo period. Fig 7 displays the distribution of radiation exposure over a 25-d period. As suggested by these data, Participant #1 typically lives in an environment with a low background radiation level around 0.003+/-0.0006 millirem (mRem) per h; however, several exposures of elevated radiation levels occurred. The majority (>90%) of the events over 0.030 mRem per h occurred during airplane flights, consistent with the expectation of increased exposure to cosmic radiation at high altitudes [36–38]. As shown in the enlarged panel, the radiation level per h generally corresponds closely with the interval and altitude of the airplane flight (typically rising to 0.038+/-0.004 mRem/h for a 35,000–39,000 feet altitude flight), an increase of 12.7-fold over home background levels). Short flights with lower cruising altitude (a flight on Day 642 with a maximum cruising altitude of 27,000 feet and the one on Day 652 with a maximum cruising altitude of 26,000 feet) yield only modest increases in radiation exposure over the background. The increased exposure at high altitudes is consistent with the well-known fact that a radiation-protective layer of atmosphere surrounds the Earth and is diminished at higher altitudes [39].
Several other interesting periods of increased radiation levels were evident. A very modest increase was evident for an entire 8-h period at the Scripps Research Institute (3-fold) and in an underground museum in Los Angeles (also 3-fold). A very substantial increase (400-fold at peak levels) in higher radiation was detected when the participant entered a hospital café. This increased exposure lasted the entire 10–15-min period in the café but was not evident upon return to the same location two times later in the same day. A likely explanation for this observation is that the radiation source was present on someone in the café, most likely an individual undergoing internal radiation therapy. We did not detect differences in HR or skin temperature during this period. Overall, these data show that very modest increases in radiation are present in several locations in the country with more substantial increases during airline flights or during chance encounters with individuals or locations with high radiation. Most importantly, it demonstrates that simple personal wearable devices can identify these levels and provide immediate feedback.
The data presented above indicate that many types of continuous physiological and activity information can be collected on a single individual on a long-term basis and can be used to measure, analyze, and guide health-related decisions. We showed that wearables can capture expected observations such as circadian fluctuation in HR and skin temperature and their changes during activity. In addition to serving as a valuable resource, we also found several interesting and important new results.
First, we found a decrease in SpO2 levels on airplane flights, including the frequent intervals (14.8%) with very low SpO2, with an adaptation toward more normal levels on long (>7 h) flights. The former has been reported previously [27–32], but, to our knowledge, the SpO2 decrease on modern aircraft (including Boeing 787) and the interesting discovery of SpO2 adaptation after approximately 7 h on long flights have not been reported. The SpO2 decrease is unlikely to be due to inactivity because similar periods of inactivity not on flights did not associate with SpO2 decrease. We suggest that the adaptation is due to altered physiology; it is unclear if frequent flying contributes to the adaptation response.
We also found a significant association of SpO2 decrease with fatigue on airline flights, which replicates findings from experiments performed in controlled laboratory conditions [40,41]. However, laboratory conditions are not a perfect replication of actual in-flight conditions, and it is important to document these changes in actual flights and in modern aircraft. This is particularly important given that the large number (approximately 2.75 billion passengers fly on commercial airlines worldwide annually [42]) and the average age of air travelers has risen and many more people with chronic diseases fly. Wearables combined with subjective and objective measures open the possibility to study a much broader range of people under real-time flight conditions and provide monitoring of the effects of reduced air pressure on individual symptoms.
Second, we found a strong association between high HR and skin temperature measurements and elevated hs-CRP levels, consistent with previous studies using nonportable devices [43]. For the 603 d of Participant #1 monitoring, elevated hs-CRP, HR, and skin temperature were evident during four periods. Interestingly, for at least one of these periods (Day 518) the participant was clinically asymptomatic, indicating that the inflammatory events were detectable by both a medical hs-CRP test and wearable devices, but not by the participant. Outside of these four periods, at low resolution, we did not detect any days with elevated hs-CRP that did not have elevated HR or skin temperature. Our high-resolution method using only HR also identifies the ill periods as well as four additional peaks and identifies the initial onset of disease for three of the four periods. For three other individuals an elevated HR was detectable during periods of high hs-CRP and illness, and skin temperature was elevated for one individual. In each case, the COH method identified an early stage of the disease. We suggest that wearable devices may be a sensitive measure for detecting certain inflammatory responses, and that in some circumstances, these may even be better than participant-reported observations.
It is possible that the use of wearables will lead to false alarms and overdiagnosis of disease. The number of false alarms will depend upon the threshold that is set, which can be personalized. It is notable that the most severe infection in this study, Lyme disease, which required physician intervention, had strong and repeated COH signals and is readily detected. Overall, we envision that these devices could be particularly powerful for individuals who are responsible for the health of others (i.e., parents and caregivers), and perhaps also for those who have historically limited health care access, including groups with low income and/or remote geography.
Of particular note was the detection of elevated skin temperature and HR as well as the decrease in SpO2 at the onset of symptoms of Lyme disease. This information was quite valuable in early diagnosis and treatment and occurred in an instance in which the characteristic “bull’s eye” rash was not observed after initial infection. Indeed, the symptoms first appeared after entry into a country in which Lyme disease was infrequent and the physician's initial recommendation of penicillin may have been an inadequate treatment. Moreover, the detection by wearables was quite robust, as outlying HR and skin temperature measurements were evident at every day of the disease. It is expected that the use of wearables for disease detection is extensible to other individuals and diseases associated with inflammation; obviously, the more serious the disease and associated inflammation, the more likely it will be detected using the portable devices. Indeed, the devices could be set to identify periods of highest inflammation (i.e., the ones that might require physician intervention) in order to reduce false alarms or avoid minor illness not requiring medical intervention.
Elevated HR is a strong predictor of cardiovascular disease and metabolic syndrome and is also associated with IR, insulin precursor presence, and the acute insulin response [44–48]. Hyperinsulinemia can trigger an increase in sympathetic activity through peripheral and central mechanisms [49–51]. Although the feedback loops are complex, this increase in sympathetic activity may contribute to the pathophysiology of IR, hypertension, and cardiovascular disease [52]. Our findings are notable in that we found a strong positive association between the difference of daytime and nighttime HR and participant SSPG levels that was independent of the effect of activity and BMI. This may indicate increased sympathetic activity reflective of complex physiological changes that are associated with IR and progression to diabetes. The fact that these differences can be measured using wearable devices raises the likelihood that this approach may someday be a useful measure for early detection of IR and risk for T2D.
Although many of the observations were originally discovered on a single person (who was also an author in this study—a potential limitation), in all cases, the results were validated on a larger cohort, demonstrating that our results can be generally applied. We also note that although up to seven devices were used simultaneously by a single person in this study, in addition to a scale to measure weight, in principle, all of the parameters measured in this study can be readily captured using two devices, a smartphone and a smartwatch, thus facilitating data collection and integration of diverse data types. Finally, we note that many more analyses of these data can be carried out, some of which are best performed using data collected from individuals operating in a more controlled setting.
In conclusion, this study demonstrates that a diverse array of measurements can be systematically obtained using portable devices and used to monitor health-related physiology and activities. These measurements are likely to be important not only in basic science research but also in a clinical setting. It is likely in the future that these devices will be used by physicians to help assess health states and guide recommendations and treatments [53,54].
The participants were enrolled in this study under the IRB protocols IRB-23602 and IRB-34907 at Stanford University; the IRB approved the study and consent forms that were used. All participants consented in writing. All clinical measurements were covered by IRB-23602, the enrollment criteria of which was 18 y of age or older. All the wearable measurements were covered by IRB-34907, with the enrollment criteria as age 13 or older. The 43 activity participants were recruited with efforts to enroll those at risk for T2D (SSPG >140 mg/dL; fasting plasma glucose >100 mg/dL, Oral Glucose Tolerance Test >140 mg/dL, Hemoglobin A1C >5.6%) along with healthy controls. Other than IR and/or moderate hyperglycemia, all participants enrolled in this study are self-reported healthy. Age, sex, and ethnicity information was available for all SpO2 participants and 38 of 43 Basis device participants and is indicated in S1 Table.
After evaluating more than 400 available wearable devices at the beginning of the study, we selected several for participants to use. The criteria for selection was (1) ability to access the raw data from the manufacturer, (2) cost, (3) overlap in measurement of at least one component with another device to assist in reproducibility, and (4) ease of use. Participant #1 wore seven portable devices for large segments of this study (Fig 1B); the remainder used a Basis device. For the SpO2 measurements, three devices (Scanadu, iHealth-finger, Masimo) were used by Participant #1; either iHealth-finger or Masimo were used by the other participants.
For the Basis data, the manufacturer securely uploaded the data to a secured cloud storage system. For other devices, the data were collected by the user’s smart phone, where the user securely transmitted the data to our repositories.
Each manufacturer and device outputs data in a unique, device-specific format. There are currently no standards and/or best practice recommendations on how the data should be recorded. Below are the data and metrics that were stored.
Additional parameters (galvanic skin response, food logging, and continuous glucose metrics) were also collected and will be the subject of another study.
For Participant #1, over 250,000 measurements were recorded daily using a combination of the MOVES App, the Basis device, and other wearables.
Many of the devices have been validated for accuracy by the manufacturer (e.g., Basis: https://258b1w36g2mmq40rp2i2rutg-wpengine.netdna-ssl.com/wp-content/uploads/2015/12/12212015_UCSF_WhitePages.pdf; http://www.mybasis.com/wp-content/uploads/2014/04/Validation-of-Basis-Science-Advanced-Sleep-Analysis.pdf). Nonetheless, we compared the SpO2 and HR measurements from the Masimo, Scanadu, iHealth-finger and Basis devices to those from our standard WA 6000 series vital signs monitor used in the clinical service laboratory at Stanford University. Measurements were taken at three or more different days. Finger measurements were made on either the right index finger or right ring (fourth proximal) finger; no detectable differences (<1%) were found when compared to the WA instrument using simultaneous measurements of the WA and wearable device, i.e., simultaneous tests were run on the wearable and WA device; as controls, finger locations were also swapped for the two devices and the WA instrument). To cover a wide range of SpO2 levels and HRs, the participant held his breath and the measurements were made simultaneously (within 2 s) on two different locations using one device and the standard instrument. For finger-based measurements, similar numbers of measurements were made switching the device locations. The comparison across devices was done by matching the time stamps. The Bland–Altman method [16,17] and Pearson correlation were applied to assess the agreement and the relationship between the wearables and the clinic devices, respectively. As shown in S1 Fig, for SpO2, 100% of the Masimo and iHealth-finger and 85% of the Scanadu measurements were within three percentage points of the WA instrument. For HR, over 95% of all portable device measurements were within three BPM of the clinical equipment (100%, 97.1%, 97.5%, and 93.5% for the Scanadu, iHealth-finger, Masimo, and Basis, respectively.) Although this percentage was slightly lower for the Basis, 100% of the Basis measurements were within the accuracy criteria of the Association for the Advancement of Medical Instrumentation for HR meters (five BPM and ±10% of the WA instrument) [55]. There was no evidence of systematic bias in the measurements (S1A–S1D Fig) with the exception of the Scanadu SpO2 measurements, in which the majority of readings were slightly higher than the clinical device and a few were much lower (S1C Fig); in all HR cases, the averages were within one BPM of the mean.
Pearson correlation analyses also revealed tight correlation of the wearables measurements with standard medical devices (R = 0.77 to 0.96, p < 0.0005; S1 Fig). The only exception was the Scanadu SpO2 measurements (R = 0.46). The Scanadu measures HR and SpO2 from the forehead, whereas all the other devices including the standard medical device record measurements from the finger. It is not clear whether our findings are due to technical differences of the device or the location of measurement [56]. Regardless, as described below, the trends for SpO2 levels (and other parameters) under different conditions are identical among each of the different devices.
In addition to the physiological measurements, we also assessed the accuracy of the activity data. First, we examined the agreement and correlation between the activity-sensing devices (Basis, MOVES, Withings) (See S1 Fig). The Pearson correlations between devices ranged from 0.74 to 0.81 (all p-values <0.00001; S1 Fig). The Bland–Altman Plots revealed that at daily step counts less than 12,000 steps, there was good agreement between the Basis and Withings devices. However, as daily step counts increased above 15,000 steps, the Basis gave higher step counts than the Withings device. Both Basis and Withings devices gave higher measurements than the MOVES device. The MOVES step measurement was compared using absolute measurements. Specifically, we manually counted 100 steps 12 separate times at three different locations (Bay area, Geneva, Uppsala) and compared the MOVES-recorded steps with the actual steps. The values recorded were found to be 0.79 +/- 0.16 SD of the actual value. To measure running distances, we compared two runs over a measured distance of 3.2 miles; the measured values were 0.94 +/- 0.04 that of the actual value. We also performed similar comparisons at different geographical locations by analyzing 13 outdoor runs at three different locations and compared the distances with those derived from Google Maps (for two locations, distance was confirmed using an automotive odometer). The values recorded were found to be 0.96 +/- 0.05 SD to that measured using Google Maps. Comparison of MOVES results with those of three runs using treadmills showed a larger difference of 0.75 +/- 0.22 SD.
To assess the agreement of measurements of steps using the Basis Peak device and MOVES and Withings applications, we compared the total number of steps per day for the 132 “nontravel days.” Time zone conversion was applied to make the three devices comparable. The Bland–Altman method and Pearson correlation were applied to assess the agreement and the relationship between the devices (S1 Fig).
We note that the devices were assessed under a limited set of conditions and that not all possible conditions were assessed.
To explore the 24-h distribution of physiological parameters, we focused on 71 “nontravel days” by excluding the days when a time zone other than the home time zone was reported by the MOVES GPS parameter. To eliminate the possible effect from jetlag, we removed the entire last traveling day and also the following 2 d after travelling. (We note the results were very similar to those when no extra days were removed, indicating that the effect of jetlag on the patterns shown in Fig 2 are small (not shown)). The mean of the physiological parameters (measured by Basis Peak) for each hour per day were reported in the heatmap Figure (S2 Fig), and the overall hourly distribution of the 71 d was summarized in box plots (Fig 2). The sleep time per hour was defined as the percentage of times designated as sleep (Basis Peak) compared to the total number of hours (71 d) in each hour window. Either the standard time or the daylight saving time was selected in the analysis, depending on the time of the year.
We binned the Basis measurements (HR, skin temperature, steps, and calories) into different activity categories (walking, running, cycling, sleep) according to the information from Basis and the MOVES app and compared the distribution in each bin (S3 Fig).
Flight information was obtained using FlightAware (https://flightaware.com/). Flight information was accessed using the FlightXML API using Python SOAP client library Suds. For Participant #1, exact take-off and landing times (within 1 min) were recorded for >95% of flights. For one out of the first 20 flights, SpO2 was measured by Masimo device only at the cruise stage, and this is the flight that does not show inverse correlation between SpO2 measurements and altitude.
Eighteen individuals participated in the flight study, and their age, ethnic background, and gender information are summarized in S1B Table. Participant #1 used Masimo, Scanadu Scout, and iHealth-finger device; Participants #16 and #44–#46 used iHealth-finger device; Participants #20, #47–#57, and #60 used Masimo device. Fig 4D shows the SpO2 distribution for all the participants (shown for Participant #1 were data recorded by iHealth-finger device).
SpO2 levels were measured by either Masimo or Scanadu Scout devices. Meanwhile, Participant #1 logged the status of “tired” and “alert,” and we compared the wearable-measured SpO2 levels between the two statuses according to the notes. To be more objective in defining fatigue, the participant also performed psychomotor vigilance test (Canadian tiredness test) (http://www.painfreesleep.ca/tiredness-test?&cid=semeOyQHbZq) in two separates flights besides the self-reported system. Specifically, this test evaluates the participant’s fatigue status by measuring the participant’s response time to a visual stimulus. For each measurement, response time to 12 visual stimuli were measured. Missed signal with response time slower than 500 milliseconds was counted as 500 milliseconds in the calculation.
The oral temperature of Participant #1 during Days 471–474 (Lyme disease infection) was measured by an oral thermometer (Day 471 8:00 a.m.: 100.7°F; Day 471 7:00 p.m.: 100.2°F; Day 472 8:00 a.m.: 98.9°F; Day 473 11:00 a.m.: 100.7°F; Day 474 3:00 p.m.: 102°F; Day 474 6:00 p.m.: 101.4°F).
To investigate the ability of wearables to predict and monitor disease, a normalization framework was developed to accommodate the dynamic change caused by different activities and make measurements comparable. To normalize resting Basis-measured HR and skin temperature, we first excluded all the measurements recorded during or immediately after exercise that usually generate large variation in physiological parameters. Specifically, all records used have step measurements of zero for at least 10 min previous to that time point (including the current minute) and are also not associated with any prediction of activity (by MOVES software, if applied), including walking, running, cycling, or flying time (by personal calendar). Data in second resolution were first converted to minute resolution by calculating the median value. After filtering the activity-related data, we further performed Z-transformation (standardize) to the measurements based on the baseline norm of sleep status and nonsleep status (predicted by Basis device). Percentage-of-outliers was defined as percentage of measurements deemed outliers for each day by comparison with the personalized, activity-specific (sleep and nonsleep at resting) mean for the overall monitoring period (the baseline value; Z-Score >2). For Participant #58, whose sleep data are missing, the data were normalized based on the personalized 24-h distribution. Data from Basis B1 and Basis Peak were normalized separately to minimize the difference between devices. Overall, a period of 679 d (from Day 63 to Day 741) was examined. In this period, Basis data were missing for 76 d, therefore the analysis was performed on the remaining 603 d in this period. To capture data on both travel and nontravel days, we defined the start and end of a day according to coordinated Universal Time, which is 7 or 8 h ahead of Pacific Daylight Time or Pacific Standard Time.
For the Day 470 flight, we assessed the Scanadu-measured SpO2 readings (flight duration time = 94 min) relative to other flights by collecting all of the SpO2 readings recorded by Scanadu during flights with similar flight time (duration time <120 min). The readings in each of the five flight stages were compared separately and the significance of the difference was assessed by two-sided Wilcoxon rank sum test.
For the analysis of illness with daily resolution, we detected abnormally elevated HR and skin temperature during the four periods reported. Specifically, we detected abnormally elevated HR (ranks #8, #2, #16, #3, and #1 out of 603 d, respectively) and skin temperature (ranks #1, #4, #10, #2, and #5 out of 603 d, respectively) for the period from Days 470–474; we detected abnormally elevated HR (ranks #7 out of 603 d) and skin temperature (ranks #12 out of 603 d) for Day 518; we detected abnormally elevated HR (Day 455 ranks #12 out of 603 d) and skin temperature (Day 456 ranks #6 out of 603 d) for the period from Days 455–456; we detected abnormally elevated HR (Day 667 ranks #4 out of 603 d) and skin temperature (Day 667 ranks #24 out of 603 d) for the period from Day 665 to Day 669.
To map inflammatory disease at higher resolution, we further analyzed the normalized HR. Specifically, we first smoothed the normalized HR using a moving average filter and then applied peak detection to identify local maxima of the smoothed signal. We used “smooth” and “findpeaks” packages in matlab to perform smooth and peak finding. To identify isolated peaks different from the global and local distribution at high confidence, we set “MinPeakHeight” to equal to two, “MinPeakDistance” to equal to “span” (3-h), and “MinPeakProminence” to equal to two. The optimized hyper-parameter “span” (3-h) was selected by training the model on Participant #1 and was applied when analyzing other individuals.
To evaluate the predictive power of the method in distinguishing the sick periods from the healthy periods, we defined a set of sick periods (positive set) based on self-reported symptoms and the relevant blood test. In the positive sets, we also included 3 d before the day when the symptom was reported or evidenced by blood work to acknowledge the fact that abnormal physiological signal might occur before the self-reported symptom. As a negative control, we followed the same rule and defined a set of periods either (1) composed of the days with normal CRP measurements or (2) composed of all days during the monitoring periods that are not included in the positive sets. We used binary scoring of each event by the presence or absence of the peak in the period. Each sick period was counted as one event. The area under the ROC curve was calculated to evaluate the classification power. We also employed cross-validation procedures to avoid overfitting to Participant #1’s data.
Of the 43 individuals tracked using Basis devices, 28 wore only Basis B1 devices, 9 wore only Basis Peak devices, and 6 wore both the Basis B1 and the Basis Peak. The Basis Peak has improved HR sensing during exercise as compared to the Basis B1; the resting HR and other parameters were comparable between the two devices. For the cohort-level analyses, 17.1 mo of Participant #1’s data were used. We used activity normalization as well as device-specific normalization, as described below, to account for potential differences between the two devices. For each of the 43 individuals, we calculated average biometric values for HR, skin temperatures, and activity. For HR and skin temperature, we used measurements occurring at time points at which there were zero steps recorded at the current time point as well zero steps recorded within the 10 s previous to that time point. These periods corresponded to activity designations of inactive, light activity, and sleep by the manufacturer’s algorithms.
The number of days recorded for each individual was calculated as the difference between the date the recording began and the date the recording ended or the date on which the data were accessed, whichever came first. We calculated the average number of steps per day for each of the 43 individuals by multiplying the average number of steps per second by the number of seconds in a day (86,400 s).
To capture daytime versus nighttime biometrics, we restricted our measurement capture window to 1 h during the day (3–4 p.m.) when our participants were awake and had taken more than 30 steps during this hour to guarantee a minimal level of activity, and compared this to 1 h during the nighttime (3–4 a.m.) when our participants appeared to be asleep and inactive with a threshold of less than five steps during this hour to ensure inactivity during sleep, but allowing for minimal measurement artifacts or limb movement during sleep.
Daily activity habit plots were created for each individual by generating smooth conditional mean lines with a 95% CI of accelerometer magnitude data by hour of day using generalized additive models (ggplot2 geom_smooth in R). Individuals were classified into one of four groups based on the peak characteristics of the curve. To automate this process, functional clustering using the R package FClust [57] was done on the activity curves to cluster members by similarity of activity curve characteristics (S2 Fig).
For the cohort that was monitored by the Basis devices, a subset of our participants had standard clinical panels (e.g., fasting plasma glucose, glycated hemoglobin [HbAlc], blood cell counts, etc.; S3 Table; performed in the Stanford clinical labs) and demographic information. The data were accessed using the Stanford Translational Research Integrated Database Environment (STRIDE) [58]. Thirty-eight participants with Basis datasets were annotated for gender (18 male and 20 female) and baseline BMI.
Twenty participants had undergone the modified insulin suppression test after an overnight fast (48). The test consisted of a 180-min octreotide (0.27μg/m2/min), insulin (0.25 μg/m2/min), and glucose (240 μg/m2/min) infusion with blood draws at minutes 150, 160, 170, and 180. Blood glucose was measured using the oximetric method, and the SSPG is the mean of the four measurements [12,13,59]. IR is defined as a SSPG ≥140 mg/dL (n = 12), and insulin sensitivity is defined as <140 mg/dL (n = 8).
We analyzed average HR and skin temperature values for men and women using the 38 Basis datasets using an unpaired, two-tailed two-sample t test with Welch correction for potential unequal variation in the two populations. Pearson correlation between the average number of steps per day and average resting HR, as well as average number of steps per day and delta BMI (year 0 [baseline] minus year 1 BMI measurements) were done using R. The evaluation of the association between steps, HR (daytime, nighttime, and difference between day and night), and SSPG was done using SAS 9.4® (SAS Institute, Inc., Cary, NC. 2013). To account for unequal variances, we used a restricted maximum likelihood approach with a robust variance estimator to estimate the regression coefficients and their 95% CIs.
|
10.1371/journal.pntd.0002787 | Epidemiology of Coxiella burnetii Infection in Africa: A OneHealth Systematic Review | Q fever is a common cause of febrile illness and community-acquired pneumonia in resource-limited settings. Coxiella burnetii, the causative pathogen, is transmitted among varied host species, but the epidemiology of the organism in Africa is poorly understood. We conducted a systematic review of C. burnetii epidemiology in Africa from a “One Health” perspective to synthesize the published data and identify knowledge gaps.
We searched nine databases to identify articles relevant to four key aspects of C. burnetii epidemiology in human and animal populations in Africa: infection prevalence; disease incidence; transmission risk factors; and infection control efforts. We identified 929 unique articles, 100 of which remained after full-text review. Of these, 41 articles describing 51 studies qualified for data extraction. Animal seroprevalence studies revealed infection by C. burnetii (≤13%) among cattle except for studies in Western and Middle Africa (18–55%). Small ruminant seroprevalence ranged from 11–33%. Human seroprevalence was <8% with the exception of studies among children and in Egypt (10–32%). Close contact with camels and rural residence were associated with increased seropositivity among humans. C. burnetii infection has been associated with livestock abortion. In human cohort studies, Q fever accounted for 2–9% of febrile illness hospitalizations and 1–3% of infective endocarditis cases. We found no studies of disease incidence estimates or disease control efforts.
C. burnetii infection is detected in humans and in a wide range of animal species across Africa, but seroprevalence varies widely by species and location. Risk factors underlying this variability are poorly understood as is the role of C. burnetii in livestock abortion. Q fever consistently accounts for a notable proportion of undifferentiated human febrile illness and infective endocarditis in cohort studies, but incidence estimates are lacking. C. burnetii presents a real yet underappreciated threat to human and animal health throughout Africa.
| Coxiella burnetii is a bacterium that can cause acute and chronic fever illness and pneumonia in humans. It is also a known cause of abortion in livestock species, and is principally transmitted to humans through contact with infected animal birth products. With growing awareness of the over-diagnosis and misclassification of malaria as the cause of fever illnesses in the tropics, including Africa, there is increased interest in the role of non-malarial causes of fever, such as C. burnetii. We performed a systematic review of the published literature on the epidemiology of C. burnetii in Africa to consolidate knowledge and identify knowledge gaps regarding the extent of this infection in humans and animals and the risk factors for infection transmission. Few studies on prevalence of infection in humans and animals used random sampling strategies, and among these only two studied linked human and animal populations. C. burnetii appears to be a common cause of severe fever illness in humans, but population-level incidence estimates are lacking. The differential risks for C. burnetii infection and potential control strategies within the various animal husbandry systems in Africa remain largely unexplored. We conclude that C. burnetii is an underappreciated threat to human and animal health throughout Africa.
| Coxiella burnetii, a zoonotic bacterial pathogen found worldwide except in New Zealand, is transmitted to humans through direct contact with milk, urine, feces, or semen from infected animals as well as inhalation of aerosolized particles from animal placentas, parturient fluids, aborted fetuses, and environmental dust [1]. While infection by C. burnetii in humans can be asymptomatic, symptomatic infection, known as Q fever, can present as an acute undifferentiated febrile illness with the possibility of focal manifestations, such as hepatitis and pneumonia. Acute disease can progress to chronic forms, such as endocarditis, in 0.5–2.0% of cases [2], [3], typically in individuals predisposed by valvular heart disease or immunodeficiency [4]. Q fever is also one of the infectious diseases that has been linked to chronic fatigue syndrome [5]. Infection by C. burnetii has been demonstrated in many animal species, but the principle reservoirs are thought to be sheep, goats, and cattle. In these livestock species infection is often asymptomatic but can cause abortion and reduce reproductive efficiency [1], [6].
Q fever has recently gained renewed attention after the largest-ever recorded outbreak which involved over 3,500 human cases in the Netherlands in 2007–2009 [7]. Recent studies in resource-limited settings have demonstrated C. burnetii as a common cause of febrile illness and community-acquired pneumonia [8]–[10]. Fever etiology research among hospitalized patients in northern Tanzania found Q fever was a more common cause of severe febrile illness than malaria [11], [12]. As control efforts have led to consistent decreases in malaria incidence throughout sub-Saharan Africa [13]–[17] the diagnosis, treatment, and control of non-malaria febrile illnesses, such as Q fever, are emerging as new public health priorities [12]. In addition to being sources for disease transmission to humans, C. burnetii infection in animals can decrease livestock productivity which can have socioeconomic and indirect health effects on humans, especially among livestock-keeping populations in resource-limited settings [18].
In light of recent findings establishing Q fever as an important cause of severe febrile illness in northern Tanzania [11], [12] and growing awareness of the potential economic impact of infection in animals, we systematically reviewed the existing literature on the epidemiology of C. burnetii infection among humans and animals in Africa. This survey aimed to consolidate knowledge and identify future research priorities for the following topics: the prevalence of C. burnetii infection in human and animal populations, including surveys of sera or shedding in body fluids; the incidence of disease due to C. burnetii in human and animal populations; risk factors for seropositivity or disease; and infection control efforts undertaken in Africa.
Nine databases were searched with the search string described in Figure 1 including all countries in the 5 United Nations (UN) sub-regions of Africa [19]. These search terms were adapted to the particular language of each database, and for those databases that did not allow the combination of Boolean operators, (q fever) OR (Coxiella burnetii) was used. Two of the databases, CABI and EBSCO Global Health, were searched with the intention to include grey literature.
Citations for all years and in all languages were compiled and de-duplicated using EndNote (Thomson Reuters, New York, USA). Abstracts were independently reviewed by two investigators (SV and MPR) using combined language competency in English, French, Spanish, and Portuguese or Google Translate (Google, Mountain View, CA, USA) and included for full-text review upon meeting predetermined criteria (Figure 2). Excluded abstracts described studies conducted outside Africa, basic science or immunology experiments, incorrect pathogens, reviews/editorials, case reports or case series, Q fever among returning travelers, periodical lay media content, diagnostic or therapeutic studies without epidemiologic data, theoretical epidemiology, duplicate data published elsewhere, textbooks/manuals, microbiologic studies without epidemiologic data, or arthropod sampling. The same criteria were applied during full-text review and to all grey literature containing sufficient information for adjudication. Cases of disagreement between the two investigators were resolved through independent review by a third investigator (JEBH).
Prevalence studies that presented evidence of current or prior infection with C. burnetii in humans and animals were included. We considered the following serologic tests and minimum antibody titer cut-offs for C. burnetii phase I and/or phase II antigen as acceptable evidence of infection in humans and animals based on expert consensus: complement fixation (CF) >1∶10 for animals [20] and ≥1∶4 for humans, microscopic agglutination test (MAT) ≥1∶4 for humans and animals, indirect fluorescent antibody (IFA) ≥1∶25 for animals and ≥1∶40 for humans, and ELISA validated against one of the above methods [20]–[23]. Capillary agglutination tests (CAT) were also accepted based on demonstrated correspondence to CF titers [23], [24]. The terms seropositive or seropositivity are used throughout to describe serologic reactions that met these titer cut-offs. For studies of pathogen shedding, confirmation of C. burnetii by nucleic acid detection, culture, or rodent inoculation was accepted. Studies of the prevalence of C. burnetii infection in humans and animals were classified by extent of study population characterization and sampling strategies, which were categorized as random (e.g., proportional, simple cluster, or simple random) or non-random. Prevalence studies that met these diagnostic criteria and used randomized sampling strategies qualified for data extraction. Prevalence studies with appropriate diagnostic criteria that used non-random sampling were included but did not undergo further data extraction.
Studies reporting disease in animals due to C. burnetii were included if the reported cases met the World Organization for Animal Health (OIE) case definition for Q fever: abortion and/or stillbirth plus confirmed presence of the bacterial agent, accomplished by 1) isolation in culture; or 2) polymerase chain reaction (PCR), in situ hybridization, or immunohistochemistry of birth products or of associated vaginal discharge [20]. Evidence of C. burnetii on placental smears with stains deemed appropriate by OIE (Stamp, Ziehl-Neelsen, Gimenez, Giemsa or modified Koster) were considered presumptive for disease and included [20]. We also included seroprevalence studies of animals with a history of abortion, as these data, although not of confirmed cases, could yield information about potential associations between C. burnetii exposure and prevalence of animal abortion.
For studies of disease in humans, acute Q fever was defined according to the US Centers for Disease Control and Prevention (CDC) case definition: a compatible fever syndrome plus four-fold rise in antibody titers to Phase II antigen or detection in clinical specimens by PCR, immunohistochemistry (IHC), or culture [25]. Phase II antigen IFA antibody titer levels for IgG≥1∶200 and IgM≥1∶50 were included as cases based on the high positive predictive value of such results [26]. Studies reporting chronic disease in humans due to Q fever were included if the reported cases met the CDC case definition for confirmed chronic Q fever: culture-negative endocarditis or infected vascular aneurysm, chronic hepatitis, osteomyelitis, osteoarthritis, or pneumonitis with no other etiology plus IFA IgG antibody to C. burnetii phase I antigen ≥1∶800 or detection in clinical specimens by PCR, IHC, or cell-culture [25].
Risk factor studies were evaluated using the same criteria for sampling design and case definitions that were applied to surveys of prevalence or disease, respectively. Risk factor analyses in prevalence studies were excluded if the prevalence study used a non-random sampling strategy.
Studies describing control efforts must have presented original data demonstrating the outcome of an intervention to decrease infection or disease incidence in human and/or animal populations.
For all qualifying studies, extracted data included study country, city or region, species, population census data when given, sample size, year of study, and diagnostic test as well as the number of seropositive or disease cases, risk factors, or control effort data where applicable. In the case of incomplete data or unclear methods, authors were contacted for further clarification when possible. Descriptive analyses of the extracted data were conducted. No quantitative meta-analysis was undertaken.
A total of 1,662 citations were identified by the search conducted on December 3, 2012. After duplicates were removed, 929 articles remained (Figure 2). Four of the six authors we contacted for further clarification responded to our inquiries. Ultimately, 100 articles describing 120 studies remained after full-text review. Forty-one articles describing 51 studies qualified for data extraction. The other 59 articles described 69 prevalence studies that did not qualify for data extraction due to non-random sampling methods (Table S1). Studies qualifying for data extraction were grouped into the following categories: 8 human and 13 animal seroprevalence; 3 animal milk shedding; 10 human disease; 7 animal abortion; and 7 human and 3 animal risk factor studies. These 51 studies spanned the years 1965–2012 and were conducted in 15 countries, mostly concentrated in the UN sub-regions of Northern, Western, and Middle Africa (Table 1 & Figure 3).
Two surveys employed a systematic sampling strategy to assess seroprevalence among linked human and animal populations. The first, from three governorates surrounding Cairo, Egypt, reported C. burnetii seropositivity in 13% of cattle, 23% of goats, 33% of sheep, 0% of buffalo, and 16% of humans in close contact with these animals [27]. The second survey, undertaken in Chad, found 80% of camels, 4% of cattle, 13% of goats, 11% of sheep, and 1% of humans in close contact were seropositive [28]. All other seroprevalence studies sampled only humans or only animal species (Table 1).
Surveys of cattle demonstrated seroprevalence ranging from 4% in Dakar, Senegal [29], to 55% around the city of Zaria, Nigeria [30], [31]. Other studies reporting cattle seroprevalence within this range were conducted in coastal Ghana (18%) [32], Cameroon's Adamawa Region (32%) [33], southern Chad (7%) [34], and South Africa's Transvaal Province (8%) [35]. Goat seropositivity ranged from 13% in Chad [28] to 23% in Egypt [27] and 24% in 8 Sudanese states [36]. Surveys of sheep revealed seroprevalences that ranged from 11% in Chad [28] to 33% in Egypt [27]. In Upper Egypt, 23% of dog sera samples indicated prior C. burnetii infection [37].
In studies of pathogen shedding in bovine milk, C. burnetii nucleic acid was detected in 22% of raw milk samples in Upper Egypt [38]. In Zaria, Nigeria, C. burnetii shedding among individual cows was reported in 63% of milk samples from extensively managed cattle and 43% of samples from semi-intensively managed cattle [30], whereas the prevalence was 26% and 22%, respectively, at the same location one year later [31]. No studies of shedding in fluids other than milk in asymptomatic humans or animals were found by our search, and no human milk shedding studies qualified for data extraction.
Seroprevalence in humans ranged from 1% in Chad [28] to 32% in a Nile Delta village in Egypt [39]. In Niamey, Niger, 10% of children ages 1 month-5 years were seropositive [40], and in Ghana's rural Ashanti Region, 17% of two-year-olds were seropositive [41]. Other surveys reported human seroprevalence at 5% in rural western Côte d'Ivoire [42], 8% among nomads sampled in rural northern Burkina Faso [43], and 5% of pregnant women attending an antenatal clinic in Dar es Salaam, Tanzania [44].
Of disease studies in humans or animals, none estimated disease incidence, and two studies [45], [46] of livestock abortions met OIE definitions for either presumptive or confirmed cases. The remaining 5 animal abortion studies were serological investigations in individuals with history of abortions [34], [47]–[50].
Of two surveys of cattle with abortions in northern Cameroon, one did not detect serological evidence of C. burnetii infection in any cattle [50], while 3% in the other study were seropositive, compared to 7% among a non-random selection of cattle without abortions [34]. In South Africa, C. burnetii was found by smear microscopy in aborted calf fetuses at all of six cattle farms sampled [46].
In the Maradi Region of Niger, 32% of goats with previous abortions were seropositive, compared to 29% of non-randomly selected goats without a history of abortion [47]. Sheep with a history of abortion in Rabat, Morocco, were more likely than those with normal births to be seropositive for C. burnetii, 33% versus 15% (p<0.01) [48]. In a survey conducted in five Tunisian governorates, 7% of sheep without past abortions were seropositive for C. burnetii compared to 12% of small ruminants with previous abortions [49]. Another Tunisian study found that 19% of small ruminants with a history of abortion had C. burnetii detected by PCR analysis of birth products or vaginal secretions [45], and in South Africa, the pathogen was found by smear microscopy in aborted lamb fetuses from all of six sheep farms sampled [46].
Human cohorts comprising individuals with infective endocarditis in Sousse and Sfax, Tunisia, as well as Algiers, Algeria, have demonstrated C. burnetii as the causative pathogen in 1–3% of cases [51]–[53]. Two studies of febrile patients in Sousse, Tunisia, serologically identified acute Q fever in 2% and 9% of hospital admissions [54], [55]. Q fever was responsible for 5% of patients with acute febrile illness hospitalized in Bobo-Dioulasso, Burkina Faso [56] as well as 3% of pediatric and 8% of adult admissions for severe febrile illness at two referral hospitals in the Kilimanjaro Region of northern Tanzania [11]. In two studies of patients admitted for community-acquired pneumonia in Yaoundé and Douala, Cameroon, 6% and 9% of persons aged >15 years had serologically-confirmed acute Q fever [57], [58]. In these studies, Q fever was the third most common etiologic agent of pneumonia, after Streptococcus pneumoniae and Mycoplasma pneumoniae. At a major hospital in Cape Town, South Africa, C. burnetii was not found to be the cause of any pneumonia cases in a 92-patient cohort [59].
Among Nigerian cattle near Zaria, no difference in seropositivity was detected for cattle managed semi-intensively versus extensively [30], [31]. In Cameroon, positive associations were found between seropositivity and cattle aged >2 years, female animals, those seen grazing with buffalo, and those for which the owner's ethnic group was recorded as Mbororo or ‘other’ when compared to Fulbe [33], [60].
In the Egypt study linking human and animal populations, rural human residents were more likely to test seropositive than those in urban areas [27]. In the linked study from Chad, human Q fever serostatus did not correlate with the proportion of seropositive animals within respective nomadic camps, and camel breeders were at higher risk for Q fever seropositivity than cattle breeders [28]. In Ghana, children of illiterate mothers had a two-fold higher risk of seropositivity compared to those of literate mothers [41]. There was no association detected between C. burnetii seropositivity and HIV serostatus in pregnant Tanzanian women [44]. In the hospitalized patient cohort in northern Tanzania, there was no difference in prevalence of acute Q fever infection in HIV-infected compared to HIV-uninfected individuals [11], and all cases of community-acquired pneumonia in the surveys of hospitalized patients in Cameroon were in HIV uninfected individuals [57], [58]. No studies of risk factors for animal disease remained after quality assessment.
No disease control studies were found by our search.
The serological data reviewed in this study reveal evidence of widespread C. burnetii infection in multiple species and multiple sites throughout Africa. However, despite evidence of the pervasiveness of this pathogen, we found only 17 studies that used appropriate case definitions to quantify disease due to C. burnetii in humans and animals. Risk factors for exposure in humans and animals have been identified in some settings, but apart from assessing for associations between HIV and Q fever, we found no other evaluation of the epidemiologic risk factors for acute disease in animals or humans. No descriptions of disease control programs appear in published literature. C. burnetii was first reported in Africa in 1947 [61], but since then, the quantity and quality of epidemiologic research for this pathogen has been limited. We identified no disease incidence estimates, and the majority of research undertaken has limited validity due to non-random sampling procedures. Further, only two investigations using random sampling procedures studied linked human and animal livestock populations.
The majority of animal prevalence studies surveyed cattle, sheep, and goats. Studies of herds in Northern Africa, the Sahel and South Africa's Transvaal Province suggest C. burnetii infection in multiple ruminant species. Seroprevalence of C. burnetii ≤13% is frequently demonstrated in cattle [27]–[29], [34], [35] and generally higher seroprevalence (11–33%) is often observed among small ruminants [27], [28], [36], [48]. However, this pattern was not observed in all studies [30]–[33]. Our review revealed wide variation in seroprevalence even in areas of close proximity such as Lake Chad, where seroprevalence in cattle ranged from 4% to 31% [28], [33]. This is consistent with findings of high regional variability within Europe [62] and highlights the importance of understanding risk factors which may operate at a local scale and may be subtle. For example, risk factor analysis revealed that cattle seropositivity differed by owner's ethnic group (Mbororo vs. Fulbe) in Cameroon, despite these groups' similar nomadic pastoralism [60]. Herds raised by unspecified ‘other ethnic groups’ had an even greater risk of infection, highlighting the need to further explore how different animal husbandry practices might modify infection risk for humans and animals. Further, the high seroprevalence observed in camels and the greater infection risk among Arab camel breeders compared to cattle breeders warrants further investigation of C. burnetii infection in camels and the potential risk factors related to camel husbandry practices [28].
Human seroprevalence was <8% [28], [42]–[44] with the exception of surveys in children and in the Nile Delta of Egypt [27], [39]–[41]. A recent survey in The Gambia found C. burnetii seroprevalence was highest in young children, although the reasons for this are still unclear [63]. In the only risk factor analysis among children found by our search, Ghanaian children with illiterate mothers were more likely to be seropositive than children with literate mothers, but no difference was attributable to other socioeconomic factors tested [41]. The study of Egyptians in close contact with animals reported a high overall seroprevalence (16%) with greater seropositivity among rural (22%) vs. urban (4%) residents [27]. In contrast, the only other linked human-animal survey studied nomads in Chad, and found a relatively low human seropositivity (1%), despite the high-risk behaviors of handling aborted animal materials and consuming raw cow's milk [28], which was shown to contain C. burnetii in 15–63% of samples from other African settings [30], [31], [38].
Only two studies [45], [46] used direct detection of C. burnetii in animal acute disease cases. The other studies investigating the potential role of C. burnetii as a cause of livestock abortions in Africa measured the seroprevalence in individuals with history of abortion as compared to individuals without history of abortion. Two cohort studies found no difference between C. burnetii seroprevalence in cattle with previous abortions compared to those without history of abortion [34], [50]. Similar studies in sheep and goats, by contrast, generally showed a higher C. burnetii seroprevalence in individuals with history of abortion compared to those without history of abortion [45], [47]–[49]. This serologic approach, however, has limited value for inferring causation in livestock abortion cases, which is compounded by non-random selection of control livestock without abortions in all but one study [30].
Q fever accounted for 2–9% of humans hospitalized for febrile illness in 3 different African sub-regions [11], [54]–[56]. Although Q fever was the third most common detected cause of community-acquired pneumonia at two of Cameroon's largest hospitals [57], [58], no cases of community-acquired pneumonia were attributed to C. burnetii after a year-long survey at a Cape Town hospital [59]. Interestingly, all of the Cameroonian patients were HIV-uninfected, and in the Tanzania severe febrile illness cohort there was no association between HIV serostatus and acute Q fever [11], [44]. The proportion of infective endocarditis cases in Africa attributed to C. burnetii was slightly lower than proportions found in European settings [64]. However, all endocarditis studies found by our review were conducted in Northern Africa [51]–[53], highlighting a key knowledge gap on the role of Q fever in endocarditis elsewhere on the continent.
We identified few studies that elucidated the epidemiologic risk factors for C. burnetii infection in humans and animals in Africa. This knowledge gap highlights the need for future studies that randomly sample linked human-animal populations in order to estimate seroprevalence and determine the dynamics of pathogen transmission. Such research requires large, representative samples as well as detailed surveys of herds and households from multiple locations, agricultural systems, and ethnic groups.
C. burnetii is clearly an important cause of human and animal disease in Africa, although illness and death have not been estimated at the population level. In animals, C. burnetii has been implicated as an etiologic agent of abortion in livestock from the most northern to the most southern reaches of the continent, but studies should include confirmatory microbiological and histological testing of abortion materials, descriptions of other disease sequelae, randomly sampled non-aborting controls, and tandem serological and bacterial shedding surveys to determine the rate of asymptomatic shedding. Estimates of economic losses due to decreases in milk production, fecundity, or birth weight are also needed. For humans, limited data suggest that Q fever frequently causes severe febrile illness in cohorts throughout Africa, yet no studies quantifying disease incidence, disability, or deaths at the population level exist. Further, aside from infective endocarditis studies, proportions and clinical features of conversion to chronic Q fever in African populations are absent.
Superseded geographic or biological terminology may have caused us to inadvertently miss pertinent research. The already remote chances of communicating with authors of older manuscripts were complicated by the absence of electronic contact information. We excluded arthropod vector studies, but surveys of invertebrates and non-domestic animals may contribute to knowledge about C. burnetii transmission. Comparisons between studies and sub-regions were restricted by changes in diagnostic methods over time, frequently small sample sizes, and the low total number of studies. The low number of studies for each research question and the heterogeneity of these studies precluded a more extensive quantitative analysis of the epidemiology of C. burnetii in Africa.
To our knowledge, this is the first systematic review of the epidemiology of C. burnetii in Africa from a ‘One Health’ perspective. Taken together, these findings suggest: 1) exposure to C. burnetii is a common finding in many animal host species across Africa, but seroprevalence varies widely by species and location, and the risk factors underlying this variability are largely unknown; 2) C. burnetii has been implicated as a cause of livestock abortion and could be responsible for substantial economic burdens, but more rigorous studies are required to determine this and other sequelae of disease in animals; 3) risk factors for human exposure to Q fever are poorly understood, but a more detailed understanding of how human exposure in different communities is linked with animal infection patterns and animal husbandry practices is clearly needed; and 4) Q fever accounts for a notable proportion of undifferentiated human febrile illness and infective endocarditis but studies describing other acute or chronic disease manifestations are scarce. The picture is complex, but the existing literature suggests that C. burnetii is found across diverse settings in Africa and presents a real yet underappreciated threat to human and animal health throughout Africa.
|
10.1371/journal.pntd.0002216 | Incidence of Rabies in Humans and Domestic Animals and People's Awareness in North Gondar Zone, Ethiopia | Rabies is a zoonotic disease that has been prevalent in humans and animals for centuries in Ethiopia and it is often dealt with using traditional practices. There is lack of accurate quantitative information on rabies both in humans and animals in Ethiopia and little is known about the awareness of the people about the disease. In this study, we estimated the incidence of rabies in humans and domestic animals, and assessed the people's awareness about the disease in North Gondar zone, Ethiopia.
The incidence of rabies in humans and domestic animals was prospectively followed up for one year period based on clinical observation. A questionnaire was also administered to 120 randomly selected dog owners and 5 traditional healers to assess the knowledge and practices about the disease. We found an annual estimated rabies incidence of 2.33 cases per 100,000 in humans, 412.83 cases per 100,000 in dogs, 19.89 cases per 100,000 in cattle, 67.68 cases per 100,000 in equines, and 14.45 cases per 100,000 in goats. Dog bite was the source of infection for all fatal rabies cases. Ninety eight percent of the questionnaire respondents were familiar with rabies and mentioned dog bite as a means of transmission. But discordant with current scientific knowledge, 84% and 32% of the respondents respectively mentioned any type of contact (irrespective of skin condition) with saliva, and inhalation as a means of transmission of rabies. Eighty four percent of the respondents relied on traditional healers for management of rabies.
The study shows high canine rabies burden, and lack of sufficient awareness about the disease and high reliance on traditional treatment that interfere with timely post exposure management. Vaccination of dogs, proper post exposure management, and increasing the awareness of the community are suggested to reduce the disease burden.
| Rabies is a fatal viral disease that affects all mammals including humans. Domestic dog is the main source of rabies for humans and livestock in developing countries. Rabies has been prevalent in Ethiopia for centuries, affecting humans and livestock. In this study we estimated the incidence of the disease in North Gondar zone, Ethiopia and assessed the people's awareness about the disease. We found a high annual rabies incidence of 2.33 cases per 100,000 in humans, 412.83 cases per 100,000 in dogs, 19.89 cases per 100,000 in cattle, 67.68 cases per 100,000 in equines, and 14.45 cases per 100,000 in goats. Although almost all interviewed people were familiar with the disease, quite a lot of them (84%) have some opinions that are incongruent with existing scientific knowledge about the cause and means of transmission of the disease. We also found high reliance on traditional healers, whose practice has not been proven effective scientifically, for the management of the disease. In conclusion, the disease poses significant public health and economic problem that warrants multi-dimensional approach towards its control. Vaccination of dogs, proper post exposure management, and increasing the awareness of the community about the disease should be considered for controlling the disease.
| Rabies is an acute encephalitis illness caused by rabies virus. Rabies virus is the prototype species of the genus Lyssavirus in the family of Rhabdoviridae. The virus affects virtually all mammals and infected species invariably die from the disease once clinical signs are manifested [1].
Rabies is endemic in developing countries of Africa and Asia, and most human deaths from the disease occur in these endemic countries [2]. Human mortality from endemic canine rabies was estimated to be 55, 000 deaths per year and was responsible for 1.74 million disability adjusted life years(DALYs) losses each year [3]. The annual cost of rabies in Africa and Asia was estimated at US$ 583.5 million most of which is due to cost of post exposure prophylaxis (PEP) [3]. Ethiopia being one of the developing countries is highly endemic for rabies. Approximately 10, 000 people were estimated to die of rabies annually in Ethiopia which makes it to be one of the worst affected countries in the world [4]. Dogs are the principal source of infection for humans and livestock [5]. In Ethiopia many households own dogs usually for guarding property. Although there are no formal studies, it is estimated that there is one owned dog per five household nationally [5]. Dog management is often poor and dog vaccination is limited to few dogs in urban centers. High population of dogs with poor management contributes for high endemicity of canine rabies in Ethiopia. In canine rabies endemic countries like Ethiopia, rabies has also significant economic importance by its effect on livestock. For example, in Africa and Asia, the annual cost of livestock losses as a result of rabies is estimated to be US$ 12.3Million [3].
In Ethiopia individuals who are exposed to rabies virus often see traditional healers for the diagnosis and treatment of the disease [5]. These widespread traditional practices of handling rabies cases are believed to interfere with timely seeking of PEP. Rabies victims specially from rural areas seek PEP treatment after exhausting the traditional medicinal intervention and usually after a loss of life from family members [5].
The available information on rabies in Ethiopia is largely based on passive reports to Ethiopian Health and Nutrition Research Institute zoonoses laboratory [6], [7], the only rabies diagnostic laboratory in the country. Passive reports usually underestimate incidence and are poor indicator of the status of the disease in countries where human and animal health information systems are inadequate [8], [9]. There is lack of accurate quantitative information on rabies both in humans and animals and little is known about the awareness of the people about the disease to apply effective control measures in Ethiopia. The objectives of the present study were, therefore, to estimate the incidence of rabies in humans and animals in North Gondar zone, Ethiopia, and assess the people's awareness of the disease in the area.
The study was ethically reviewed and approved by Gondar University Research and Publication Office. Oral informed consent was obtained from each study participant after reading written consent form. The use of oral consent was approved by ethical committee of the office considering the fact that most of the study participants didn't read and write to give their consent in writing. The interviewers confirmed the participants' oral consent by signing on the respective consent form for each interview as per ethical review guideline of the office. The consent form mainly explains about the purpose of the study, the risks and benefits of participation in the study, conditions of confidentiality and the right to refusal or withdrawal from the study, and has a signature of the participant confirming his/her informed consent.
The study was carried out in North Gondar administrative zone, Ethiopia. The zone's capital, Gondar town, is located at 12°35′N latitude and 37°29′E longitude. North Gondar zone has 16 districts, an area of 45,944.6 square kilometers, a human population of 2.9 million, and a livestock population of 2.4 million cattle, 2.36 million small ruminants and 0.3 million equines [10], [11].
In Ethiopia administrative zones comprise several districts and districts in turn comprise several kebeles (district sub units). Two stage cluster sampling technique was used to select sample for the study from North Gondar zone. Selection from the first cluster (districts) was done by judgment sampling and selection from the next cluster (kebeles) was done randomly. Accordingly, Gondar town district (which mainly contains town kebeles) and Dabat district (which mainly contains rural kebeles) were selected judgmentally to represent the urban and rural part of the zone, respectively. Six out of the total of 12 kebeles from Gondar town district and 10 out of the total of 29 kebeles from Dabat district were selected randomly by lottery system, and finally all households in the selected kebeles were included in the study. The human and livestock populations of the selected kebeles were taken from existing statistical sources [10], [11]. As the size of dog population was not available, census of dog population in the selected kebeles was done at the beginning of the study.
The incidence study was done based on a prospective follow up of suspected and exposure cases of rabies for one year from April 2009 to March 2010 in the human and domestic animal populations of the selected kebeles. The rabies suspected cases were those humans and animals showing symptoms consistent with rabies (encephalitis with spasm in response to sensory stimuli, change of temper, vocalization, drooling, paralysis and other neurological signs) without a known source of exposure. Rabies exposure cases were on the other hand those animals and humans that were exposed to a known rabies suspect case. Exposure was defined according to World Health Organization (WHO) guideline [12] and included both minor exposure and severe exposure. According to the guideline, minor exposure refers to nibbling of uncovered skin, minor scratches or abrasions without bleeding while severe exposure refers to single or multiple transdermal bites or scratches, licks on broken skin, or contamination of mucous membrane with saliva of infected animals.
Follow up data were collected by resident enumerators, recruited one from each selected kebele. The enumerators were all at least high school graduates in terms of qualification. They were trained on how to do the follow up, collect the required data and the precaution to be taken when dealing with suspected cases. They conducted their data collection task under close supervision of the investigators throughout the course of the study. Before the start of the follow up, the enumerators conducted census of the dog population by going house to house in the selected kebeles and did awareness work about the study to get the cooperation of the community during the follow up. Ownerless dogs were not included in the census but their number could not be significant as the enumeration was done immediately after stray dog elimination campaign in Gondar town district, and no significant number of ownerless dogs was expected in rural kebeles of Dabat district. Once the incidence follow-up had begun the enumerators fortnightly rounded the households in the selected kebeles to note occurrence of any suspect or exposure cases and to remind the community to report any occurrence of cases in between the rounds. When cases were encountered, they were recorded with the relevant information and the follow up continued at least for 6 months after the date of their initial record. The enumerators collected data from suspect and exposures cases in a format prepared according to the WHO guideline for rabies case reporting and surveillance [13]. Human exposures were immediately referred to health organizations for appropriate measures. Final diagnosis of cases for the incidence calculation both in humans and animals was based on history of exposure, clinical signs of the disease and its fatal end. Based on WHO case classification those cases diagnosed from the follow up of suspected cases constituted suspected rabies cases and those form exposed cases constituted probable rabies cases [14]. Submission of samples for laboratory confirmation was not possible due to lack of facilities in the region.
A questionnaire that aimed at collecting information on the people's knowledge and practices about rabies was administered by face to face interview to 120 randomly selected dog owners, 10 from each kebele of Gondar town and 6 from each kebele of Dabat district. More individuals were selected from Gondar town kebeles because of their bigger population size. Additionally 5 rabies traditional healers, who were referred by the dog owner respondents as the main providers of traditional treatment against rabies in the study area, were also interviewed. The questionnaire was designed to collect information about the respondents' knowledge of the disease, its cause and means of transmission, and treatment and prevention practices. The respondents' knowledge was validated based on their description of the disease's typical clinical and epidemiological features like neurological signs, salivation, and primarily a disease of dog that is transmissible to humans and other animals.
The data collected for both incidence and awareness study were entered into Microsoft access 2010. The data were checked for their completeness and consistency, and those incomplete and inconsistent were corrected when possible and removed otherwise. Description of the results from the cleaned data was done by descriptive statistics like percentages. The estimated incidence of the disease in humans and different species of domestic animals was calculated by dividing the clinically diagnosed cases (suspected rabies cases and probable rabies cases) by the population at risk. The incidence estimates were expressed as cases per 100, 000 individuals at risk per year
In the one year follow up of rabies incidence in humans and domestic animals in North Gondar zone, the highest estimated incidence was observed in dogs (412.83 cases per 100, 000 per year) followed by equines (67.68 cases per 100, 000 per year), and the estimated incidence in humans was 2.33 cases per 100, 000 per year (Table 1).
During the follow up period, a total of 55 cases of rabies virus exposures (32 in humans and 23 in animals) were recorded from which 16 (3 humans and 13 animals) ended with fatality. Majority (58.18%) of exposures were in humans followed by cattle (25.45%). Thirty seven (67.27%) of the exposure reports were due to bite by suspected dogs. The rest 18 (32.72%) cases were by contact of broken skin with saliva of suspected dogs and livestock, and even human in one case. Majority of exposure cases were recorded in the rural district of Dabat. Distribution of exposure cases with districts and species are presented in Table 2.
A detail follow up of the 16 fatal rabies cases of humans and animals who were diagnosed following a known exposure history revealed that the source of exposure for all of the fatal cases was dog bite. Clinical details of the fatal human and animal cases are presented in Table 3.
Almost all (98%) of the 120 questionnaire respondents were familiar with the disease and gave it slightly different local names (e.g. ‘Kelebat’, ‘Likefit’, ‘Yebed wusha beshata’) which all mean madness. One hundred three (86%) of the respondents ascribed starvation and thirst as causes of the disease in dogs; and 8(7%) included prolonged exposure to sun heat as a cause. All of the respondents who know the disease mentioned bite as a means of transmission while 101(84%) them stated any type of contact (irrespective of the skin condition) with saliva of affected individual can transmit the disease. Thirty eight (32%) of the respondents, all of whom from Dabat district, included inhalation as means of transmission.
Regarding the practices with rabies virus exposed humans, 101 (84%) respondents used traditional medicine when they feel exposed to the disease. When it is disaggregated with district, use of traditional medicine was 100% in Dabat and 65% in Gondar town. According to the respondents, most of the traditional medicines were prepared from herbs with additional spiritual rituals in some cases, and in 20 (17%) of the cases, the treatment was preceded by diagnosis. Two of the 5 traditional healers interviewed claimed that they can diagnose whether a person is exposed or not to rabies virus, and provide such diagnostic service for their clients. The traditional healers claimed wisdom acquired from sacred scriptures as basis for their diagnosis. The respondents did not give precise information about similar treatment and diagnostic practice in animals. When it is used, it was usually holy water, and only few respondents mentioned herbs as treatment of rabies in animals.
While all respondents indicated a preventive treatment for dogs, no respondent claimed similar treatment for humans or livestock. In almost all cases the preventive treatment was constituted from herb and given for dogs by cow milk at age of 2–6 months.
Dog vaccination practice assessment revealed only 24 (20%) of the total respondents, all of them form Gondar town, vaccinate their dogs regularly. Other 13 (11%) vaccinated their dogs once in vaccination campaigns during rabies outbreaks. The rest including all respondents from Dabat district had never vaccinated their dogs against rabies. The main reasons for not vaccinating were lack of awareness about dog vaccination in Dabat district and lack access and cost of vaccine in Gondar town district.
The incidences estimated in this study were based on only clinical diagnosis of the disease. This can be considered as the limitation of the incidence study. However, because of its obvious symptoms and invariably fatal consequence, estimating the incidence of rabies based on clinical diagnosis in endemic area would not that much compromise the reliability of the estimate.
Ethiopia has been considered among the most rabies affected country in the world with an estimated annual occurrence of 10, 000 cases of human rabies which is equivalent to 18.6 cases per100, 000 people [4]. In the 1980's Bogel and Motschwillor [15] had also reported 12 cases per million people which made Ethiopia the second worse affected by rabies next to India. The present estimate of 2.33 per 100, 000 in North Gondar zone is also high and lies between the two national estimates mentioned in the preceding sentences. These previous national estimates were based on passively acquired secondary data. In developing countries accurate estimates of rabies from secondary data are difficult to obtain because of poor surveillance system and inadequate regional laboratories [16]. The incidence estimate from the present study, which is based on active follow up, would be therefore a relatively accurate estimate of the disease's burden in the study area. When compared to other incidence reports in East Africa, it is comparable to a report of 2.5 cases per 100, 000 in Kenya [8] but lower than 4.9 cases per 100, 000 reported in Tanzania [17]
Studies on the incidence of rabies in dogs in Ethiopia are nonexistent except some reports based on suspected samples that are submitted to Ethiopian Nutrition and Health Research Institute rabies laboratory [6]. In the present study a high annual incidence of about 412.83/100, 000 was determined in dogs. This estimated incidence is higher than a similar study in Chad (140/100, 000 dogs) [9] but lower than a report from Kenya (860/100, 000 dogs) [8].
Data on the incidence of rabies in livestock are limited and generally considered as a sporadic occurrence [1]. In this study, incidences of 19.89 cases per 100, 000 and 67.68 cases per 100, 000 were recorded in cattle and equines, respectively. A comparable incidence of 12.3 cases per 100, 000 in cattle was reported in agro-pastoral area of Tanzania [18]. In terms of economic impact rabies is major concern in cattle. In literatures this economic effect is considered significant in areas where bat transmitted rabies is common [1]. But in areas where the dog rabies is abundant and uncontrolled like in Ethiopia, the economic impact of rabies in livestock cannot be underemphasized. Rabies in equine could also have public health significance. During a discussion with the community in this study, the authors anecdotally learnt a story of a rabid donkey that had transmitted fatal rabies to his owner through bite.
From the total of 55 human and animal exposure cases, the 16 developed the disease and died. The rest 29 cases remained normal up to the end of their follow up duration of 6 months. This could probably be due to minor degree of exposure. It is also possible that the suspected exposing animal might not shed the virus at the time of exposure (due to intermittent shedding) or might not rabid at all. Majority of exposure cases were recorded in the Dabat rural district. This was because rural households have more livestock than urban households and some of the exposed cases in the rural district were from handling of rabies suspected livestock which increased their risk of exposure. When the details of the 16 fatal cases are seen, they were all due to dog bite confirming dog as the important reservoir and source of infection for livestock and humans. This is consistent with previous reports in Ethiopia where 95% fatal cases of human rabies were associated with dog bites [6]. In developed countries where dog rabies is controlled the main rabies cycle is associated with wild carnivores [19]. But in Ethiopia wild carnivores including the endemic Ethiopian wolf (Canis simensis) are threatened by spillover of virus from domestic dog [20].
The incubation periods were variable in different species and even within species which is not unexpected as the incubation period varies from few weeks to several years depending on the amount of virus in the inoculum and site of inoculation [12]. In this study incubation period of 35 days was observed in woman bitten at head region which was much shorter than 93 days for the other woman bitten on the leg. The shortest incubation period was recorded in cattle followed by in horse and then in humans (Table 3). But generally longer incubation periods were observed in livestock as compared with literature information [1], [21]. For example, the average incubation period of 37.4 days in cattle seen this study was more than double of the 15.1 days observed in an experimental study [21]. The fatal human cases were all given post exposure treatment within 36 hours using sheep brain tissue anti-rabies vaccine but it was without effect. This could be associated with many factors like vaccine quality, storage or delivery which warrants investigation.
It was found that almost all respondents know the disease by slightly different names all meaning mad dog disease. Despite the fact that the community is familiar with the disease, many misconceptions about how it is caused and transmitted were observed. Although bite was correctly implicated as a means of transmission of the disease by all respondents, any direct or indirect saliva contact (irrespective of skin condition) by 84%, and inhalation by 34% of the respondents were also considered as means of transmission. Based on this conception of exposure, humans were subjected to crude traditional treatments which sometimes have serious negative consequences to their health. However a mere contact of saliva with intact skin do not pose risk of rabies virus exposure and exposures to rabies virus by inhalation are evidenced only under exceptional circumstances like inside bat caves where millions of bats congregate and create virus laden aerosol in the air [1]. Most of respondents believe that the disease in dogs is caused by starvation, thirst and prolonged exposure to sun heat. This view could be given probable explanation in relation to the notion of asymptomatic rabies carrier dogs in which stressors like starvation and thirst might induce development of clinical rabies in these carrier dogs. But the notion of asymptomatic rabies carrier dogs by itself is a contentious issue [22], [23], [24], and the association of stressors to the development of clinical rabies might be a farfetched claim.
Eighty four percent of the respondents were found to use traditional treatment following rabies virus exposure. Traditional treatment usage was more prevalent in Dabat districts than in Gondar town district indicating that people in the town use more of modern treatment because of either easy access or better awareness. While the traditional treatment is mostly made of herbs, the diagnosis is based on wisdom which the healers claimed to get it from sacred scriptures. The treatment using herbs, irrespective of its effect, is at least amenable to scientific scrutiny but the diagnosis part lacks scientific plausibility.
The use of traditional treatment by 84% of respondents shows high reliance on this unproven medication. Deressa et al. [5] noted that most fatal human rabies cases recoded in Ethiopia were mostly helped exhaustively by traditional healers. So much has to be done to reduce the high reliance of victims on traditional treatment by raising their awareness and increasing availability of post exposure anti-rabies vaccines. There are preliminary reports of anti-rabies activities of some commonly used herbs) in the traditional anti-rabies treatment practice in Ethiopia. For example, Deressa et al. [25] using mice model reported significant increase in survival time of rabies infected mice with treatment of chloroform and aqueous extracts of Salix subserrata leaf, and chloroform and methanol 80% extracts of Silene macroselen root as compared with control group. But more research is needed to identify and prove traditional medicines and practices that could be reliably used in dealing with rabies problem. Until then only scientifically proven methods of preventing the disease i.e. avoiding risk of exposure and urgent delivery of post exposure prophylaxis for those exposed should be promoted and applied.
Traditional treatment for animals was not as common as in humans. The most consistent and most surely uttered practice in animals was prophylactic treatment of puppies at young age so that they would never get the disease. This has been affirmed both by the users and the healers which again needs scientific scrutiny.
Dog vaccination practice was generally very low and totally nonexistent in rural district of Dabat. In Gondar town district, where there was better awareness, lack of access and cost of vaccine was raised as problem. Raising awareness about dog vaccination and improving access and affordability of the vaccine should be considered in control of the disease as dogs are the main reservoir of the disease.
In conclusion the study shows high incidence of rabies in both humans and animals that could be of significant public health and economic burden. Dogs are the main source of infection for both humans and livestock rabies. The lack sufficient awareness about the disease and high reliance on traditional treatment seen in this study could interfere with timely post exposure management and be an obstacle for control of the disease in the study area. In light of these findings vaccination of dogs, proper post exposure management and increasing the awareness of the community about the disease are suggested to reduce the disease burden.
|
10.1371/journal.pbio.1002125 | Dynamic Endothelial Cell Rearrangements Drive Developmental Vessel Regression | Patterning of functional blood vessel networks is achieved by pruning of superfluous connections. The cellular and molecular principles of vessel regression are poorly understood. Here we show that regression is mediated by dynamic and polarized migration of endothelial cells, representing anastomosis in reverse. Establishing and analyzing the first axial polarity map of all endothelial cells in a remodeling vascular network, we propose that balanced movement of cells maintains the primitive plexus under low shear conditions in a metastable dynamic state. We predict that flow-induced polarized migration of endothelial cells breaks symmetry and leads to stabilization of high flow/shear segments and regression of adjacent low flow/shear segments.
| The question of how blood vessel networks achieve their branching patterns is key to our understanding of organ formation as well as diseases that involve vascular anomalies. Regression (or pruning) of blood vessel segments is required for functional vascular branching patterns; however, the molecular basis for this is poorly understood. Here we investigate remodeling of vascular networks in the mouse retina and in zebrafish and focus on the cellular components of the endothelium—the cell layer that lines blood vessels. We use high-resolution imaging to map and analyze endothelial cell orientation in relation to blood flow direction during vascular remodeling. We identify sequential steps that characterize blood vessel regression through endothelial cell migration, finding no evidence for predicted endothelial cell death in the retina. Combining endothelial cell mapping with computational modeling of flow-induced shear forces allows a systems-level prediction of endothelial cell migration patterns that drive vascular remodeling. Our work establishes how local differences in blood flow drive endothelial cells to orientate and migrate against the direction of flow. We show that the dynamic and polarized migration of endothelial cells leads to the regression of segments under low flow and the stabilization of segments under high flow. We propose that strong flow functions as an “attractor” for endothelial cells, while poorly perfused vessels are less “attractive,” thereby promoting regression of non-functional vessel segments.
| The formation of a functionally perfused and hierarchically branched network of blood vessels is essential for vertebrate development, tissue growth, and organ physiology [1]. Together, vasculogenic vessel assembly and angiogenic sprouting establish the major axial vessels and form a rough draft of a network, which undergoes extensive remodeling to become functional. Also, in the adult, previously quiescent and functional networks can be reactivated, expanded to meet changing metabolic demands, or remodeled, as a consequence of injury or local occlusion. A large number of mouse mutants present defects in vascular remodeling [1,2], yet surprisingly little is known about the cellular principles and the molecular control of remodeling. One critical aspect of remodeling is segment regression, in which previously present connections between two vessel segments are lost. Endothelial cell death has been identified as a major mechanism of programmed regression of the ocular hyaloid vessels [3] and pupillary membrane [4], while in the rat retina, vessel regression occurs without evident cell death [5]. Dynamic imaging has confirmed these distinctions. In the pupillary membrane, network regression is associated with apoptosis-mediated flow restriction [4]. By contrast, in the zebrafish brain, real-time imaging showed that endothelial cells move out of the regressing branch and rarely undergo apoptosis [6,7]. Molecular and physical signals appear to be jointly involved in the process: delta-like ligand 4 (Dll4)/Notch signaling is required for vessel remodeling in the mouse retina, and vessel constriction promotes branch regression [8]. Low or fluctuating flow appears to predetermine branch regression, and enhanced flow protects vessel branches from regression [6]. Our previous work in mouse and zebrafish illustrated that an imbalance in Notch and Wnt/β-catenin signaling due to loss of the Notch-regulated ankyrin repeat protein (Nrarp; Q91ZA8) leads to premature vessel regression, likely as a consequence of reduced cell proliferation [9]. How physical forces and signaling pathways collectively stabilize or disrupt vessel connections remains unknown.
Here we investigate with high resolution the cellular mechanisms contributing to vessel regression in mouse and zebrafish. We find that vessel regression in mouse developmental angiogenesis is largely cell-death independent. We demonstrate that, rather, vessel regression involves dynamic rearrangement of endothelial cells, which migrate from regressing vessel segments to integrate in neighboring vessels. We propose that developmental vessel regression involves four discrete steps: (1) selection of the regressing branch, (2) lumen stenosis, (3) endothelial cell retraction, and (4) resolution of the regressing vessel segment. At the cellular level, we observe junctional arrangements similar to those found during vessel anastomosis, suggesting that vessel regression resembles morphologically anastomosis in reverse. Furthermore, we propose that endothelial cell nucleus-to-Golgi axial polarity predicts migration patterns at sites of vessel regression in vivo, and that differential flow/shear patterns in juxtaposed vessels drive asymmetries in cellular movements, thereby promoting stabilization of high-flow and regression of low-flow vessel segments.
Remodeling of primitive vascular networks through substantial regression of vessel segments is detectable as empty type IV collagen (Col.IV) matrix sleeves (Figs 1A and S1). The number of regression points per vascularized area increased only slightly as remodeling progressed during postnatal stages (Fig 1B), suggesting that regression profiles have a limited lifetime and therefore do not accumulate. Over the period analyzed, vessel regression was proportional to the total area vascularized. Programmed, vessel regression is mediated by endothelial cell apoptosis and correlates with macrophage activity [3,10]. To analyze whether developmental vessel regression in the retina involves endothelial cell death, we quantified the number of apoptotic endothelial cells (cleaved caspase 3) at different postnatal stages. We observed a mean of 78.1 ± 7.0 events (±SD, n = 8) in a whole 6-d post-natal (P6) retina, comprising approximately 16,000 endothelial cells (Fig 1C–1F and S1 Data). Although the total numbers of apoptotic endothelial cell events per retina increased over time, the ratio of apoptotic endothelial cells and the numbers of endothelial cell per vascularized retinal area remained surprisingly constant (Fig 1D–1F and S1 Data). Moreover, at P6, only 4.82% ± 0.76 (mean ± SD, n = 5) of the abundant regression profiles were associated with active endothelial cell apoptosis (Fig 1D and S1 Data). Thus, 95% of the regression events were not directly associated with endothelial apoptotic events, suggesting that vessel regression initiation or completion is largely unrelated to apoptosis in physiological vascular development in the mouse retina.
Analyzing endothelial cell configurations in regressing vessels by co-staining for intercellular adhesion molecule 2 (ICAM2) (P35330) (marking the apical/luminal endothelial cell membrane [11]), Col.IV, and isolectin B4 (IB4), we identified disrupted lumen as the first visible sign of vessel regression (Fig 2A). It was demonstrated in zebrafish that vessel regression in the brain vasculature was influenced by vessel perfusion [6]. Indeed, following perfusion of rhodamin-conjugated concanavilin-A in mouse pups, we observed that lumen disconnections were preferentially observed in rhodamin-negative vessel segments (S2A Fig). Co-labeling with vascular endothelial (VE)-cadherin (P55284) or zona occludens protein 1 (ZO1) (P39447) illustrated that the usual continuous junctions lining stable vessels as parallel lines are disrupted in branches with interrupted lumen. Instead, the junctions form isolated ring structures, often surrounding a patch of apical endothelial membrane without contact to the lumen in neighboring vessels (Fig 2B and 2C; S1 and S2 Movies). Such junctional arrangements surrounding apical membrane patches have been previously reported in early stages of lumen formation during anastomosis in the dorsolateral anastomotic vessel (DLAV) of zebrafish embryos [12]. At regression sites, they are additionally surrounded by a continuous Col.IV basement membrane (Fig 2B), suggesting that this configuration of lumen and cell junctions represents an intermediate step common to both anastomosis and regression. Indeed, by mosaic single-cell labeling using low-dose tamoxifen-induced Cre-mediated activation of membrane enhanced green fluorescent protein (eGFP) expression, we observed that endothelial cells in regressing branches extend numerous filopodia, similar to fusing endothelial tip cells, representative of activated endothelium (Fig 2D and 2E and S2 Movie). This suggests that endothelial cells could be actively migrating within the remodeling vascular plexus, and not only at the vascular sprouting front. In agreement, looking at the patterns of endothelial cell distribution in chimeric mouse retinas, we observed a lack of cohesive clonal expansion of proliferating endothelial cells (S2B Fig) [13]. These observations, although derived from static images in the retina, are nevertheless consistent with the idea that rearrangements of endothelial cells contribute to remodeling and possibly drive regression.
To gain a better understanding of directionality and coordination of cell movements, we investigated endothelial cell polarity along the axis of vessel segments in the remodeling plexus. In vitro, endothelial cells position their Golgi apparatus ahead of the nucleus in the direction of migration [14]. Using the endothelial-specific transcription factor Erg to label endothelial nuclei, the Golgi marker Golgi integral membrane protein 4 (Golph4) (Q8BXA1), together with lumen and Col.IV labeling, we determined endothelial cell nuclear shape and Golgi location in the mouse retina at P6 (Figs 3A and S3A). We established the distance between the center of mass of the endothelial nuclei and the position of the corresponding Golgi apparatus as a vector representing axial cell polarity (Fig 3A). Live imaging in transgenic zebrafish embryos confirmed the dynamic correlation between axial Golgi polarization and directional endothelial cell migration (S3 Movie). The Golgi apparatus of two anastomosing tip cells generally pointed towards the point of contact (Fig 3A), while in regressing vessels the Golgi position arrangements were reversed, suggesting that cells migrate away from each other during regression (Fig 3A and 3B). We therefore propose that regression in developing vascular plexuses is a cell migration–driven process, resembling the cellular events occurring during anastomosis in reverse order.
To directly observe cell dynamics during the process of vessel regression, we studied regression of intersegmental vessels (ISVs) during remodeling of arterial to venous ISVs in the zebrafish embryo. After the first angiogenic phase, ISVs are originated from endothelial sprouts arising from the aorta. A second angiogenic phase occurs at a later stage in development, in which secondary sprouts arising from the posterior cardinal vein (PCV) either connect with the arterial ISVs, triggering disconnection from the aorta, or instead form precursors of the zebrafish lymphatic system (Fig 4A) [15]. We generated mosaic endothelial expression of membrane-bound eGFP in Tg(kdrl:mCherry-CAAX) embryos to observe the dynamics of single endothelial cells during regression of the connection of ISVs to the aorta (Fig 4B and S4 Movie). Where venous sprouts connected to the ISV (Fig 4Bii), we observed subsequent disconnection and retraction of the arterial cells from the aorta (Fig 4Biii, iv). Similar to the mouse retina observations, the disconnection of the ISVs occurred without evident endothelial cell death (S4 Movie). Cell tracking using a zebrafish transgenic line labeling endothelial nuclei (Fig 4C and S5 Movie) also revealed migration of cells with no sign of apoptosis during and after regression. In the shown example, the regressing cell proliferates after regression (Fig 4Cv), confirming that cells involved in regression remain active and viable. When comparing vessel regression in the mouse retina and the zebrafish ISV, we could observe striking similarities in the cellular arrangements during the different phases of vessel regression (Fig 4D). On the basis of these observations, we schematized the cellular and junctional rearrangements underlying vessel branch regression (Fig 4E). We propose that vessel regression entails four distinct steps: (1) an initial selection step, which precedes and triggers the morphological alterations during regression; (2) a stenosis step, in which the lumen is focally constricted or collapsed; (3) a retraction step, in which endothelial cells migrate and retract processes, associated with junctional remodeling; (4) a resolution step, which comprises the final loss of any endothelial processes in this branch, leaving only basement membrane and pericyte(s) behind (Figs 4E and S4).
Stimulated by our analysis of the axial endothelial polarity in vessel regression and the importance of haemodynamics in regulating blood vessel pruning in the zebrafish brain [6], we took advantage of the newly developed approach for the computation of haemodynamic forces in mouse retinal vascular networks [16] to investigate the correlation between endothelial axial polarity and blood flow patterns in the mouse retina (Figs 5A and S5). Interestingly, in the retinal plexus, axial polarity vectors were largest in high flow vessels, such as arteries and arterioles, with very little variance and vectors pointing exclusively against the direction of blood flow (Figs 5B, 5C, and S6). Also in veins, polarity was generally directed against the flow, but the vectors were, in general, smaller (Figs 5B, 5C, and S6). Surprisingly, even closer to the retinal sprouting front and distant from the feeding arteries, where flow and shear levels are predicted to be low through our simulations, endothelial axial polarity is still significantly directed against flow (Fig 5A–5C, S1 Data, and S6 Fig). Linear regression analysis identified a strong correlation between increasing wall shear stress and polarization (analyzed as scalar product of polarity and shear vectors, S6 Fig and S1 Data). In order to better understand the relationship between flow-induced shear and vascular parameters across the retina, we performed unbiased combinatorial quantitative analysis of wall shear stress, cell density, and branchpoint density as a function of the distance from the optic nerve (Fig 5D, S1 Data). With increasing distance from the optic nerve, branchpoint density increased, reflecting the transition from the more remodeled to the unremodeled plexus. In parallel, the wall shear stress levels decreased with distance from the optic nerve. However, surprisingly, endothelial cell density, measured as nuclei/μm2, was highest closest to the optic nerve and decreased towards the periphery. Thus, although the increase in branch points signifies a more ramified vascular plexus in the developing periphery, endothelial cell density is lower in this region. Conversely, the highest wall shear stress found in the most central and remodeled area correlated with higher cell density (Fig 5D, S1 Data), indicating that the transition from primitive to remodeled plexus is not driven by cell loss. Given that minimal proliferation is detected in these more central areas (Fig 1E), the increased density of endothelial cells closer to the optic nerve independently argues against endothelial cell apoptosis as a driver of vascular remodeling. Instead it would be consistent with cells incorporating into higher flow segments as low-flow segments regress. To directly observe endothelial axial Golgi-to-nucleus polarity under the influence of blood flow, we analyzed transgenic zebrafish embryos expressing transiently mCherry-GM130, a Golgi-specific protein, during the process of lumenization and blood flow onset in intersegmental vessels. Endothelial cells in ISVs without continuous lumen showed variable axial polarities, with cells positioning their Golgi apparatus in the direction of migration, occasionally directed towards the aorta, however, mostly directed towards the DLAV (Fig 5E and 5F and S6 Movie). Interestingly, as ISVs formed a continuous lumen and flow is established, endothelial cells redirected their axial polarity towards the aorta and against the blood flow direction. Thus endothelial cells respond dynamically to the onset of flow and rapidly redirect their Golgi against the direction of flow (Fig 5E and S6 Movie). Quantification of axial polarization in ISVs demonstrates that during the sprouting phase arterially connected endothelial cells show dorsal polarization, which is reversed in mature arterial ISVs, to point against the flow direction (Fig 5F, S1 Data). In stable venous ISVs, axial polarity points dorsally, i.e., against the flow direction (Fig 5F, S1 Data).
Interestingly, looking at regions in the retinal vascular network showing coordinated endothelial axial polarities—around arteries and first order branches—with clear differences in wall shear stress levels between adjacent vessel segments, we observed a strong correlation amongst the lower wall shear stress vessels segments and the presence of endothelial cells with very low axial or misaligned polarity vectors (Fig 5G; S5 and S6 Movies).
Based on the present observations, we propose that endothelial cells migrate and rearrange dynamically, not only in sprouts, as shown previously [13,17,18], but also in the newly formed and the remodeling plexus. In the primitive plexus, this migratory behavior lacks an overt directionality and is thus balanced, enabling a symmetric distribution of cells throughout all segments, thus forming a uniform primitive plexus. When flow creates sufficient high shear forces on the endothelial luminal surface, this new directional force breaks the symmetry and drives polarization against the blood flow direction. This polarization directs migration of cells in low-flow or oscillatory flow segments towards the high flow segments, thus destabilizing the segment. As a consequence, there will be a net movement of cells out of the low-flow branch into the higher flow branch, thus leading to regression of the former and stabilization of the latter. Similarly, live-imaging in zebrafish brain vasculature demonstrated that regressing vessel segments exhibit low flow, which decreased irreversibly prior to the onset of regression [6]. Interestingly, the set value for shear stress when vessels enter the regression program is variable, but seems to depend on shear stress levels on juxtaposed vessel segments [6]. Taken together, we propose that increasing flow asymmetry between juxtaposed vessel segments is the trigger for developmental vessel regression (Fig 6). At the regressing segment, low-flow conditions are insufficient to establish strong continuous cell polarity within the segment. Where these cells connect to and sense higher flow in neighboring segments, the resulting polarity will lead to “attraction” of cells into the high-flow segment. In principle, axial polarity can be a component of directional cell migration or a consequence of the shear forces exerted onto the cell. Thus the migratory polarization and flow-induced polarization may be distinct events. Recent work by the Siekmann team, however, identified movement of endothelial cells in a remodeling plexus from the vein to the artery [19], thus against the predicted direction of flow. Therefore, it is also possible that the migration and flow-induced polarity events are tightly linked. The observed vessel stenosis could also be a trigger of poor perfusion, and thus polarity, or the consequence of the migratory behavior and attraction of the cells out of this segment and into the neighboring one.
Intriguingly, we noted an association of apoptotic events with long regressing vessel segments, especially when disconnecting from retinal arteries. We hypothesize that apoptosis during developmental vessel regression might be associated with a failure of endothelial cells to integrate into neighboring vessel segments.
The observed lumen stenosis may result from active endothelial contraction [8], RhoA over-activation [20], or passive lumen collapse, and conceivably could also be triggered by endothelial cell retraction. In contrast, the programmed regression of fetal ocular vessels is triggered by induced single endothelial cell apoptosis, leading to flow stasis, followed by synchronous endothelial cell apoptosis [3,4]. Similarly, experimental oxygen-induced vessel regression involves widespread endothelial cell apoptosis in the retinal vasculature [21]. Therefore, two distinct mechanisms for initiation and completion seem to be operating during vessel regression depending on the context, extent and biological requirement; (1) endothelial cell apoptosis for programmed regression of entire networks, and (2) endothelial cell migration for angiogenic remodeling. Defining the molecular mechanisms regulating each step will be critical to fully understand the process of vessel regression.
The profound motility and rearrangement of endothelial cells in the immature vascular plexus [6,13,22] implies that endothelial cells need to coordinate their cellular movements in order to maintain vessel integrity and vessel connections. In a wide range of developing tissues, the orientation and coordination of cells is dependent on intrinsic and extrinsic factors, such as morphogen gradients, extracellular matrix adhesion, cell junctions, physical forces, and cell-to-cell communication, which all participate in the correct cell polarization and coordinated cell migration [23,24]. Here, we show that endothelial cell Golgi polarity predicts migration patterns at sites of vessel regression in vivo. In the primitive plexus, where flow is low, endothelial axial polarity is less apparent, suggesting that cell movement is less directional or less collectively aligned. We hypothesize that this movement of cells maintains the primitive plexus in a metastable state of symmetry, with cells evenly distributed throughout all vessel segments.
The emerging conceptualization of vessel regression favors a model in which endothelial cells proliferate to provide sufficient numbers to support formation of the primitive plexus and are then rearranged and re-used in the process of making a functional vascular plexus to meet regional demands. Several observations support this model: (1) low number of apoptotic endothelial cells associated with regressing vessel segments; (2) a surprisingly similar number of endothelial cells per vascularized area before and after remodeling, yet with highest cell densities in central and more remodeled areas; (3) cells actively migrate from regressing vessel segments and integrate in the juxtaposed vessel network; (4) decreased proliferation of endothelial cells lead to excessive vessel regression, with cells stretching over long distances [9]. In the chick yolk sac, vessels disconnecting from the vitelline arteries are re-used for establishing new vessel connections with neighboring veins [25], and also in the mouse yolk sac, endothelial cells move from smaller into larger caliber vessels, contributing to remodeling [26].
What drives the rearrangements of cells in the primitive plexus and how flow in one segment initiates regression in another is poorly understood. Recent results show that differential VE-cadherin dynamics drive cell rearrangements in the vascular sprout [18]. Cells with higher vascular endothelial growth factor (VEGF) signaling and lower Notch activity show increased mobility by displaying a larger mobile fraction of VE-cadherin at their junctions. Whether this also holds true for events during regression is unclear. However, given that Notch is also active in remodeling [8,27], VE-cadherin is a component of endothelial cell-to-cell and fluid shear stress force sensing [28], and that VE-cadherin is implicated in coordinating endothelial polarity in collective migration [14], it is tempting to speculate that rearrangements in the primitive plexus involve Notch signaling as a driver of local differences.
All studies and procedures involving animals were in strict accordance with the European and United Kingdom Regulation for the Protection of Vertebrate Animals used for Experimental and other Scientific Purposes Animal procedures in accordance with the Home Office Animal Act 1986 under the authority of the Project License PPL 80/2391. Suffering of the animals was kept to a minimum; no procedures inflicting pain were performed.
For perfusion fix experiments, P6 pups were anaesthetized via IP injection of 0.1 ml/10 g of Ketaset/Hypnovel mix. Mouse pups were then perfused, via left ventricle intracardiac puncture, with room temperature PBSa, followed by 1% PFA solution, and finally perfused with Rhodamine labeled Concanavalin A (Vector Labs) in PBSa (0.05 mg/ml) at room temperature for 10 min. Eyes were thereafter collected for further analysis. For chimeric retina experiments, at 3.5 d post-coitus (dpc), embryos from PDGFb-iCreER; Rosa26mTmG mice were used to isolate ES cells, which were cultured in standard ES cells media with the MEK inhibitor PD0325901 (Stemgenet), as described previously [29]. ES cells were characterized and injected into 3.5-dpc staged Balb/cOlaHsd wild-type embryos and re-implanted into pseudopregnant foster females using standard protocols [30]. Newborn pups were injected intraperitoneally with Tamoxifen (Sigma; 20 μl/g of 1 mg/mL solution) at P3 before eyes were collected at P6. For endothelial proliferation assessment in the retina, mouse pups were injected IP 4 h before collection of eyes with 20 μl/g of EdU solution (0.5 mg/mL; Invitrogen, C10340).
Oxygen-dependent vessel obliteration was achieved using two different regimes of hyperoxia. At P4 (regime 1) or P7 (regime 2) pups were place in 70% oxygen chamber until P6 (regime 1) or P12 (regime 2). Animals were sacrificed immediately after hyperoxia treatment and processed for retinal vasculature analysis.
Animal procedures were performed in accordance with the Home Office Animal Act 1986 under the authority of project license PPL 80/2391.
The following transgenic zebrafish strains were used: Tg(fli1a:eGFP) [31]; Tg(fli1a:nEGFP)y7 [32]; and Tg(Fli1a:dsRedEx)um13 [33]. Transgenic zebrafish embryos Tg(fli1a:eGFP)y7 were injected with 45 ng/μl of plasmid DNA pT2Fliep-mCherryGM130 at one-cell stage. Embryos were raised at 28°C and screened for transient expression at ~30 hpf. Positive embryos were anaesthetized in 1x tricaine (0.08%) and mounted in a 35 mm glass bottom petri dish (0.17 mm, MatTek), using 0.7% low melting agarose (Sigma) containing 0.08% tricaine and 0,003% PTU. Time-lapse analysis was performed using a Leica TCS SP5, an Andor Revolution 500 spinning disk or a Zeiss LSM710 confocal microscope, with 20x or 40x objectives.
Primary and secondary antibodies are listed in S1 Table. At desired stage of development, mouse eyes were collected and fixed 5 h in ice-cold 2% paraformaldehyde (PFA) in PBS for 5 h at 4°C. Thereafter retinas were dissected in PBS. Blocking/permeabilisation was performed using Claudio’s Blocking Buffer (CBB), consisting of 1% FBS (Gibco), 3% BSA (Sigma), 0.5% triton X100 (Sigma), 0.01% Na deoxycholate (Sigma), 0,02% Na Azide (Sigma) in PBS pH = 7.4 for 2–4 h at 4°C on a rocking platform. Primary and secondary antibodies were incubated at the desired concentration in 1:1 CBB:PBS at 4°C overnight in a rocking platform. When using two rabbit primary antibodies, such for Erg and Golph4 immunofluorescence images in Figs 3, 5, S4, and S5, after first incubation with Erg primary and corresponding secondary, an additional step of incubation with donkey anti-rabbit Fab fragments (1:100, Jackson’s Laboratories) followed by 15 min fixation with 4% PFA was performed prior the incubation with Golph4 primary antibodies, in order to avoid intensive cross-reaction between the two primary antibodies. DAPI (Sigma) was used for nuclei labeling. Retinas were mounted on slides using Vectashield mounting medium (Vector Labs, H-1000). For imaging we used a Carl Zeiss LSM780 scanning confocal microscope.
Complete high-resolution three-dimensional (3-D) rendering of whole mount retinas were acquired using a LSM780 laser-scanning microscope (Zeiss). Tiled scans of whole retinas were analyzed with Imaris (Bitplane) or ImageJ. Proliferation of endothelial cells was measure by quantifying the total number of endothelial cell nuclei (labeled by Erg immunostaining) positive for EdU staining in 3–5 20x objective images, in regions containing the sprouting front, and dividing by the total area of vascularized tissue. Quantification of apoptosis in regression profiles was measured as the number of regression profiles positive for cleaved caspase-3 and divided by the total number of regression profiles in regions used for quantification, and given as percentage.
Details of the computational rheology model used to study flow patterns in the developing mouse retina can be found in [34]. Briefly, retinal vascular plexuses were stained for ICAM2 and imaged following the above-mentioned protocol. The resulting images were post-processed with Photoshop CS5 (Adobe) in order to isolate the luminal region of interest, which was further processed with MATLAB (The MathWorks, Inc.) in order to extract the image skeleton and compute vessel radii along the network. Based on the computed image skeleton and radii, a three-dimensional triangulation of the plexus luminal surface was generated with VMTK (Orobix srl). The computational fluid dynamics software package HemeLB (see [17] for more details) was used to compute high-resolution estimates of pressure, velocity, and shear stress across the domain. Flow visualization was generated with Paraview (Kitware, Inc.), and post-processing of the results was performed with custom-made Python scripts [34].
|
10.1371/journal.ppat.1004691 | Antibiotic Modulation of Capsular Exopolysaccharide and Virulence in Acinetobacter baumannii | Acinetobacter baumannii is an opportunistic pathogen of increasing importance due to its propensity for intractable multidrug-resistant infections in hospitals. All clinical isolates examined contain a conserved gene cluster, the K locus, which determines the production of complex polysaccharides, including an exopolysaccharide capsule known to protect against killing by host serum and to increase virulence in animal models of infection. Whether the polysaccharides determined by the K locus contribute to intrinsic defenses against antibiotics is unknown. We demonstrate here that mutants deficient in the exopolysaccharide capsule have lowered intrinsic resistance to peptide antibiotics, while a mutation affecting sugar precursors involved in both capsule and lipopolysaccharide synthesis sensitizes the bacterium to multiple antibiotic classes. We observed that, when grown in the presence of certain antibiotics below their MIC, including the translation inhibitors chloramphenicol and erythromycin, A. baumannii increases production of the K locus exopolysaccharide. Hyperproduction of capsular exopolysaccharide is reversible and non-mutational, and occurs concomitantly with increased resistance to the inducing antibiotic that is independent of the presence of the K locus. Strikingly, antibiotic-enhanced capsular exopolysaccharide production confers increased resistance to killing by host complement and increases virulence in a mouse model of systemic infection. Finally, we show that augmented capsule production upon antibiotic exposure is facilitated by transcriptional increases in K locus gene expression that are dependent on a two-component regulatory system, bfmRS. These studies reveal that the synthesis of capsule, a major pathogenicity determinant, is regulated in response to antibiotic stress. Our data are consistent with a model in which gene expression changes triggered by ineffectual antibiotic treatment cause A. baumannii to transition between states of low and high virulence potential, which may contribute to the opportunistic nature of the pathogen.
| Acinetobacter baumannii has gained notoriety as a cause of hospital-acquired infections that are difficult to treat due to extensive antibiotic resistance. While the microorganism rarely causes disease in the community, it commonly infects patients receiving antibiotics. The factors intrinsic to the bacterium that enable growth in the presence of antibiotics are not well characterized. Furthermore, the consequences of subinhibitory antibiotic concentrations on A. baumannii disease are unknown. Here we examined the K locus, a bacterial disease determinant responsible for the production of protective surface polysaccharides, and asked whether this determinant also contributes to antibiotic resistance. We found that K locus polysaccharides facilitate resistance to multiple antibiotics, and, unexpectedly, that the bacterium responds to certain antibiotics at subinhibitory concentrations by increasing production of capsule, the principal K locus polysaccharide. This augmented production of capsule, which is mediated by upregulation of K locus gene expression, increased the ability of the bacterium to overcome attack by the complement system, an important anti-pathogen host defense, and result in lethal disease during experimental bloodstream infection in mice. Our studies indicate that A. baumannii increases its disease-causing potential in the setting of inadequate antibiotic treatment, which may promote the development of opportunistic infections.
| Hospital-acquired infections with multidrug resistant (MDR) bacteria pose increasingly difficult challenges for patient care. These diseases are often intransigent to initial empiric broad-spectrum antibiotic therapy, delaying effective treatment and resulting in significant morbidity and mortality in already vulnerable patient populations. An emerging cause of such troublesome infections is Acinetobacter baumannii. This organism is responsible for a spectrum of diseases in susceptible individuals, including hospital-acquired pneumonia, sepsis, and wound infections [1]. A. baumannii is especially problematic in intensive care units (ICUs), where it is now among the 5 most common pathogens associated with ventilator-associated pneumonia in US hospitals [2–5]. Moreover, these infections are associated with alarming increases in drug resistance rates. A recent survey reported that most hospital-acquired A. baumannii infections are MDR [5], and strains resistant to all clinically useful antibiotics are emerging [6–8]. Observations such as these have led the Infectious Diseases Society of America and Food and Drug Administration to designate A. baumannii a high priority target for new antibiotic development [9,10]. Novel approaches to treat A. baumannii are urgently needed.
An understanding of the pathobiology of A. baumannii infections would facilitate the development of novel control strategies. These infections typically target critically ill, hospitalized patients with indwelling devices [1], in whom they can be found as colonizers before the onset of disease [11,12]. In addition, multivariate analyses have consistently identified prior or inappropriate antibiotic treatment as an independent risk factor for A. baumannii nosocomial diseases [13–15]. How antibiotics modulate host susceptibility to infection is not well understood, although several mechanisms are possible, including indirect effects on the host, such as reduction in competitive, drug-susceptible populations within the patient microbiota and/or modulation of innate immune defenses, as well as direct effects on A. baumannii physiology and virulence potential.
Regarding bacterial factors that contribute to pathogenicity, the A. baumannii envelope is associated with many of the determinants of virulence in mammalian disease models [16–22]. Among these, capsular exopolysaccharide has emerged as a universal virulence factor owing to several observations. In a study of greater than 40 A. baumannii patient isolates, almost all expressed a surface capsule [23]. Two recent bioinformatics studies of the genomes from a large number of clinical isolates have identified a sequence-variable gene cluster (the K locus) with predicted capsule biosynthesis functions [24,25]. Experiments with mutants deficient in certain of these functions, including polymer assembly and export [20] or subunit biosynthesis [17], have demonstrated roles for A. baumannii capsular exopolysaccharide in growth within soft tissue infection sites [20], lethality in a mouse septicemia model [17], defense against serum killing [17,20], and biofilm modulation [17]. Because of its importance in animal infection models and its immugenicity, the capsule has been proposed to be a target for protective antibody-based interventions [26].
How the complex surface polysaccharides determined by the K locus contribute to intrinsic antibiotic resistance is poorly characterized. The high intrinsic resistance of A. baumannii and related Gram-negative organisms such as the Pseudomonads is largely due to a cell envelope consisting of lipopolysaccharide (LPS) with low permeability to hydrophobic agents [27], “slow” porins with low permeability [28], and multiple drug efflux pumps [29]. Although not previously explored with A. baumannii, recent evidence supports a role for capsule structures in contributing to microbial defenses against macromolecular antibiotics such as antimicrobial peptides [30–35]. The A. baumannii outer envelope structures including capsule may lie at the interface of pathogenesis and drug resistance.
Here, we have examined the A. baumannii K locus with a view toward elucidating its contributions to the opportunistic nature of the bacterium. We determined how the capsular exopolysaccharide and major LPS glycoform defined by the K locus facilitate resistance to antibiotic treatment, and how antibiotics at doses insufficient to block bacterial replication cause regulatory changes in K locus expression that result in increased capsule production. We show, moreover, that this hyperproduction augments the ability of the organism to survive in the presence of host complement and cause virulent infections in mice. By uncovering conditions in which the protective exopolysaccharide capsule is hyperproduced, these studies have implications for the pathogenesis of A. baumannii infections in the antibiotic-treated patient.
The organization of the A. baumannii K locus responsible for capsule production is shown in Fig. 1A. While the gene composition of the K locus can exhibit substantial diversity across isolates, it generally encompasses a modular architecture including two highly conserved flanking units. At the 5’ end of the cluster, a highly conserved capsular polysaccharide assembly and export unit forms one universal module (Fig. 1A black arrows). This module includes wzc (A1S_0049, also referred to as ptk), having an E. coli ortholog that encodes a critical component of the export machinery controlling high-order Group 1 capsule assembly [36]. Control of capsule assembly by Wzc and its orthologs depends on a critical N-terminal tyrosine kinase domain [37]. A transposon insertion in wzc of A. baumannii strain 307–0294 completely abrogates capsular exopolysaccharide production [20]. A second highly conserved module composed of simple UDP-sugar synthesis genes lies at the 3’ end of the cluster (Fig. 1A, white arrows). The intervening region (Fig. 1A, gray arrows) contains several genes specific to each capsule type but invariably contains an initiating glycosyltransferase itrA (A1S_0061, also referred to as pglC) necessary to construct the glycan repeat-units that make up the high-order capsular exopolysaccharide as well as O-linked protein glycans [17,25].
In order to probe the functions of the complex polysaccharides determined by the A. baumannii K locus, we constructed deletions of key K locus genes in reference strain 17978 via homologous recombination. We first constructed strains harboring a deletion of wzc. After repeated attempts, we were able to isolate four such deletion mutants, each of which had the rough colony morphology expected for acapsular bacteria. Consistent with the model that these strains harbor second-site mutations required for viability, three of the strains had delays in growth kinetics compared to the WT strain (S1A Fig; isolates 16, 17 and 27) that were not rectified after replacement with the WT wzc allele (S1B Fig), although the colonies had a WT-appearing, smooth colony morphology. In the case of the fourth strain (isolate 9), when a WT wzc allele was re-introduced, a highly mucoid phenotype resulted, and we mapped the causative point mutation to bfmS (A1S_0749). The bfmS mutant generally suppressed the growth defect caused by the absence of wzc, as we could easily isolate ∆wzc strains in the bfmS mutant background. Further analysis of bfmS and its effects on capsule synthesis is presented in a subsequent section of this work. We conclude that loss of wzc in the 17978 background is likely to be lethal in the absence of suppressor mutations. ∆wzc isolates 16 and 27, for which reintroduction of wzc resulted in WT colony morphology, were utilized in the analyses that follow.
Decreased viability with mutations in complex polysaccharide synthesis pathways in other systems is thought to involve accumulation of toxic intermediates or the sequestration of essential lipid carriers [38]. We therefore generated additional deletions that would be predicted to bypass these lethal processes. We constructed a deletion, termed ∆KL3, that eliminates both the capsule export module as well as biosynthesis genes of the K locus (A1S_0049 through A1S_0061, see Fig. 1A)[17,25]. Mutants were also constructed by deleting the initiating glycosyltransferase itrA, or galU (A1S_0062), the first gene of the simple UDP-sugar synthesis module that encodes a predicted UTP-glucose-1-phosphate uridylyltransferase [24,25]. Multiple independent isolates of these mutants, each of which exhibited the expected rough colony morphology, were easily obtained, and the isolates had growth kinetics that matched that of WT (S1C, D Fig).
Like most A. baumannii isolates, strain 17978 displays a thin capsule (Fig. 1B, WT), and we assessed the various K locus mutants using India ink staining. The K locus mutants lacked the smooth, uniform capsule of the WT and were instead associated with a patchy, thinner outline (Fig. 1B). Reintroduction of the corresponding WT allele in the chromosome of the ∆wzc, ∆itrA, and ∆galU mutants restored the WT-appearing, smooth capsular halo (Fig. 1B). We next examined the polysaccharides produced by each strain after fractionation of cell lysates by SDS-PAGE and staining with the polysaccharide-specific dye, alcian blue (see Materials and Methods)[39,40]. With the WT strain, 4 major bands were detected (Fig. 1C). Band 3 depends on the genes of the K locus (Fig. 1C and S1E Fig) and has a migration pattern consistent with capsular exopolysaccharide [17]. Bands 1 and 2 co-migrate with LPS [41,42], with band 1 consistent with truncated, deep rough glycoforms of LPS [19]. Band 2 is consistent with LPS containing substituents that require the function of K locus proteins (Fig. 1C and S1E Fig). The galU deletion alters the migration of band 2 (Fig. 1C white arrow), consistent with a partially truncated LPS intermediate. Band 4 represents an additional high-molecular weight (MW) polysaccharide that is not dependent on the K locus, which may be poly-N-acetyl-glucosamine (PNAG) [43].
Stocks of commonly studied A. baumannii strains (17978, 19606, and 17961) generate mucoviscous colony variants at a high frequency that alter band 3 content (Fig. 2A and C). These mutants are associated with viscous, sticky strings when lifted with a toothpick. Two of the three mutants displayed capsules with increased thickness upon inspection with India ink (Fig. 2B), and lysates of all three mutants contained a very high MW polysaccharide (Fig. 2C, “cell” panel, filled arrowhead) that replaced the band 3 capsular polysaccharide of WT. The majority of this polysaccharide was found in culture supernatants (Fig. 2D, “sup” panel, filled arrowhead). Genome sequencing indicates that each mucoviscous variant has a SNP in the wzc regulatory autokinase domain (Fig. 2D, white arrows). To test the model that these mutations alter Wzc activity, we probed the A. baumannii lysates with anti-phosphotyrosine. We note that, when a phosphotyrosine signal was detected in the resulting western blots, only a single band was present; this band depended on Wzc and matched its size (Fig. 2E, open arrowhead), indicating that it represents Wzc autophosphorylation. Consistent with the mutational changes altering Wzc activity, the mucoviscous mutants have lowered autophosphorylation levels compared to the WT strain (Fig. 2C, open arrowhead). Expressing the wzc point mutants on a plasmid in a ∆wzc background reconstituted the production of loosely associated, very high MW polysaccharide (Fig. 2F). These results indicate that autophosphorylation of Wzc negatively regulates capsule polymer chain length.
Engineered mutations of key Walker Box residues (K547, D649) or the C-terminal phosphorylation site tyrosines (Fig. 2D, black arrows) in Wzc support the model that the autokinase activity negatively regulates capsule production. Each of these alleles caused very high MW polysaccharide production (Fig. 2F), although autophosphorylation was detected in some of the engineered Walker box mutants (Fig. 2F). Together these data demonstrate that the principal A. baumannii exopolysaccharide is a high MW capsule whose assembly is controlled by Wzc.
We next interrogated the roles of K locus polysaccharides in defense against antibiotics using the deletion mutants with specific deficiencies in these structures. Studies with diverse pathogens have shown that a polysaccharide capsule can confer resistance to antimicrobial peptides [30–35] and other large antibiotics [32]. We tested whether mutants deficient in the A. baumannii capsular exopolysaccharide have increased sensitivity to the antimicrobial peptide colistin (Col), a critical last-line antibiotic for MDR A. baumannii, as well as the bulky, hydrophobic antibiotics erythromycin (Em) and rifampicin (Rif). To quantitate antibiotic resistance, we enumerated the growth of bacterial populations on agar containing serial dilutions of each antibiotic. With this assay, the ∆wzc and ∆itrA mutants, which are completely deficient in capsule but have an intact LPS, consistently displayed increased sensitivity to Col, and reintroduction of the respective WT allele restored resistance to near WT-levels (Fig. 3A and Table 1). Compared with Col, resistances to Rif and Em were affected to a lesser extent and less consistently across the mutants (Fig. 3A and Table 1). Resistance to the small antibiotic chloramphenicol (Cm) was unaffected (Fig. 3A and Table 1). We note that because the ∆itrA mutant is impaired in both capsule production and O-linked protein glycosylation [17], we cannot exclude the possibility that a loss of protein glycans contributed to the phenotype of decreased antibiotic resistance observed with this strain.
We next tested the mutants altered in production of both the exopolysaccharide capsule and the K locus-dependent LPS glycoform. Similar to the strains deficient in only capsule, the ∆galU mutant, which is deficient in capsule and expresses a partially truncated LPS, showed increased sensitivity to Col, along with mild increases in sensitivity to Rif and Em; these phenotypes reverted to WT upon reintroduction of the galU gene (Fig. 3B and Table 1). By contrast, the ∆KL3 mutant demonstrated a large increase in sensitivity to Col, Rif, and Em (Fig. 3B and Table 1), consistent with the phenotypes of Gram-negative mutants lacking core LPS sugars [16,44,45]. Moreover, at concentrations of Em and Rif that permitted 100% colony forming efficiency (CFE) with ∆KL3 (Fig. 3B, arrows), growth inhibition was observed (Fig. 3C). Mutants with altered envelope permeability due to LPS defects usually are not hypersusceptible to small, hydrophilic antibiotics, which generally enter the cell though porin channels [46]. Consistent with this idea, ∆KL3 had no increased susceptibility to the small antibiotic Cm (Fig. 3B and Table 1). These data demonstrate that the K locus polysaccharides facilitate intrinsic resistance to multiple classes of antibiotics, with the exopolysaccharide capsule contributing principally to defense against antimicrobial peptides.
While performing antibiotic sensitivity experiments we made the unexpected observation that growth on Cm and Em causes A. baumannii to assume a hypermucoid state (Fig. 4A). Hypermucoidy was not seen with Rif or Col, but was seen with Cm and Em across many clinical A. baumannii isolates. The colonies were highly mucoid but were not associated with viscous strings when picked, indicating a phenotype distinct from the mucoviscous Wzc kinase mutants above. Hypermucoidy occurred over a wide range of antibiotic concentrations at sub-MIC including concentrations that that showed no loss of CFE (Em, 2–16 μg/ml, and Cm, 8–64 μg/ml; see Fig. 3), suggesting that the mucoid colonies were not mutants. The mucoid phenotype is lost upon restreaking onto plates lacking antibiotics, and reappears upon restreaking onto Cm plates (Fig. 4B); this holds true after ten such passages, further supporting that the phenomenon is non-mutational. Hypermucoidy was dependent on the K locus genes and the ability to produce capsular exopolysaccharide, because the K locus deletion mutants prevented this phenotype, and reintroduction of the respective WT genes restored it (Fig. 4A).
To begin to analyze the connection between antibiotic exposure and exopolysaccharide production, we first examined the effects of sub-MIC antibiotics on polysaccharide levels. Using Cm as prototype, we treated logarithmically growing cells with concentrations (10 and 20 μg/ml) that are below the MIC (128 μg/ml; see Fig. 3) but that inhibit growth in liquid culture by about 50 and 65%, respectively (Figs. 4E and 4F). Cells challenged with Cm at 10 μg/ml displayed slightly thickened capsules as assessed by India ink (Fig. 4C). Furthermore, both cell-associated and cell-free capsular exopolysaccharide were increased approximately two- to three-fold based on SDS-PAGE fractionation (Fig. 4D and 4E). Of note, the migration pattern and apparent MW of the capsular exopolysaccharide was identical with untreated cells, indicating that polymer length was unchanged (Fig. 4D). The induction by Cm of capsular exopolysaccharide was dependent on the K locus genes (S2A Fig), and increased production of K locus-independent polysaccharides was not observed.
We next determined the kinetics of capsular exopolysaccharide induction in broth culture by Cm by analyzing fractionated culture lysates and supernatants over multiple post-treatment time points. With untreated cultures, cell-associated capsular exopolysaccharide levels generally peaked upon entry into post-exponential phase and then steadily declined (Fig. 4F, black squares, and S2B Fig). In contrast, cultures treated with Cm resulted in increased capsular polysaccharides as early as one hour after drug addition with continued accumulation over time (Fig. 4F, gray squares, and S2B Fig). Cell-associated LPS levels, however, appeared unaffected by Cm treatment (Fig. 4F and S2B Fig). These results indicate that Cm treatment induces a rapid and specific increase in capsular exopolysaccharide content.
The reversible plating phenotype and rapid kinetics of capsular exopolysaccharide induction with drug treatment support the idea that this phenomenon is regulatory and non-mutational. To understand the extent to which mutant subpopulations may be involved in the exopolysaccharide hyperproduction phenomenon, we performed two additional tests. First, we determined if mutants constitutively hyperproducing capsular exopolysaccharide were selected after drug treatment under the conditions used in Fig. 4D-F by analyzing multiple lineages treated with drugs. As diagrammed in Fig. 5A, 4 independent lineages of WT A. baumannii not previously exposed to antibiotics were grown with Cm for 20h and then passaged twice on solid media lacking antibiotics. Cells were then reinoculated into liquid broth and again grown with or without Cm. Supernatants from cultures before (d2) and after (d5) passage on solid media without antibiotics were analyzed for differences in the levels of basal and drug-induced cell-free capsular exopolysaccharides to discern whether drug treatment selected for constitutive exopolysaccharide hyperproducers. As seen in Fig. 5A, the levels of capsular exopolysaccharide in the d5 cultures not treated with Cm were identical to those at d2, as were the levels of induced capsular exopolysaccharide in the Cm-treated cultures. In addition, all colonies appearing during passage on media without antibiotics on d3–4 were observed to be WT-appearing and non-mucoid. These results indicate that the phenotypic increase in capsular exopolysaccharide levels by Cm addition at sub-MIC was not the result of an outgrowth of exopolysaccharide hyperproducing mutants, and instead was due to a highly reversible, population-wide regulatory response.
Second, to determine if subpopulations of mutationally drug-resistant clones were selected, bacteria from d2 (Fig. 5A) were also analyzed for high-level resistance to Cm by quantifying the number of bacteria able to form colonies at increasing concentrations of drug. Pretreatment with sub-MIC Cm (Fig. 5B, gray squares) resulted in an increase in the proportion of the population able to grow on Cm at 32–256 μg/ml compared to cells pregrown without the antibiotic (Fig. 5B, black squares). A similar effect was observed with ∆KL3 bacteria, indicating that the increased resistance was independent of the K locus polysaccharides (Fig. 5B, circles). The colonies arising after plating Cm-pretreated WT A. baumannii on Cm at 32–256 μg/ml (arrowheads) were then tested for mutational resistance at 100 μg/ml Cm. The increased growth on 32–64 μg/ml Cm was not mutational, because all clones isolated from these concentrations behaved as the parental WT when retested for high-level Cm resistance at 100 μg/ml Cm (open arrowheads). This phenotypic increase in CFE by sub-MIC Cm may be related to a number of possible mechanisms, one of which may be upregulation of efflux activities as shown with P. aeruginosa [47]. By contrast, the increased growth on 128–256 μg/ml Cm, which is above the MIC, was mutational because most or all of the clones at these concentrations had high-level Cm resistance when retested for growth on Cm100 plates (blue arrowheads). These data indicate that exposure to sub-MIC Cm stimulates multiple bacterial responses including increased capsule production and induction of K locus-independent resistance mechanisms. Furthermore, these responses are largely non-mutational.
Augmented capsular exopolysaccharide expression in response to environmental triggers has implications for multiple facets of disease with A. baumannii. We first asked whether capsule hyperproduction increases resistance to the peptide antibiotic Col, to which, unlike Cm, full intrinsic resistance depends on the presence of capsular exopolysaccharide. Increasing capsule production by transient pre-treatment with sub-MIC Cm, however, conferred no additional protection from the bactericidal activity of Col (S3 Fig).
Because capsule determines evasion of killing by the complement system [17,20], an important host defense mechanism, we tested the hypothesis that capsule hyper-production upon antibiotic exposure can confer increased resistance to complement killing. A. baumannii cells were grown in the presence or absence of sub-MIC Cm and then incubated with rabbit serum as a source of complement, followed by serial dilution and plating onto medium without antibiotics to determine viable counts. In the absence of Cm pretreatment, incubation in serum resulted in 2–3 logs of killing with WT bacteria, and approximately 5 logs of killing with the ∆itrA mutant which lacks capsule but has an intact LPS (Fig. 6A). This killing was consistent with a previous study using human serum [17], and was entirely dependent on active complement components because incubation with heat-inactivated serum resulted in complete survival in all cases (Fig. 6A). Notably, pre-treatment of WT bacteria with sub-MIC Cm conferred increased survival compared to untreated cells, resulting in high-level serum resistance (Fig. 6A). The ability to acquire high-level serum resistance required capsular exopolysaccharide, because pre-treatment of the ∆itrA mutant with sub-MIC Cm did not result in a commensurate level of protection (Fig. 6A). These data demonstrate that treatment with an antibiotic at sub-MIC levels sufficient to augment capsule production increases resistance to the lethal effects of complement.
Based on this result, we predicted that antibiotic-treated bacteria would be more virulent during bloodstream infection. To evaluate virulence, bacteremia was established in mice by intraperitoneal injection of A. baumannii [48]. In experiments analyzing the survival of infected mice with this model, we found that a dose of approximately 108 WT bacteria grown under standard laboratory conditions was unable to cause lethality (Fig. 6B), consistent with previous reports demonstrating the low virulence of strain 17978 [49,50]. When enhanced capsule production was induced by pre-treating bacteria with sub-MIC Cm, however, a significant decrease in host survival was observed at the same dose (Fig. 6B; P = 0.001). We hypothesized that this elevated virulence was the outcome of increased bacterial loads in the bloodstream and deep tissues of the host. To test this, we analyzed bacterial burdens in the blood and spleen of additional groups of mice infected for 12 hours. Consistent with the above hypothesis, antibiotic pre-treatment of bacteria resulted in a >3 log increase in bacterial counts in the bloodstream and a >2 log increase in bacterial counts in the spleen (Fig. 6C, D). Moreover, in these experiments, all mice infected with control bacteria appeared healthy, but the vast majority of mice infected with antibiotic pre-treated bacteria displayed signs of illness at the 12-hour endpoint (Fig. 6E). Together, these data demonstrate that exposure to antibiotics that enhance capsule production manifests in increased virulence during systemic infection.
To understand how induction of capsular polysaccharide production by antibiotics occurs, we analyzed the effects of sub-MIC Cm on K locus gene transcripts after 2 hours of drug exposure, focusing on 3 representative K locus genes critical for capsule synthesis: wzc; galU; and gnaA, the first gene in the unit divergently transcribed from the export module (see Fig. 1A) that encodes a UDP-D-GlcNAc to UDP-D-GlcNAcA dehydrogenase [24,25]. Compared to untreated cells, exposure to sub-MIC Cm resulted in significantly increased transcripts for each of these genes (Fig. 7A). The transcriptional increase was modest with 10 μg/ml Cm but commensurate with the increased levels of capsular exopolysaccharide detected at 2 hours with this condition (Fig. 4F). Since capsular exopolysaccharide is expected to accumulate with Cm treatment during post-exponential phase, these early increases in transcript levels are also in agreement with the overall increases in exopolysaccharide levels observed after prolonged exposure in post-exponential phase cultures (Fig. 4E).
We explored the possibility that exopolysaccharide hyper-production could additionally involve post-transcriptional changes in K locus gene expression. While Wzc appeared to be the principal protein modified by phosphotyrosine in cell lysates, we hypothesized that sub-MIC Cm may result in novel phosphotyrosine modifications that could contribute to enhanced capsule biosynthesis activities. Treatment with neither low nor high concentrations of Cm, however, resulted in any detectable novel phosphotyrosine signals after lysates were analyzed by western blotting (S4 Fig).
In E. coli, Cm and Em exposure induces changes in the transcription levels of cold shock genes [51–54]. This is in contrast to other protein synthesis inhibitors such as aminoglycosides that induce an opposing response virtually identical to that observed with heat-shock [53]. We have observed that the aminoglycosides gentamicin and streptomycin do not result in hypermucoid colony morphology or substantial increases in exopolysaccharide production, consistent with the model that antibiotics inducing cold shock are associated with enhanced exopolysaccharide synthesis. To test this model, we identified orthologs of the cold-shock genes hscA and cspA [52,55] in A. baumannii and probed their transcription levels after treatment with sub-MIC Cm. As shown in Fig. 7B, Cm at 20 μg/ml results in transcriptional induction of these genes, while treatment with 10 μg/ml results in an increase in only hscA transcript at the time point analyzed, consistent with cold shock gene induction coinciding with K locus gene induction.
We noted that the hypermucoid phenotype assumed by WT cells growing with sub-MIC Cm or Em was highly similar to the hypermucoid plate phenotype that resulted when a WT wzc allele was re-introduced to ∆wzc isolate 9, which contains a suppressor mutation that mapped to bfmS. BfmS is the receptor histidine kinase (HK) component of bfmRS, a highly conserved TCS in A. baumannii initially identified as a locus critical for biofilm formation and production of pili [56], and shown to modulate porin localization [57]. In addition, results from two genome-wide screens have indicated a role for BfmRS in promoting A. baumannii survival within the mammalian host [22,58]. Because of the genetic interaction between wzc and bfmS, in which mutation of bfmS suppresses the toxicity of ∆wzc, and the highly similar plate phenotypes between a bfmS strain and sub-MIC antibiotic treatment, we sought to characterize how BfmRS affects K locus expression in response to antibiotics.
A group of spontaneous hypermucoid mutants was isolated that had amino acid changes in the bfmRS locus (Table 2 and Materials and Methods). In BfmS, these mutations truncate the protein or are predicted to disrupt key motifs within the HK domain (Fig. 8A). In BfmR, the mutated residue (Fig. 8A) corresponds to the exact position in the close E. coli ortholog RstA required for phosphotransfer [59], with the result that the mutant is expected to be phosphorylation-negative. These observations all fit a model in which the BfmS HK negatively regulates capsule expression through phosphorylation of its cognate receiver. This model predicts that deletion of bfmS would lead to constitutive capsule hyperproduction, while deletion of bfmR or the complete two-gene operon would not.
We tested this model by constructing the respective deletions and analyzing their effects on polysaccharide production. As predicted by the model, ∆bfmS resulted in a hypermucoid plate phenotype (Fig. 8B) and slightly thickened capsule (Fig. 8C), as did the truncation allele affecting the HK domain, bfmS1–467. Deletion of bfmS in another A. baumannii strain, 19606, also resulted in a hypermucoid plate phenotype (S5A Fig). By contrast, ∆bfmR and ∆bfmRS, which showed identical phenotypes, resulted in rough colony morphologies with no appreciable increase in capsule by India ink staining (Fig. 8B, C; ∆bfmR shown in S5B, C Fig). Fractionation of polysaccharides by SDS-PAGE revealed that capsular exopolysaccharide levels were consistent with these phenotypes. Compared to WT, ∆bfmS and bfmS1–467 resulted in increased cell-associated and cell-free capsular polysaccharides, while ∆bfmRS caused an overall decrease (Fig. 8D, E). These changes in polysaccharide abundance were due to transcriptional changes in K locus gene expression, as demonstrated by qRT-PCR analysis of wzc, gnaA, and galU gene transcripts (Fig. 8F). Together these results demonstrate that the BfmRS TCS sets the level of K locus gene expression, and are consistent with a model in which phosphorylation of BfmR by its cognate sensor HK negatively regulates its ability to induce capsular exopolysaccharide production.
We next determined whether bfmRS was required for exopolysaccharide induction in response to sub-MIC Cm treatment. The presence of ∆bfmRS blocked the hypermucoid response on plates containing Cm, while the WT bfmRS allele restored the response to Cm (Fig. 9A). Although there was a partial increase in capsular exopolysaccharide levels in the ∆bfmRS strain in the presence of Cm, the mutant lacked the robust response observed in a strain harboring an intact bfmRS locus (Fig. 9B, C). Transcription analysis revealed that during culture of the ∆bfmRS with the antibiotic there was little to no early induction of K locus gene transcripts (Fig. 9D). Induction of cold shock gene transcripts was similarly muted, although in the case of hscA we detected a partial response (Fig. 9E). Introduction of the WT bfmRS allele restored these transcriptional responses (Fig. 9D, E). These results indicate that bfmRS contributes to the transcriptional induction of K locus and cold shock gene expression upon exposure to sub-MIC Cm.
With this report, we have delineated the roles of K locus genes in the production of surface polysaccharide structures and analyzed their contributions to the growth of A. baumannii in the presence of antibiotics. We demonstrate that the high-order capsular exopolysaccharide and LPS glycoform dependent on the K locus facilitate intrinsic resistance against diverse antibiotics, consistent with their roles in providing a barrier function. Unexpectedly, our study uncovered that production of the A. baumannii capsular exopolysaccharide actively responds to antibiotic treatment. At sub-MIC, the translation inhibitors Cm and Em augment the production of capsular exopolysaccharide in a rapid and reversible manner, and strikingly this results in enhanced resistance to complement-dependent serum killing and increased virulence during disseminated infection in mice. We determined that capsular exopolysaccharide hyperproduction upon Cm exposure is a regulated process involving transcriptional increases in the expression of genes responsible for its biosynthesis and export. These changes in capsule production occur alongside regulatory changes that promote resistance to the inducing antibiotic independent of the capsule. In further support of the idea that the capsular exopolysaccharide is stress responsive, we observed that its production is also regulated by a TCS, bfmRS, signals from which facilitate early transcriptional induction of K locus genes by antibiotics. Antibiotic-polysaccharide interactions have multiple implications for the opportunistic nature of the pathogen, as discussed below.
These results are consistent with previous literature supporting roles for sub-MIC antibiotics in altering bacterial environmental responses [60], including defense against stressors. An earlier report with the environmental bacterium Acinetobacter venetianus RAG-1 showed that Cm induces the synthesis and secretion of the exopolysaccharide emulsan [61]. In that organism, the exopolysaccharide is determined by a gene cluster with similar overall organization to those in A. baumannii [62], suggesting a common mode of gene regulation upon antibiotic stress that may be a defining feature of the genus. In E. coli, a range of translation-inhibitor antibiotics including Cm, Em, and aminoglycosides augments poly-N-acetyl-glucosamine (PNAG) production and biofilm mass by a mechanism involving the second messengers ppGpp and c-di-GMP [63]. Aminoglycoside antibiotics also induce biofilm formation in some clinical strains of P. aeruginosa [64,65], although an opposite response has been observed with the macrolide azithromycin at sub-MIC [66]. In the case of Klebsiella pneumoniae, ciprofloxacin (a fluoroquinolone) and ceftazidime (a ß-lactam) were shown to increase capsule production [67]. It appears that across bacterial species, a range of physiological responses exists with different antibiotics that may influence behaviors important in the development and control of drug-resistant infections.
We have begun to uncover a regulatory circuit governing capsular exopolysaccharide production that involves multiple components. The BfmRS TCS appears to set the level of expression of K locus genes by controlling their transcription. According to our model, under basal conditions the transcription-promoting effects of BfmR are negatively regulated by phosphorylation signals from BfmS. Upon loss of these signals, as is the case with the bfmS null mutants and possibly with certain antibiotic stresses, phosphoregulation of BfmR is relieved and K locus gene expression is transcriptionally activated. Transcriptional control may account for only part of K locus gene induction upon antibiotic exposure, because a ∆bfmRS strain still displays a partial increase in capsular exopolysaccharide in response to sub-MIC Cm.
An additional component implicated in capsule regulation and antibiotic sensing is the ribosome. Cm and Em both target the 50S ribosomal subunit, and in doing so produce lesions in the ribosome functionally similar to that experienced with temperature downshift; it has been proposed that ribosome sensing of such a stress results in altered second messenger signaling (e.g., ppGpp) and thereby altered transcriptional responses which in E. coli mirror the cold shock response [53]. We therefore tested whether cold shock genes are similarly induced under capsule-stimulatory conditions in A. baumannii. We found that sub-MIC Cm indeed induces at least a subset of cold-shock gene homologs, and that the bfmRS mutant is altered in cold-shock gene expression and unable to fully induce cold shock gene transcripts early in response to sub-MIC Cm. These findings are consistent with a regulatory network controlling capsular exopolysaccharide that involves interactions between bfmRS and cold shock genes.
That the production of a major virulence determinant, the K locus capsular exopolysaccharide, is responsive to antibiotic stress has a possible clinical correlate in the observation that a major risk factor for opportunistic A. baumannii infections is prior or inappropriate treatment with antibiotics. We found that conditional hyperproduction of capsular exopolysaccharide induced upon exposure to sub-MIC levels of Cm allows the bacterium to overcome a key humoral defense of the host—killing by the complement system. These data add to previous work with other microorganisms on the ability of antibiotics to antagonize the bactericidal effects of serum [68]. Capsular exopolysaccharide in A. baumannii may block the function of serum complement by one or more possible means that remain to be investigated, including decreased binding of early complement components or decreased access of late complement complexes to sites required for efficient lysis. In support of the hypothesis that increased serum survival results in a more virulent organism, antibiotic-exposed bacteria hyperproducing capsular exopolysaccharide were more virulent and resulted in higher levels of bacteremia than control inocula not treated with antibiotics after intraperitoneal injection. These data are consistent with previous studies with other pathogens showing that overproduction of capsular exopolysaccharide increases virulence [69–71]. Our results are also consistent with a recent study by Bruhn and colleagues [49], which found that lethal A. baumannii disease was associated with isolates that have the ability to establish high bacterial loads in blood very early after inoculation. Our experiments involving pretreatment of bacteria with an antibiotic prior to inoculation allow an isolate that is deficient at establishing early high bacterial loads in the bloodstream to mimic the behavior of highly virulent isolates. In total, our data suggest a model wherein stimulation of enhanced capsule production, as in a patient receiving inappropriate antibiotic treatment, facilitates a transition from a low-virulence, colonizing state to one of higher virulence and invasiveness, increasing the likelihood of opportunistic disease. Whether increased protection from phagocyte killing, a prominent feature of the aforementioned studies, also contributes to the increased fitness of hyper-encapsulated A. baumannii during systemic infection is yet to be determined.
An additional consequence of hyperproduction of exopolysaccharides due to sub-MIC antibiotic exposure is that it is predicted to increase resistance of the organism against desiccation [72] and detergents [73,74]. As a consequence, exposure to low-level antibiotics should promote persistence of A. baumannii in the hospital environment. Although Cm is rarely used in developed countries, macrolides related to Em are widely used in varied settings, and are often included in empiric treatment of hospital-acquired infections for patients with multiple comorbidities or immunocompromised states [75]. Furthermore, other antibiotics more commonly administered in ICU settings such as carbapenems induce a mucoid state [76]. Therefore, inappropriate antibiotic administration may facilitate colonization in environmental reservoirs in which the organism poses the greatest risks.
In summary, these studies demonstrate a connection between regulation of the pathogenicity-determining capsular exopolysaccharide of A. baumannii and antibiotic exposure, and identify a two-component regulator that is critical for modulating this interface. These data are consistent with a model wherein environmental cues such as antibiotic stress directly alter the virulence potential of the bacterium, with possible influences on the development of nosocomial disease. Disrupting the regulatory loops that connect stress sensing with pathogenicity in A. baumannii may represent a strategy to control infections with these opportunists.
Bacterial strains used in this study are listed in S1 Table. A. baumannii reference strain ATCC 17978 was used throughout unless otherwise noted. Bacteria were grown in Lysogeny Broth (LB) or on LB agar plates. Carbenicillin (100 μg/ml), gentamicin (Gm, 15 μg/ml), and/or kanamycin (Km, 30 μg/ml) were used in selection of recombinant strains. Antibiotics were purchased from Sigma. Assay subcultures of A. baumannii were grown without antibiotics in the case of chromosomal markers, and with the appropriate antibiotics in the case of plasmids. Cultures were grown at 37°C in flasks with orbital shaking at 250 RPM or in tubes rotated at maximum speed with a roller drum. Growth as optical density (OD) was monitored by measuring absorbance at 600nm.
Plasmids used in this study are listed in S1 Table, and oligonucleotides (purchased from Integrated DNA Technologies) are listed in S2 Table. All constructs were sequenced (Genewiz) before introduction into A. baumannii via electroporation [77] or conjugation. The ∆wzc, ∆KL3, ∆bfmS, ∆bfmR and ∆bfmRS deletions were constructed as allelic replacements with the aacC1 GmR cassette by amplifying 1–2kb of 17978 genomic flanking DNA, three-way ligation with pUC18 or its derivatives, insertion of the SacI or KpnI fragment of pFGM1 containing aacC1, and subcloning the resulting tripartite construct into pSR47S. pSR47S constructs were crossed into the A. baumannii chromosome via homologous recombination and GmR KmR merodiploids were isolated. GmR KmS double recombinants were isolated by sucrose counterselection. Markerless, in-frame deletions of itrA and galU were constructed as above, but without the insertion of aacC1. KmS double recombinants were isolated from KmR merodiploids as above. All deletion mutants were verified by colony PCR.
Point mutations were engineered in wzc by cloning the gene into pUC19, introducing mutations with primers covering native or engineered restriction sites via inverse PCR and self-ligation, or by amplifying and substituting a mutated gene fragment, followed by subcloning into pJB1801. Marker rescue of wzc was achieved by amplifying the gene and flanking regions, subcloning into pSR47S, and isolating double recombinants in EGA106 as above. Single-copy itrA and galU constructs for complementation tests were generated by cloning the genes into miniTn7 elements carried on pUC18T-miniTn7T-Gm. itrA was cloned with an upstream tetp promoter (amplified from pBR322), and galU was cloned with an upstream fragment containing the K locus core promoter region (amplified from 17978 genomic DNA). Insertion of the miniTn7 elements was performed by four-parental mating as described [78], but the helper plasmid pTNS3 [79] was used and exconjugants were selected on Vogel-Bonner base medium plates [80] containing Gm15. Insertion into the chromosomal attTn7 site was verified by PCR as described [78]. A single-copy bfmRS construct for complementation tests was generated by amplifying and subcloning the operon into pEGE148. The resulting plasmid was crossed into EGA251 at the ∆bfmRS::aacC1 locus.
Illumina whole-genome sequencing and analysis were performed as in [81]. Polymorphisms identified by genome sequencing of EGA2MV, EGA57, EGA127 and WT strains were used to guide PCR analysis of the corresponding loci in additional spontaneous mutants. Mucoviscous variants containing mutations in wzc were isolated after reviving lyophilized stocks obtained from ATCC. Mucoid bacteria containing mutations in bfmRS were isolated either during construction of ∆wzc or of unrelated recombinant strains. In the latter case, while attempting to isolate double recombinant strains from a sacB+ merodiploid, we recovered a series of mutants that retained sacB but had adventitious mutations, subsequently mapped to bfmRS, that permitted growth on sucrose and resulted in constitutively mucoid morphology on LB plates.
Bacterial capsules were visualized by the wet-film India Ink method of Duguid [82]. Images were acquired on a Zeiss Axiovert 200m microscope with 100x/1.3 lens.
Extracts of A. baumannii polysaccharides from whole-cell lysates were prepared by the method of Hitchcock and Brown [83] with the following additional steps to decrease DNA and RNA content. Cells were pelleted, resuspended with lysis buffer [60mM Tris, pH 8; 10mM MgCl2; 50μM CaCl2; 20μl/ml DNase and RNase; and 3mg/ml lysozyme], and incubated at 37°C for 1 hour followed by vortexing and 3 liquid nitrogen/37°C freeze-thaw cycles. Additional DNase and RNase were added and the samples were incubated at 37°C for 30 min. SDS was added to 0.5%, followed by incubation at 37°C for an additional 30 min. The samples were then processed with boiling and proteinase K treatment as previously described [83]. Polysaccharides in culture supernatants were precipitated in 75% ice-cold ethanol overnight, followed by pelleting, air-drying, resuspending with SDS sample buffer at a volume normalized based on A600, and boiling for 5 minutes. Commensurate amounts of cell lysates and supernatant precipitants were loaded in SDS-PAGE.
Samples were separated on 4–20% BioRad TGX Tris-glycine gels and stained overnight with alcian blue as in [40]. Gels were imaged with a G-box QX4 with white-light converter (Syngene). Band intensity was quantified with GeneQuant (Syngene) via the manual band quantification method, using inter-lane spaces as background. For detection of phosphotyrosines, lysates were blotted onto PVDF membranes, incubated with 4G10 monoclonal antibodies (Millipore; 1:1000 dilution) followed by HRP-conjugated goat anti-mouse (Invitrogen), and developed with ECL-plus substrate (Perkin Elmer). Blots were stripped and re-probed with antiserum raised against Bacillus subtilis isocitrate dehydrogenase (ICDH).
Bacteria grown to early post-exponential phase were normalized to OD 2, serially diluted in PBS, and spotted onto LB agar plates without antibiotic or containing serial dilutions of antibiotic. Plates were incubated at 37°C overnight and the resulting CFU were enumerated. Colony forming efficiency is defined as (# CFU on the antibiotic test plate x dilution factor)/(# CFU on the antibiotic-free plate x dilution factor).
Bacteria were grown to mid-log (OD 0.4–0.5), divided into two equal volumes to which 0 or 10 μg/ml of Cm was added, and grown for an additional 4.5 hours. Approximately 105 bacteria diluted in PBS were then incubated with 30% baby rabbit complement serum (AbD Serotec) for one hour at 37°C. Control reactions were performed with serum inactivated by heating (56°C for 30 minutes). Reactions were stopped by placing on ice, and viable bacterial counts were determined by plating serial dilutions on LB agar followed by overnight incubation at 37°C.
8–10 week old female C57BL/6 mice obtained from Jackson Laboratories were used for intraperitoneal infections. Bacterial cultures were grown with or without induction by sub-MIC Cm as in serum bactericidal assays. Infections were initiated by intraperitoneal injection of approximately 108 bacteria suspended in 100ul PBS into groups of mice (7–8 mice per group for survival studies; 5 per group for analyses of bacterial counts). To analyze bacterial counts, mice were euthanized after 12 hours of infection; spleens were removed aseptically and homogenized in 1ml PBS, and blood was collected via cardiac puncture followed by immediate four-fold dilution with ice-cold PBS/10mM EDTA. Viable counts were determined by plating serial dilutions as above.
RNA was extracted via the RNeasy kit (Qiagen), DNase treated with the DNA-free kit (Ambion), and reverse transcribed with Superscript II Reverse Transcriptase (Invitrogen). cDNA was amplified with SYBR Green Master Mix (Applied Biosystems) via a StepOnePlus system according to the manufacturer’s instructions. Melting curves were obtained with each experiment to confirm reaction specificity, and reactions were performed with parallel samples lacking reverse transcriptase to verify the absence of signal from residual genomic DNA. Target amplification efficiency with each primer pair (S2 Table) was determined by obtaining a standard curve with a dilution series of cDNA and was found to be >97% in each case. Fold change in gene expression was calculated by using the 2-∆∆Ct method [84] with 16S as endogenous control.
|
10.1371/journal.pbio.1001194 | Elimination of the Vesicular Acetylcholine Transporter in the Striatum Reveals Regulation of Behaviour by Cholinergic-Glutamatergic Co-Transmission | Cholinergic neurons in the striatum are thought to play major regulatory functions in motor behaviour and reward. These neurons express two vesicular transporters that can load either acetylcholine or glutamate into synaptic vesicles. Consequently cholinergic neurons can release both neurotransmitters, making it difficult to discern their individual contributions for the regulation of striatal functions. Here we have dissected the specific roles of acetylcholine release for striatal-dependent behaviour in mice by selective elimination of the vesicular acetylcholine transporter (VAChT) from striatal cholinergic neurons. Analysis of several behavioural parameters indicates that elimination of VAChT had only marginal consequences in striatum-related tasks and did not affect spontaneous locomotion, cocaine-induced hyperactivity, or its reward properties. However, dopaminergic sensitivity of medium spiny neurons (MSN) and the behavioural outputs in response to direct dopaminergic agonists were enhanced, likely due to increased expression/function of dopamine receptors in the striatum. These observations indicate that previous functions attributed to striatal cholinergic neurons in spontaneous locomotor activity and in the rewarding responses to cocaine are mediated by glutamate and not by acetylcholine release. Our experiments demonstrate how one population of neurons can use two distinct neurotransmitters to differentially regulate a given circuitry. The data also raise the possibility of using VAChT as a target to boost dopaminergic function and decrease high striatal cholinergic activity, common neurochemical alterations in individuals affected with Parkinson's disease.
| The neurotransmitters dopamine and acetylcholine play opposite roles in the striatum (a brain region involved in motor control and reward-related behaviour), and their balance is thought to be critical for striatal function. Acetylcholine in the striatum has been linked to a number of functions, including control of locomotor activity and response to drugs of abuse. However, striatal cholinergic interneurons can also release glutamate (in addition to acetylcholine) and it is presently unclear how these two neurotransmitters regulate striatal-dependent behaviour. Previous work has attempted to resolve this issue by ablating cholinergic neurons in the striatum, but this causes loss of both cholinergic and glutamatergic neurotransmission. In this study, we created a novel genetic mouse model which allowed us to selectively interfere with secretion of acetylcholine in the striatum, while leaving total striatal glutamate release intact. In these mice, we observed minimally altered behavioural responses to cocaine, suggesting that striatal glutamate, rather than acetylcholine, is critical for cocaine-induced behavioural manifestations. Furthermore, elimination of striatal acetylcholine release affects how striatal output neurons respond to dopamine, by up-regulating dopaminergic receptors and changing behavioural responses to dopaminergic agonists. Our experiments highlight a previously unappreciated physiological role of cholinergic-glutamatergic co-transmission and demonstrate how a population of neurons can use two distinct neurotransmitters to differentially regulate behaviour.
| The striatum is the major input gateway to the basal ganglia. Striatal activity plays important roles in controlling motor functions and goal-directed and reward-related behaviours [1]–[4]. The striatum is the brain region mostly affected in motor diseases, such as Parkinson's disease (PD), Huntington's disease, and dystonia [5]. Medium spiny GABAergic neurons (MSN), activated by corticostriatal glutamatergic inputs, are the major output neurons for the striatum; these neurons are regulated extensively by the classical neurotransmitters dopamine and acetylcholine (ACh) [1],[2],[4],[6]. These two neurotransmitters have reciprocal relationships, regulating each other's release at different levels, and they generally have opposing actions in the direct and indirect striatal pathways [1],[5],[7]–[9]. Regulation of MSNs by dopamine has received considerable attention, largely due to the well-known effects of reduced dopamine levels leading to motor symptoms in PD [10] and the role of dopamine in the effect of drugs of abuse [11].
In contrast to the widely known effects of dopamine in the striatum, we know considerably less about how ACh shapes striatal function. Cholinergic neurons form a small population of aspiny and large striatal interneurons that provide the sole source of cholinergic innervation to MSNs [12],[13]. These neurons fire constantly and therefore ensure relatively high levels of extracellular ACh. To maintain high levels of transmitter release, cholinergic neurons transport ACh synthesized in the cytoplasm into synaptic vesicles, a process which requires the activity of the vesicular acetylcholine transporter (SLC18A3, VAChT [14],[15]), the last cholinergic-specific step for ACh-mediated neurotransmission [16]. A variety of muscarinic receptors [17], as well as nicotinic subtypes of receptors [18]–[21], involved in controlling striatal function add complexity to unravelling the role of endogenous ACh in the striatum. To make matters more difficult, several central cholinergic neurons express both VAChT and distinct vesicular glutamate transporters (VGLUTs) and thus are able to store and release both ACh and glutamate [22]. Striatal cholinergic neurons express VGLUT3 [23]–[26] and simultaneously release glutamate and ACh [27]. It is unknown, however, if cholinergic neurons can use both neurotransmitters to regulate striatal function.
Elimination of cholinergic neurons in the striatum, using ablation strategies, indicated that these neurons have a role in regulating spontaneous and cocaine-induced locomotor activity, as well as its rewarding properties [28]–[31]. These neurons have the capacity to release both ACh and glutamate; therefore, non-selective manipulations of striatal cholinergic neurons can affect both VAChT and VGLUT-mediated neurotransmission. Interestingly, mice null for VGLUT3 phenocopy many of the behavioural alterations found in mice that had their accumbens cholinergic neurons ablated [25]. However, because VGLUT3-null mice also presented a 40% decrease on acetylcholine release, it is difficult to discern the individual effects of these two neurotransmitters. Therefore, the specific roles of ACh for striatal function have not yet been addressed.
To investigate the possibility that cholinergic neurons can use these two distinct neurotransmitters differentially to regulate striatal circuitry, we generated a novel mouse line in which we selectively eliminated ACh release by deleting the VAChT gene in the striatum. Our results reveal specific roles for ACh release in regulating dopamine receptor-mediated locomotor responses, but suggest that some of the previous functions attributed to these neurons are related to their ability to release glutamate.
To address specific roles of ACh release in striatal function we generated a VAChT floxed mouse line (VAChTflox/flox, [32]), as constitutive VAChT knockout mice do not survive birth due to impaired breathing [16]. The addition of lox P sites did not change VAChT expression at the mRNA and protein levels when compared to wild-type control mice. VAChTflox/flox mice had normal levels of VAChT and other pre-synaptic cholinergic markers. In addition locomotor activity, grip-strength, and fatigue were identical in VAChTflox/flox mice and wild-type mice [32].
In order to selectively eliminate VAChT in the striatum, we used the D2-Cre bacterial artificial chromosome (BAC) transgenic mouse line generated by GENSAT [33], which expresses the enzyme Cre recombinase under the control of regulatory elements of the D2 dopamine receptor (D2R). Details related to this mouse line, including control experiments demonstrating that the expression of Cre has no effects on the parameters studied here, are presented in Experimental Procedures and Figure S6. To test whether Cre was expressed in striatal cholinergic neurons, we crossed D2-Cre mice to Rosa26 reporter mice (Rosa26-YFP mice), in which the Rosa26 locus expresses YFP once Cre-mediated recombination has occurred (Figure 1a). We found that in D2-Cre;Rosa26-YFP mice almost 100% of striatal cholinergic neurons identified with an antibody against CHT1 also showed Cre-recombination (YFP staining 98% co-localization, Table S1). We did not detect co-localization of YFP in cholinergic neurons in the penduculopontine nucleus or in motoneurons in the brainstem (Figure S1 and Table S1). Partial localization of YFP in cholinergic neurons was detected in the basal forebrain, albeit to a much lower extent than in the striatum (approx. 50%, Figure 1b and Table S1). We therefore intercrossed D2-Cre mice to VAChTflox/flox mice to generate mice with selective elimination of VAChT in the striatum (VAChTD2-Cre-flox/flox) or control mice (VAChTflox/flox). Genotyping for these lines is shown in Figure S2. VAChTD2-Cre-flox/flox mice were born in the expected Mendellian ratio and did not present overt phenotypes. We found no gross morphological alterations in the striatum or other brain sections stained with hematoxylin/eosin in VAChTD2-Cre-flox/flox mice compared to control mice (unpublished data).
To assess the degree of Cre-mediated recombination we evaluated the expression of VAChT in the striatum of VAChTD2-Cre-flox/flox. As expected, based on the observations with the D2-Cre;Rosa26-YFP mice, both mRNA and protein levels for VAChT were almost abolished in the striatum of VAChTD2-Cre-flox/flox (Figure 2a,d,g). In contrast, choline acetyltransferase (ChAT) and the high-affinity choline transporter (CHT1) protein levels were not altered (Figure 2e and f). There was no difference in VAChT protein expression levels in the hippocampus of VAChTD2-Cre-flox/flox mice when compared to controls (Figure 2h and i). Accordingly, release of [3H]-ACh was abolished in striatal slices from VAChTD2-Cre-flox/flox mice depolarized with high KCl, whereas it was identical to controls in hippocampal slices (Figure 3a and b).
Acetylcholine can modulate glutamate release via pre-synaptic nicotinic receptors in projection glutamatergic nerve-terminals [34]. In addition, striatal cholinergic neurons can also release glutamate [27]. Therefore, we examined if there was any effect of VAChT elimination on glutamate release. Isolated nerve terminals were obtained from striatal tissue of VAChTD2-Cre-flox/flox and control mice and glutamate release was stimulated by KCl. We did not detect changes in glutamate release from isolated nerve terminals in VAChT-deficient mice compared to controls (Figure 3c). It should be noted, however, that this method does not separate terminals containing VGLUT3 from nerve terminals containing other VGLUTs, and therefore only reflects global changes in glutamate release. Moreover, VGLUT3 mRNA expression by qPCR did not differ in VAChTD2-Cre-flox/flox mice compared to control mice (Figure 3d). These results suggest that overall glutamate release is not grossly altered in these mice.
Because we detected the presence of Cre-mediated recombination in motoneurons in the spinal cord (Figure S1), which could affect the behavioural performance in VAChTD2-Cre-flox/flox mice, we examined the cholinergic system in the spinal cord of VAChTD2-Cre-flox/flox mice. We did not find alterations in mRNA levels for VAChT in the spinal cord of VAChTD2-Cre-flox/flox mice (Figure S3). However, we detected an increase in ChAT mRNA and protein levels in the spinal cord. Surprisingly, there was also about a 50% decrease in VAChT protein levels. Previous experiments showed that up to a 50% decrease in the expression of VAChT in the spinal cord is well tolerated in mice and does not alter motor function [35],[36]. In agreement with these previous results, VAChTD2-Cre-flox/flox mice showed no difference in grip-force strength (Figure S4a, t(47) = 1.702, p = 0.095) or fatigue (detected by the Wire-hang task, Figure S4b, Mann-Whitney, T(13) = 49, p = 0.710). Interestingly, we also found that relative to controls, VAChTD2-Cre-flox/flox mice showed no deficit in motor performance or motor learning assessed using the rotarod test (Figure S4c, Repeated Measures ANOVA reveal no difference between the two genotypes with respect to time to fall, F(1,261) = 0.0000409, p = 0.995; both sets of mice improved their performance, F(9,261) = 41.614, p<0.001; and there was no interaction between genotype and session, F(9,261) = 1.333, p = 0.220). These results show that despite a decrease in the levels of VAChT in the spinal cord there were no detectable changes in motor function. The rotarod experiments also suggest that VAChTD2-Cre-flox/flox mice are physically fit and that motor learning does not appear to depend on striatal cholinergic activity.
Next, as a further control experiment, we determined if the elimination of ACh release in the striatum could interfere with cognitive performance that is believed to be generally independent of striatal function. We used object recognition memory, a task that is thought to be dependent on the hippocampus [37],[38] and perirhinal cortex [39], and has been previously shown to be sensitive to global decreases in VAChT levels [35],[36],[40]. In this test VAChTD2-Cre-flox/flox mice performed identically to controls, suggesting that important cognitive functions are preserved in this new mouse line (Figure S4d, two-way ANOVA revealed no effect of genotype, F(1,16) = 0.651, p = 0.431, a significant effect for object, F(1,16) = 21.559, p<0.001 and no Object × Genotype interaction, F(1,16) = 0.0185, p = 0.893).
Previous experiments have shown that the density of cholinergic neurons in the accumbens, as well as expression of ChAT, is decreased in the post-mortem brain of schizophrenic individuals [41],[42]. Moreover, partial ablation of cholinergic striatal neurons caused alterations in sensorimotor gating [43]. Therefore, we used habituation to acoustic startle and pre-pulse inhibition to assess sensorimotor gating, but found no effects of elimination of striatal VAChT on these parameters (Figure S5). These results show that decreased striatal ACh release does not cause sensorimotor gating dysfunctions in these animals and likely in schizophrenia as well.
There are controversial views regarding the role of striatal cholinergic neurons in locomotion. Previous experiments in which cholinergic neurons in the nucleus accumbens were ablated indicated that loss of these neurons caused hyperlocomotion and increased sensitivity to the locomotor effects of cocaine [29]–[31]. However, more recent experiments using an optogenetics approach failed to detect an increased locomotor activity in mice in which striatal cholinergic neurons were acutely silenced [44]. In agreement with the latter, we found no differences in locomotor activity when we compared VAChTD2-Cre-flox/flox mice to controls (Figure 4a). The dynamics of total horizontal activity (Figure 4a and 4b, t(48) = 0.1464; p = 0.884) or counts of vertical activity (unpublished data, t(24) = 1.027; p = 0.315) were essentially identical in the two strains. Importantly, in control experiments D2-Cre mice did not differ in locomotor activity from respective wild type mice (Figure S6).
It has been shown that VGLUT3-null mice present hyperactivity, which was attributed to decreased ACh release from striatal cholinergic neurons due to decreased filling of synaptic vesicles with ACh [25]. Because these experiments with VGLUT3-null mice were performed in the initial hours of the dark cycle, we reproduced these conditions with a new cohort of our mice. The VAChTD2-Cre-flox/flox mice were no more active than their control counterparts during the first hours of the dark cycle (Figure 4c and d, repeated measures ANOVA shows no main effect of genotype, F(1,1593) = 0.321, p = 0.576, significant effect of time, F(59,1593) = 14.411, p<0.001 and no interaction Genotype × Time, F(59,1593) = 0.947, p = 0.591; total activity was not different, Mann-Whitney, T(29) = 213, p = 0.431). Finally, we also tested inter-session habituation by investigating locomotor activity in 3 consecutive days in the open-field (Figure 4e). We observed that both genotypes habituated similarly to the open-field. Repeated measures ANOVA confirmed that the general activity was the same for both genotypes (genotype factor, F(1,58) = 0.932, p = 0.342). The activity decreased over the day (day factor, F(2,58) = 10.244, p<0.001) and both genotypes habituated to the environment at comparable rates (interaction between genotype and day, F(2,58) = 1.506, p = 0.230). Evidently, deletion of VAChT in the striatum does not affect general spontaneous activity or compromise the capacity to habituate to a new environment.
Previous experiments in mice in which cholinergic interneurons were ablated suggested that decreased ACh levels increase sensitivity of mice to the locomotor effects of cocaine [29]–[31]. However, these experiments did not separate the effects of VAChT and VGLUT3-mediated transmission. Interestingly, VGLUT3-null mice are also more sensitive to the locomotor effects of cocaine, a result that was attributed at least in part to a decrease in striatal ACh release [25]. Due to the surprising observations of normal locomotor activity in VAChT-deficient mice, we investigated the specific effects of the elimination of VAChT-mediated neurotransmission on the actions of cocaine. Administration of 5, 20, or 40 mg/kg of cocaine increased locomotor activity in VAChTflox/flox mice and VAChTD2-Cre-flox/flox mice (Figure 5c, two-factor ANOVAs show a significant effect of genotype, F(1,51) = 6.531, p = 0.014, significant effect of treatment, F(3,51) = 15.611, p<0.001, and no Genotype × Treatment interaction, F(3,51) = 0.983, p = 0.381). There was no difference between the two genotypes in their ability to increase activity in response to cocaine-injected i.p. at 5 mg/kg dose (Figure 5a, 5 mg/kg, repeated measures ANOVAs show no effect of genotype, F(1,322) = 0.201, p = 0.661, significant effect of time, F(23,322) = 12.820, p<0.001, and no Time × Genotype interaction, F(23,322) = 1.373, p = 0.121). Paradoxically, at 20 mg/kg VAChTD2-Cre-flox/flox mice showed a smaller effect of cocaine in locomotor activity than controls (Figure 5b, 20 mg/kg, repeated measures ANOVA shows significant effect of genotype, F(1,480) = 11.345, p<0.001, significant effect of time, F(23,480) = 9.464, p<0.001, and no Time × Genotype interaction, F(23,480) = 0.945, p = 0.537). Analysis of total activity counts showed a clear effect of genotype (Figure 5c, Mann-Whitney, T(23) = 166.000, p<0.05). At 40 mg/kg both genotypes showed similar responses (Figure 5c, t(14) = 0.980, p = 0.344), suggesting that lack of striatal VAChT altered the response to 20 mg/kg of cocaine, but overall did not cause increased sensitivity to locomotor effects of cocaine.
Cocaine increases firing of striatal cholinergic neurons [44] and the release of ACh in the striatum [45]–[47]. Previous experiments have suggested that striatal cholinergic neurons also play important roles in the rewarding effects of cocaine. Indeed, optogenetic silencing of striatal cholinergic neurons seemed to attenuate the response of cocaine in a conditioned-place preference (CPP) paradigm. Because these experiments did not separate the contribution of ACh from that of glutamate and to determine if there was a causal link between ACh release and expression of cocaine-induced CPP, we performed CPP experiments with VAChTD2-Cre-flox/flox mice. We were unable to obtain reliable CPP with either genotype at 5 mg/kg of cocaine (unpublished data). In contrast, at 20 mg/kg we detected robust CPP in both genotypes (Figure 6a, repeated measures ANOVAs show no effect of genotype, F(1,10) = 0.443, p = 0.521, significant effect of treatment, F(1,10) = 86.033, p<0.001, and no Genotype × Treatment interaction, F(1,10) = 0.0118, p = 0.916). In these experiments we used an extended protocol [48] with consecutive injections of cocaine in alternate days. We repeated the short protocol used before in the optogenetic experiments [44] with only one injection of cocaine (20 mg/kg), but we were unable to detect place preference in control or VAChTD2-Cre-flox/flox mice (unpublished data). In addition, neither extinction of CPP nor relapse, measured as a reinstatement of CPP by a priming injection of cocaine after extinction, were altered in mice without striatal VAChT (Figure 6b, repeated measures ANOVAs show no effect of genotype, F(1,7) = 0.00057, p = 0.982, significant effect of treatment, F(1,7) = 7.457, p = 0.029, and no Genotype × Treatment interaction, F(1,7) = 9.67×10−5, p = 0.992). Therefore, there was no difference in CPP response for the two genotypes.
Behavioural sensitization protocols for cocaine likely reflect altered synaptic plasticity in response to the drug [49], which manifests as an increase in the locomotor effects of cocaine. In a separate group of mice, we measured behavioural sensitization to 10 mg/kg of cocaine (Figure 7) and found that repeated treatment with this dose of cocaine seems to cause slightly higher locomotor activity in VAChTD2-Cre-flox/flox mice, but the relative increase in behavioural sensitization was not different between genotypes (Figure 7a,b, repeated measures ANOVAs show a significant effect of genotype, F(1,16) = 4.902, p = 0.042, significant effect of treatment, F(1,16) = 33.855, p<0.001, and no Genotype × Treatment interaction, F(1,16) = 0.496, p = 0.491). Thus, elimination of striatal ACh release caused a small change in the dose-response profile of cocaine-treated mice in intermediate doses: a slight increase in activity is observed at 10 mg/kg, whereas a decrease in locomotor response is observed at 20 mg/kg in mutant mice.
The balance between acetylcholine-dopamine is important in a number of conditions, including PD; therefore we further investigated dopaminergic function in VAChTD2-Cre-flox/flox mice. For that, we first determined the concentration of dopamine and metabolites in the striatum of VAChTD2-Cre-flox/flox mice and compared these to control mice. In general there were no major changes in dopamine and metabolites in these mutant mice (Table 1). However, the ratio between dopamine and DOPAC as well as dopamine and HVA were significantly changed, showing that dopamine turnover is decreased by 25% (p<0.001), suggesting potential relatively minor alterations in dopamine dynamics or metabolism.
To further assess dopaminergic function, we performed qPCR analysis for D1R and D2R expression in the striatum. We detected an increase in the expression of D1R and D2R mRNAs in the striatum of VAChTD2-Cre-flox/flox mice compared to control mice (Figure 8a and b, D1R, t(12) = 2.756, p<0.05, D2R, t(14) = 2.300, p<0.05). In contrast, D2R mRNA expression in the midbrain was not altered (Figure 8c). G-protein coupled receptors (GPCRs) can have agonist-independent effects; hence, altered expression of such receptors could modulate behaviour even in the absence of neurotransmitter release. We thus also investigated the expression of cholinergic receptors. Figure 8d–f indicates that expression of M1 and M2 muscarinc receptors (mAChR1 and mAChR2) was unchanged, whereas M4 muscarinic receptors (mAChR4) showed increased expression (t(12) = 3.678, p<0.05). Homozygous mice expressing D2-BAC-GFP construct present some dopaminergic phenotypes [50], however control experiments show that the heterozygous D2-Cre mice used here do not present any of the phenotypes associated with selective elimination of striatal VAChT (Figure S6). Because dopamine receptor expression was normal in D2-Cre mice, we conclude that these molecular alterations are due to the loss of ACh release.
To confirm the increased alteration of dopamine receptors in the striatum of VAChTD2-Cre-flox/flox mice we initially performed Western blots. Unfortunately, we were unable to obtain a reliable D1 antibody that showed specific detection of D1R (unpublished data). However, we obtained a D2 antibody that labelled only one major band with the correct molecular mass (Figure 9a). Quantification of immunoblots confirmed increased expression of D2R (Figure 9a,b, t(14) = −3.628, p<0.01). In order to provide an independent measure of D1R activity and test if D1-mediated responses would be altered in VAChT-eliminated mice, we used pharmacological magnetic resonance imaging (phMRI) [51],[52]. phMRI is a variant of functional magnetic resonance imaging that indirectly detects neuronal activity using blood oxygenation level-dependent (BOLD) MRI signal changes [53] to detect functional effects of pharmacological agents in intact systems in vivo with high temporal and spatial resolution. A 9.4T anatomic MRI of the mouse brain (Figure 9c) was used to outline regions of interest in the striatum and cortex. The average difference in BOLD effect between striatum and cortex (Figure 9d) indicates that there is increased neuronal activation in the striatum in VAChTD2-Cre-flox/flox mice following injection of SFK 81297 (3 mg/kg, Figure 9d). The change in the BOLD response after administration of the selective D1R agonist SKF 81297 relative to baseline (prior to injection) was then compared between the two genotypes. Saline administration prior to SKF 81297 did not alter BOLD signal (unpublished data). In contrast, injection of SKF 81297 lead to a slow increase in striatal BOLD response (area under the curve) in VAChTD2-Cre-flox/flox mice compared to control mice following injection of the D1R agonist (Figure 9d and e, p<0.01).
To test if the increased expression/sensitivity of D1R and D2R has direct behavioural consequences, we investigated the effects of the selective dopaminergic agonists SKF 81297 (D1R agonist) and quinpirole (D2R agonist) on locomotor activity. VAChTD2-Cre-flox/flox mice had significantly higher locomotor responses to two doses of SKF 81297 (Figure 10a and b, two-factor ANOVAs revealed significant effect of genotype, F(1,66) = 11.654, p<0.01, significant effect of drug concentration, F(3,66) = 34.476, p<0.001, and significant Drug Concentration × Genotype interaction, F(3,66) = 4.277, p<0.01, Tukey post hoc test showed significant differences with SKF 81297 doses of 3 mg/kg (p<0.01) and 8 mg/kg (p<0.001)). Moreover, VAChTD2-Cre-flox/flox mice also showed enhanced inhibition of locomotion in response to low doses of the D2R-selective agonist quinpirole (Figure 10c and d, ANOVA showed significant effect of genotype, F(1,111) = 12.543, p<0.001, significant effect of drug, F(4,111) = 42.223, p<0.001, but the interaction was not significant, F(4,111) = 2.052, p = 0.092). Analysis of locomotion in response to the individual doses showed a significant difference for 0.005 and for 0.01 mg/kg quinpirole dose (p<0.01). Taken together, these data reveal important alterations in the expression and function of striatal dopamine receptors in VAChTD2-Cre-flox/flox mice.
Here we present a series of evidence that delineates the role of released ACh from those of VGLUT3-dependent glutamate release from striatal cholinergic interneurons. Our data provide a novel perspective on the function of striatal cholinergic neurons suggesting the possibility that they can use distinct neurotransmitters to regulate striatal circuitry. We found that elimination of VAChT in the striatum, without disruption of VGLUT3, did not cause overt disruptions or alterations in several behavioural tasks previously thought to be dependent on striatal ACh release, such as motor learning, sensorimotor gating, and spontaneous locomotor activity. However, we uncovered a novel form of regulation of MSNs by cholinergic tone, and found that selective silencing of striatal ACh release results in an increase in the responses to D1R and D2R agonists. In contrast to the effects of direct dopamine receptor agonists, we found that overall these mice do not show increased locomotor response to cocaine. Similarly, sensitization and rewarding effects of cocaine did not seem to be dependent on striatal release of ACh. Thus, our results significantly depart from previous studies in which the specific contributions of striatal ACh release (mediated by VAChT) were not separated from those of glutamate release (mediated by VGLUT3). These data suggest that VGLUT-3 dependent glutamate release may influence locomotor activity and responses to cocaine considerably more than VAChT-dependent ACh release. Our data suggest that targeted approaches aimed at inhibiting VAChT activity in the striatum may potentially provide a novel strategy to enhance dopaminergic signalling, without causing other major behavioural disturbances.
Our studies in VAChTD2-cre-flox/flox mice indicated that elimination of ACh release in the striatum does not seem to play a major role in motor function and motor learning, at least for acrobatic motor skills in the rotarod test. This observation is also in agreement with previous experiments in striatal cholinergic neuron-ablated mice that presented no deficiency in rotarod performance [31]. However, we cannot completely exclude more subtle effects of ACh in fine motor tuning and motor tasks. For example, the chronic nature of elimination of ACh release in our experiments may lead to adaptations in motor behaviour. Future experiments using VAChTD2-Cre-flox/flox mice and more sophisticated motor behavioural tests may be necessary to pinpoint possible roles for striatal ACh in motor learning and performance.
There are multiple lines of evidence that pharmacological modulation of cholinergic receptors regulates locomotor activity. It is known that muscarinic antagonists increase locomotor activity and M1 and M4 muscarinic receptor KO mice are hyperactive [54]–[57]. Moreover, we have recently observed that mice with a significant decrease in VAChT expression in the whole forebrain show hyperactivity [32]. The present work provides compelling evidence for more selective roles of the neurotransmitter ACh in the striatum, indicating that decreased striatal expression of VAChT does not cause overt motor consequences. These results may be of particular importance, since there have been reports that in Huntington's disease VAChT levels are decreased in the striatum [58]. Our data suggest, however, that this alteration is unlikely to contribute to gross motor symptoms observed in Huntington's disease. Cholinergic neurotransmission in brain regions other than the striatum may still play a role in control of locomotion.
Previous attempts to assess the function of cholinergic neurons in the striatum were performed following the ablation of cholinergic neurons using immunotoxin-mediated cell targeting. Injection of toxin targeting transgenic cholinergic neuron in the accumbens led to an 80% decrease in ChAT-positive neurons [30]. Elimination of cholinergic neurons in the accumbens by this means inhibited certain forms of reward-related learning; however, it also induced hyperactivity and increased sensitivity to the locomotor and the rewarding effects of cocaine, including increased sensitivity in the CPP test to low doses of cocaine [28],[29],[31]. In contrast, recent experiments using an optogenetic approach to inactivate or activate cholinergic neurons in the accumbens found no effects of inactivation of these neurons on locomotor activity, albeit their silencing prevented the response to cocaine in a CPP test [44]. Thus, elimination of cholinergic neurons in the accumbens seemed to increase sensitivity to cocaine-induced CPP [29], whereas optogenetics silencing of these neurons blocked cocaine-induced CPP [44]. The reason for the different outcome in these two experiments is not entirely clear at the moment, but could be related to the chronic versus acute nature of the manipulations. Although in our experiments we have targeted the whole striatum, rather than only the accumbens, we did not detect major alterations in cocaine-induced CPP, suggesting that the above effects obtained with neuronal ablation or by optogenetics manipulation may be linked not to loss of cholinergic transmission per se but rather to suppression of glutamate release from cholinergic neurons.
While an optogenetic approach provides a novel paradigm to acutely activate or inactivate populations of neurons, it is unlikely that this method can separate VAChT from VGLUT3-dependent neurotransmission as selectively as that which can be achieved using VAChTD2-Cre-flox/flox mice. Interestingly, recent data have shown that cholinergic neurons in the habenula secrete both ACh and glutamate (mediated by VGLUT1), and release of either of these neurotransmitters appears to depend on the frequency of stimulation [22]. Basal forebrain neurons in culture release both ACh and glutamate [59]. Importantly, recent work shows that optogenetics stimulation of striatal cholinergic neurons can evoke synaptic glutamatergic neurotransmission onto MSNs, with predominant activity over NMDA receptors [27]. The co-release of glutamate with dopamine has also been described [60],[61], suggesting that interpretation of the roles of dopaminergic neurons will also need to take into account glutamate co-release. Therefore, the co-release of glutamate with classical neurotransmitters may be a more common mechanism than previously appreciated and may have a broad impact in circuitry control. However, we cannot discard the possibility that other neuromodulators released from cholinergic neurons, such as ATP or peptides, could also play a role as co-transmitters.
The role of VGLUT3 in striatal function is far from being fully understood [62]. Interestingly, with respect to striatum-related behaviour, VGLUT3-null mice show hyperactivity and increased response to the locomotor effects of cocaine [25]. Therefore, mice lacking VGLUT3 show a phenotype that is remarkably similar to that of mice in which cholinergic neurons in the accumbens were targeted by an immunotoxin [29],[30]. Experiments in VGLUT3-null mice suggested that the absence of VGLUT3 causes a decrease in striatal cholinergic tone. VGLUT3 is used by the striatal vesicles to facilitate VAChT-mediated ACh storage in synaptic vesicles [25],[62]. However, measurements of ACh release in VGLUT3-null mice have indicated only a modest reduction, by 30% to 40% [25], compared to almost 100% inhibition in VAChTD2-Cre-flox/flox mice. It is unlikely that 40% reduction in ACh release observed in VGLUT3-null mice can be responsible for the hyperactive phenotype. Indeed, independent mouse lines with a 50% decrease in VAChT expression, and concomitant reduction of ACh release [16],[36],[63], did not present increased locomotor activity in the open field [16],[35]. We conclude that the locomotor phenotypes observed previously in striatal cholinergic neuron-ablated mice [29],[31] and in VGLUT3-null mice [25] are either a consequence of the disruption of VGLUT3-mediated neurotransmission or the combination of reducing both glutamatergic and cholinergic activity simultaneously from these neurons. Future experiments using VAChTD2-Cre-flox/flox mice, VGLUT3 floxed mice, and double knockouts will be necessary to provide an assessment of independent effects of VGLUT3-mediated neurotransmission in the striatum.
Although we have focused on striatal-related behaviours, the extent by which alterations in VAChT expression in other brain regions in VAChTD2-cre-flox/flox mice may contribute to these phenotypes should also be taken into account. We did not detect Cre-expression in cholinergic neurons in the penduculopontine area, for example (Figure S2), which harbours groups of cholinergic neurons that project to the midbrain and thalamus and could influence striatal function. However, we cannot completely exclude the possibility that cholinergic neurons in other brain regions would not be targeted in our mouse line. At the same time, as the phenotypes described here seem to be mainly striatal specific and cholinergic interneurons provide the almost exclusive source of cholinergic tone in the striatum, it is unlikely that other groups of cholinergic neurons would have contributed to the observed behaviours.
Elimination of cholinergic neurotransmission in the striatum did not cause hyperlocomotion, however the responses to direct activation of dopamine receptors were substantially increased. Both behavioural and phMRI analysis indicated an increased response to D1R agonist. Western blot analysis also showed selective increase of D2R expression in the striatum. Moreover, in addition to the increased D2R levels in the striatum, which likely reflect a combination of pre- and post-synaptic receptors, we also uncovered increased D2-like receptor pre-synaptic activity, revealed by the increased sensitivity of VAChTD2-Cre-flox/flox mice to low doses of quinpirole. Certainly, we cannot rule out that changes at the level of receptors play a more complex role in regulating locomotor activity in VAChTD2-Cre-flox/flox mice. Indeed, GPCRs may have agonist-independent activity [64],[65]. The locomotor effects of cocaine seem to depend mainly on inhibition of the dopamine transporter [66]. However, acetylcholine can affect release of dopamine via distinct nicotinic receptors [19], as well as regulate both dopamine release and activity of MSNs, via distinct muscarinic receptors [56],[57],[67]. The fact that both D1R and D2R had increased expression in the striatum would suggest that VAChTD2-Cre-flox/flox mice should be more responsive to dopamine and might present increased spontaneous locomotor activity or cocaine-induced locomotion or CPP. However, this was not the case. It is likely that cell-autonomous compensatory mechanisms related to disrupted cholinergic function significantly altered striatal circuitry, preventing such a simple relationship. For example, because M4 muscarinic receptors seem to specifically regulate D1R-mediated signalling [56],[57],[68], it is possible that the increased expression of M4 receptors we detected in the striatum could counterbalance D1R-mediated responses in vivo, leading to unaltered locomotor activity. Moreover, because D2-like pre-synaptic receptors may be more active in VAChTD2-Cre-flox/flox mice, elimination of ACh release in the striatum may also affect pre-synaptic control of dopamine release. The slightly decreased turnover of dopamine in mice without striatal VAChT supports the notion of direct consequences of reduced cholinergic tone at the level of dopaminergic terminals. Thus, behavioural analysis of VAChTD2-Cre-flox/flox mice indicates that control of locomotor function and response to cocaine mediated by dopamine might become more complex in the absence of cholinergic tone. Future experiments will be needed to evaluate direct consequences of elimination of either acetylcholine or glutamate neurotransmission originating from striatal cholinergic neurons on dopamine transmission.
The present data provide direct and indirect evidence that striatal cholinergic neurons can use two different neurotransmitters to regulate striatal function. Hence, re-evaluation of previously attributed functions of striatal cholinergic tone is warranted. The data indicate that VGLUT3-mediated glutamatergic neurotransmission originating from cholinergic neurons may have greater influence on striatal function than previously envisioned. The behavioural consequences of selective elimination of VAChT, and thus cholinergic transmission, in the striatum are remarkably minimal, at least for the locomotion control by the striatal complex. One intriguing phenotype uncovered in mutant mice is an increase in dopamine receptors' expression and function without major alterations in cocaine-induced behaviours. Our experiments provide evidence that targeting VAChT in the striatum can up-regulate dopamine receptors and thus could be used in conditions of dopamine deficiency and abnormally increased cholinergic activity, as found in individuals with PD.
The isolation of a VAChT genomic clone has been described previously [36]. The genomic clone was used to construct a gene-targeting vector in which we added LoxP sequences flanking the VAChT open reading frame and a TK-Neo cassette. Generation of VAChTflox/flox mice is described elsewhere [32], and the construct is shown in Figure S2. Briefly, after removal of the TK-Neo cassette, one LoxP sequence was present 260 bp upstream from the VAChT translational initiation codon, and a second LoxP sequence was located approximately 1.5 kb downstream from the VAChT stop codon and within the second ChAT intron. Note that this vector is distinct from that previously used for generation of VAChT KD mice [36].
D2-Cre mice (Drd2, Line ER44) were obtained from the GENSAT project via the mutant mouse regional resource centers. VAChTD2-Cre-flox/flox mice were generated by crossing VAChTflox/flox with the D2-Cre mouse line. We then inter-crossed VAChTD2-Cre-flox/wt to obtain VAChTD2-Cre-flox/flox mice. Because these mice were apparently normal and fertile, we bred VAChTD2-Cre-flox/flox mice and VAChTflox/flox to obtain all the mice used in the present study. These mice were backcrossed to C57BL/6J mice for five generations. Unless otherwise stated, all control mice used were VAChTflox/flox littermate mice without the Cre transgene.
After the completion of this work we were made aware that the BAC used to generate D2-Cre mice carried an extra gene, ttc2, and a recent report suggests that homozygous D2-GFP mice, generated using the same BAC construct, are hyperactive and show a number of dopamine-related phenotypes [50]. However, as these authors point out, their experiments cannot discern if the phenotypes uncovered are due to the BAC positioning insertion or to the extra copy of ttc2. We confirmed that the D2-Cre indeed have increased expression of the TTC2 mRNA (unpublished data). However, heterozygous D2-Cre mice showed no locomotor phenotype. Moreover, these mice showed normal levels of D1R, D2R, and M4-muscarinic receptors (Figure S6). Hence, neither the phenotypes nor the molecular changes observed in VAChTD2-Cre-flox/flox mice are due to the BAC transgene.
Rosa26-YFP mice (B6.129X1-Gt(ROSA)26Sortm1(EYFP)Cos/J, stock number 006148) were obtained from Jackson Laboratories. Animals were housed in groups of three to four mice per cage without environment enrichment in a temperature-controlled room with 12-h light–12-h dark cycles, and food and water were provided ad libitum. Mouse stocks were SPF, however experimental subjects were kept in a conventional mouse facility.
All studies were conducted in accordance with the NIH and the Canadian Council of Animal Care (CCAC) guidelines for the care and use of animals with approved animal protocol from the Institutional Animal Care and Use Committees at the University of Western Ontario (protocol number 2008–089). Only male mice were used for the behavioural studies, and they were at least 12 weeks old. Mice were randomly assigned to distinct experimental groups. Only mice used for evaluation of spontaneous locomotor behaviour were used in other tasks.
For the immunofluorescence experiments we followed a protocol previously described [35],[69]. For mRNA analyses tissue samples were frozen in a mixture of dry ice/ethanol and kept at −80°C until used as described [70]. Immunoblotting was performed as described elsewhere [36],[71],[72].
Slices were obtained from the striatum and hippocampus of control and test mice, labelled with [3H]methyl-choline, and the release of labelled ACh was determined essentially as described [73] except that 33 mM KCl was used as a depolarizing stimuli.
Striatal synaptosomes were prepared by the method of [74],[75] as previously described [76]. Glutamate release was followed continuously using a fluorimetric method [77] exactly as previously described [78].
All behavioural experiments were performed between 9 a.m. and 4 p.m. in the light cycle, essentially as previously described [16],[35] except the spontaneous activity of the first hours of the dark cycle was done from 7 p.m. to 10 p.m.
The dissected brain tissues were homogenized in 0.2 M perchloric acid with 100 µM EDTA-2Na. Samples were spun in a microcentrifuge at 12,000 rpm for 15 min at 4°C. Samples of the supernatant were then analyzed for norepinephrine (NE), dopamine (DA), and its two metabolites, 3,4-dihydroxyphenylacetic acid (DOPAC) and homovanillic acid (HVA) by NoAb BioDiscoveries (Mississauga, ON).
The HPLC used was an Eicom EP-700 with electrochemical detection (Eicom ECD-700). To elute catecholamines from the reverse phase column (3.0×100 mm SC-3ODS column, Eicom), a mobile phase consisting of 0.1 M citric acetate buffer pH 3.5 with 5 mg/ml EDTA-2Na, 220 mg/L sodium octane sulfonate, and 22% methanol was used.
These experiments have been described in [79]. Briefly, animals were acclimatized 3–5 times to the startle boxes (Med Associates). Habituation of startle was measured using 30 startle pulses (20 ms, white noise, 115 dB on a 65 bd white noise background) with an inter-trial interval of 20 s. Subsequently, prepulse inhibition was measured by displaying 50 startle stimuli with either no prepulse (pulse alone), a 75 dB (4 ms white noise) prepulse preceding the pulse by either 30 ms or 100 ms, or a 85 db prepulse (30 ms or 100 ms interval). Each of the five trial types were displayed 10 times in a pseudorandomized order. PPI is expressed as the average startle response to the respective prepulse trials in relation to the pulse alone trials.
A Grip Strength Meter from Columbus Instruments (Columbus, OH) was used to measure forelimb grip strength essentially as described [36]. For the wire-hang test each mouse was placed on a metal wire-grid, which was slowly inverted and suspended 40 cm above a piece of foam as previously described [36]. The time it took for each mouse to fall from the cage top was recorded with a 60 s cut-off.
The rotarod task followed a previously described protocol [36].
The CPP protocol was modified from [80]. Briefly, CPP was performed in a three chamber apparatus containing two large compartments with differences in visual and tactile cues, separated by a neutral area. In day 1 (habituation), mice were placed in the central compartment and allowed free access to the entire apparatus for 30 min. The time spent in each compartment was recorded. On days 2–7 (the conditioning phase), mice received alternating injections of cocaine or vehicle and were immediately confined into one of the two large compartments for 30 min. A combination of unbiased and biased allocation was used. On day 8 (test day) mice were once again allowed free access to all three compartments for 30 min, and the time spent in each compartment was recorded. For the CPP extinction and reinstatement, the protocol previously described [80] was followed.
Behavioural sensitization was performed as described [57].
The general procedure was previously described, but for analysis we used Anymaze [35].
Mice (VAChTD2-cre/flox/flox, N = 5; control, N = 4) were anesthetized with 4% isofluorane and maintained at 1.5% isoflurane during the MRI scanning. Two intraperitoneal (I.P.) catheters (26 gauge, Abbotcath) were used for injection of saline and SFK. The catheters were secured in place with subcutaneous sutures. Catheters were connected to polyethelyne tubing (PE50, VWR, Canada) and to syringes containing saline and SFK for remote injection during imaging. Mice were placed in a custom built frame designed to secure the skull and minimize respiration induced movement during image acquisition. Mice were imaged on a 9.4 Tesla small animal MRI scanner (Agilent, Palo Alto, CA) equipped with a two-channel surface coil (diameter = 2 cm). A fast low angle shot (FLASH) pulse sequence was used to acquire anatomical images (field of view = 19.2×19.2 mm2, matrix = 128×128, repetition time = 50 ms, echo time = 11 ms, flip angle = 11°, and 10 averages). Respiratory gated lower resolution FLASH images were also acquired for pharmacological imaging (field of view = 19.2×19.2 mm2, data matrix = 64×64, repetition time = 15 ms, echo time = 7 ms, flip angle = 11°, and 1 average) to measure blood oxygen level–dependent (BOLD) signal changes. Seven contiguous axial slices (500 μm thick) covered the brain. Each animal received two injections: first, an injection of 0.5 ml physiological saline (0.9%) administered over a 30 s period (control), and second, SFK 81297 (3 mg/kg), diluted in 0.5 ml physiological saline, also administered over a 30 s period (drug). For the control experiment, images were acquired for 8 min prior to saline injection and then for 20–50 min following injection. For the drug experiment, images were acquired for 8 min prior to drug injection and then for 80–180 min after injection. Throughout the imaging session, body temperature and respiration rate was monitored every 10 min using the MR-Compatible, Model 1025 monitoring system (Small Animal Instruments Inc., Stony Brook, NY). Temperature was maintained at 37.5°C using a warm air blower, and respiration rate ranged from 45–66 (mean 54 BPM). Following imaging, mice were euthanized by cervical dislocation while still under isoflurane anaesthesia.
To limit the influence of global motion on the functional result, the signal intensity difference between striatum and cortex was used to examine the effect of SFK 81297 on the striatum as a function of time. A single slice transecting the striatum was chosen for analysis in each animal (Figure 9c). BOLD signal change was expressed as the percentage change relative to the average baseline signal (first 50 images) prior to drug injection.
Data are expressed as mean ± SEM. Sigmastat 3.1 software was used for statistical analysis. Comparison between two experimental groups was done by Student's t test or Mann-Whitney Rank Sum Test when the data did not follow a normal distribution. When several experimental groups were analyzed, we used two-way analysis of variance (ANOVA). For locomotion experiments we used ANOVA with repeated measures, and when appropriate, a Tukey post hoc comparison test was used. For pharmacological MRI, the area under the curve of the signal time course was compared between VAChTD2-cre/flox/flox mice and control mice using a Student's t test.
|
10.1371/journal.ppat.1006451 | A human endogenous retrovirus-derived gene that can contribute to oncogenesis by activating the ERK pathway and inducing migration and invasion | Endogenous retroviruses are cellular genes of retroviral origin captured by their host during the course of evolution and represent around 8% of the human genome. Although most are defective and transcriptionally silenced, some are still able to generate retroviral-like particles and proteins. Among these, the HERV-K(HML2) family is remarkable since its members have amplified relatively recently and many of them still have full length coding genes. Furthermore, they are induced in cancers, especially in melanoma, breast cancer and germ cell tumours, where viral particles, as well as the envelope protein (Env), can be detected. Here we show that HERV-K(HML2) Env per se has oncogenic properties. Its expression in a non-tumourigenic human breast epithelial cell line induces epithelial to mesenchymal transition (EMT), often associated with tumour aggressiveness and metastasis. In our model, this is typified by key modifications in a set of molecular markers, changes in cell morphology and enhanced cell motility. Remarkably, microarrays performed in 293T cells reveal that HERV-K(HML2) Env is a strong inducer of several transcription factors, namely ETV4, ETV5 and EGR1, which are downstream effectors of the MAPK ERK1/2 and are associated with cellular transformation. We demonstrate that HERV-K(HML2) Env effectively activates the ERK1/2 pathway in our experimental setting and that this activation depends on the Env cytoplasmic tail. In addition, this phenomenon is very specific, being absent with every other retroviral Env tested, except for Jaagsiekte Sheep Retrovirus (JSRV) Env, which is already known to have transforming properties in vivo. Though HERV-K Env is not directly transforming by itself, the newly discovered properties of this protein may contribute to oncogenesis.
| Nearly half the DNA of mammals consists of reitarated, selfish elements that can move and amplify within the genome. With time, some of these elements are recruited by the host and the proteins they encode are used to fulfill physiological functions, whereas other elements have conserved some of their pathological properties and contribute to the development of diseases. The human HERV-K(HML2) elements originated from an ancestral infection of the primate germline by an infectious retrovirus that has been maintained and amplified in the human lineage. It is associated with several pathologies in modern humans, in particular cancer of the breast, germline and skin. We show that the HERV-K(HML2) envelope protein is able to activate a major cellular signalling pathway often involved in human cancers, and that its expression promotes a series of cellular changes that are characteristic of cancer development. Altogether, this study indicates that the expression of HERV-K(HML2) elements is not only a marker of cancer, but can also directly participate to tumourigenesis via the newly discovered oncogenic properties carried by the envelope protein.
| Retroviruses are responsible for a broad range of diseases in animals and humans, the most common of which is the development of cancers. The mechanisms by which they contribute to oncogenesis are diverse and include: (i) insertional mutagenesis, due to activation of cellular proto-oncogenes by inserted proviruses, (ii) immunosuppression, by an immunosuppressive domain conserved in most retroviral envelope proteins and (iii) direct oncogenic activity, with some retroviruses encoding proteins with transforming activities leading to tumour formation. For example JSRV causes the development of contagious lung tumours in sheep [1], and the Env protein alone has been shown to be responsible for the formation of the tumours in vivo [2]. It is also able to transform cell lines [3–8] and induce lung tumour formation in mice [9]. The transforming pathways involved are many, and depend on the direct action of Env itself, as well as the Env-receptor interaction [1].
Endogenous retroviruses (ERVs) are the remnants of past retroviral infections, which have been captured by the host during the course of evolution. They occupy around 8% of the human genome and are similar to the proviral forms of integrated retroviruses from which they derive. Whilst most ERVs are defective and have degenerated over time, others have retained some or all of their open reading frames (ORFs) and can encode potentially pathogenic viral proteins [10–13]. These elements are normally suppressed in healthy tissues but expression has been reported in animal and human cancers [14–17]. The HERV-K (HML2) family (hereafter shortened to HERV-K) is remarkable in that it has recently amplified in humans and many of its ORFs are intact, making it the largest contributor of retroviral-derived proteins in the genome [18]. Expression of the associated proteins and viral particles has been detected in cell lines as well as in human cancers, including melanoma, breast and ovarian cancers [19–22]. In addition, reports indicate that HERV-K expression is important for the transformed phenotype of several cell lines. For example, in melanoma, downregulation of HERV-K Env by siRNA decreases the tumorigenic potential of the A-375 cell line [23] and HERV-K Env expression in the TVMA-12 cell line is necessary for the transition from a adherent to a non-adherent phenotype [24]. In several breast cancer-derived cell lines, HERV-K expression was also recently shown to be important for cell motility and growth, both in vitro and in vivo [25].
In this study, we investigated whether HERV-K Env could have oncogenic properties and be involved in the transformation process of the cells where it is expressed. We demonstrate that its expression induces epithelial to mesenchymal transition (EMT), leading to an increase in cell motility. We also show that HERV-K Env activates the ERK1/2 pathway and promotes the expression of transcription factors involved in oncogenesis.
As HERV-K Env expression has been reported in several human cancers, including breast, we investigated whether it could have a causal role in the transformation process. For this, we used the non-transformed breast epithelial MCF10A cell line, widely used to study the process of oncogenesis [26, 27]. Stable long-term expression of the HERV-K Env was obtained following lentiviral transduction. A control population was generated using an empty vector. After selection, transgene expression in the populations was measured by qRT-PCR using primers located in a region common to both vectors, and were found to express similar levels of lentiviral transcripts. The HERV-K Env population alone was also found to express HERV-K Env transcripts at high but still physiological level (S1 Fig), and as expected, we also detected expression of Rec transcripts that are produced from the Env construct through internal splicing [28, 29]. Interestingly, the HERV-K Env populations displayed an altered morphology (Fig 1A), changing size (Fig 1B) and becoming more elongated. They also lost their typical acinus organisation, and are dispersed in the plate, a phenotype reminiscent of that observed on cells treated with TGFß, a known inducer of EMT. EMT is a process during which cells change their identity, and is important both in normal development and in epithelial cancer progression. It is characterized phenotypically by a loss of cell polarity and cell-cell adhesion, and at the molecular level by a decrease of E-cadherin, an increase of N-cadherin and fibronectin, as well as the induction of a few key transcription factors (mainly Snai1, Snai2, Zeb1, Zeb2) [30]. We quantified these markers in the different populations. As expected, TGFß-treated cells showed an increase of the expression of the mesenchymal markers fibronectin and N-cadherin, as well as EMT-associated transcription factors Snai1&2 and to a lesser extent Zeb1 (Fig 1C). Interestingly, HERV-K Env, but not the control, also modified the expression of EMT-associated genes, with a significant increase in the levels of fibronectin and N-cadherin, and a decrease in E-cadherin. However, the most induced transcription factor was Zeb2, unlike in TGFß treated cells. We also tested the different populations for changes in their motility properties using transwell assays. As shown in Fig 1D, the HERV-K Env expressing populations displayed an increase in cell migration and invasion, similar to what is observed with TGFß-treated cells (Fig 1E). Altogether the data obtained with the MCF10A cells indicate that HERV-K Env possesses oncogenic properties, and modifies the cell physiology to induce a process related to, but not identical to TGFß-induced EMT.
To further characterise HERV-K Env effects on cell physiology, we used the embryonic kidney 293T cell line. We searched for genes whose expression is modified by HERV-K Env using a non-biased transcriptomic approach. 293T cells were transfected in duplicates with a vector expressing the HERV-K Env or a control plasmid, under conditions adjusted to minimise cell toxicity due to the transfection (Fig 2A). Of note, the two expression vectors are identical except for nt 6759 and 6762 that introduce premature stop codons in the control vector, leading to the production of a much shorter protein (102 aa instead of 699), in order to rule out any effect due to the Env RNA per se. The transcriptomes were compared at 24 and 48 hours post transfection using whole-genome microarrays. Genes with statistically significant changes in expression (≥2-fold) were further analysed. At 24 hours post-transfection, no significant difference between the two conditions was observed, probably due to very low protein expression levels at this time point. At 48 hours, we found 86 genes with modified expression levels (69 down-regulated, 17 up-regulated). After checking, the down-regulated genes proved to be genes induced by the transfection itself. The expression level of these genes increased between 24h and 48h after transfection, but to a lower extent in the HERV-K Env cells than in the control, leading to a seemingly downregulation of these genes by HERV_K Env when only the 48h expression data are taken into account. These genes were not investigated further.
Amongst the identified upregulated genes (Fig 2B), there was a strong excess of transcription factors (7 out of 17, with five in the top six most upregulated genes). Interestingly, they include EGR1, ETV4, ETV5 and FosB that have been associated with EMT and tumour aggressiveness in several cancers [31–35]. We confirmed the induction of the top five genes from the list in the same samples by qRT-PCR (Fig 2B). We noted some variations in the measurements between microarray and qRT-PCR data, but this is likely due to the specific sequences of the primers/probes used in each technique. We also verified that the induction of transcription factors was specific and not observed with another retroviral Env protein using an expression vector for the Amphotropic Murine Leukemia Virus Env as a control in an independent series of experiments. As shown in Fig 2C, qRT-PCR confirmed the increased expression of the four transcription factors (EGR1, ZCCHC12, ETV4 and ETV5) by HERV-K Env.
We noticed that the majority of genes activated by HERV-K Env are involved in the ERK1/2 MAPK pathway (Fig 3A). Indeed, EGR1, ETV4 and ETV5 are direct targets of ERK1/2. DUSP6, which we found induced at a lower level, is a secondary target involved in the negative retro-control of the pathway. This strongly suggested that HERV-K Env is an inducer of the ERK1/2 pathway. We tested this hypothesis by assessing the phosphorylation of ERK1/2 following transient transfection of 293T cells (Fig 3B). As shown, cells transfected with HERV-K Env show a marked phosphorylation of ERK1/2, not seen with the controls, while total amounts of ERK1/2 are similar with all plasmids.
To investigate the specificity of this activation, we transfected a panel of retroviral Envs, encompassing several genera, in 293T cells and measured the expression of EGR1, ETV4 and ETV5, as well as the phosphorylation of ERK1/2. As shown in Fig 4A & 4B, no other retroviral Env was able to induce both ERK1/2 phosphorylation and transcription factor expression, except JSRV Env. Interestingly, JSRV Env’s ability to activate ERK1/2 has been described in other cellular models and has been linked to its strong oncogenic effect [36–38]. This similarity suggests that HERV-K Env possesses oncogenic properties as well. Of note, the closely related MMTV Env was unable to activate the ERK1/2 pathway, indicating that the ability to activate the ERK1/2 pathway is not a general property of betaretroviral Envs. Finally, EGR1, but not ETV4 and ETV5, was also induced by the deltaretroviral HTLV-1 Env, consistent with previous data reporting the transactivation of EGR1 promoter by the accessory protein Tax [39] which is also produced from the HTLV-1 Env expression vector. Accordingly, HTLV-1 Env does not induce the phosphorylation of ERK1/2 (Fig 4A).
Due to the similarities with JSRV Env, we tested if expression of HERV-K Env was directly transforming in a classical transformation assay using rat 208F cells. [40, 41] HERV-K Env was unable to induce the formation of transformed foci in this assay, but an otherwise fully infectious functional endogenous allele of JSRV (enJSRV-18) was also negative in this assay (Fig 4C). Furthermore, unlike HERV-K Env, the endogenous JSRV Env was also unable to activate ERK1/2 or induce expression of the transcription factors (Fig 4D+4E). Functional expression of all Env constructs was confirmed by assessing by their ability to produce infectious pseudotyped viral particles (Fig 4E).
All the experiments described above have been performed with a consensus HERV-K Env, which theoretically corresponds to the protein in the progenitor virus responsible for the insertion of all modern human-specific HERV-K proviruses [42]. In order to assess the relevance of our observations in modern-day people, we tested the effect of six previously described natural “alleles” of HERV-K Env present in humans on ERK1/2 activation (Fig 5A). Five out of these six, namely K108, K109, K113, K115, K17833 show some transcription factor induction activity (Fig 5B). Among these, the three that correspond to the most functional HERV-K Env proteins when tested for other classical virological properties (Fig 5C) also induce significant ERK1/2 phosphorylation in our assay, suggesting that the ability of the endogenous Envs to activate the signalling pathway is linked to the canonical properties of the infectious retrovirus.
Two proteins are produced from the HERV-K Env-Rec plasmid: the envelope glycoprotein and the accessory protein Rec. The latter consists of two exons and is completely contained within the Env ORF. We wanted to ascertain which of these two proteins mediates the effects seen earlier. We therefore used two previously described vectors [43]: one expressing only Rec, and the other HERV-K Env without Rec (through silent point mutations targeting the splicing sites) (Fig 6A). Expression of the Rec protein did not induce expression of the transcription factors or phosphorylation of ERK1/2, whereas the Env alone was as efficient as the consensus Env-Rec construct to activate signal transduction (Fig 6B & 6C).
We then set about mapping the domains in HERV-K Env required for ERK1/2 activation. Like other retroviral Envs, it is processed by cellular proteases into two subunits, SU, expressed at the cell surface, and TM, containing a single-pass transmembrane domain (Fig 6D). Using previously described mutants [43] (see Fig 6D), we found that the soluble SU (mut 4) completely lost the ability to induce the expression of the transcription factors and the phosphorylation of ERK1/2 (Fig 6E & 6F), indicating that the TM moiety of Env is required. HERV-K Env deleted of its cytoplasmic tail (mut 1) showed some activity, but was markedly decreased. The cytoplasmic tail is therefore important for the activation of the ERK1/2 pathway, either directly or through modification of the intracellular localisation of HERV-K Env, but other domains are also involved.
We also assayed HERV-K Env for potential effects on other cell signalling pathways. We first tested for activation of NFκB, which is often implicated in transformation, using the degradation of IκBα as a sign of activation of the pathway. None of the Envs that we tested had any effect on IκBα levels, unlike TNFα stimulation (Fig 7). HERV-K Env therefore does not modify cell physiology through the NFκB pathway. We then tested for p38 activation (Fig 7). As previously reported, JSRV Env induces the phosphorylation of p38 [36]. The role of p38 activation in JSRV pathogenesis is not clear, but it has been shown to have an inhibitory effect on ERK1/2 signalling. We found that HERV-K Env also activates p38, but our microarray data indicates that ERK1/2 is the major signalling pathway activated by HERV-K Env. It is possible that the activation of p38 by JSRV and HERV-K Envs is a mechanism to regulate the level of ERK1/2 activation. Finally, we tested the PI3K/AKT pathway, previously shown to be important for JSRV Env mediated transformation in several cellular models [6, 8, 40, 41, 44, 45]. Preliminary experiments showed that AKT is constitutively phosphorylated in 293T cells, whatever the culture conditions. We therefore tested for AKT activation in HeLa cells. As shown, HERV-K Env had no effect on the phosphorylation state of AKT, unlike JSRV Env (Fig 7).
Finally, using a series of inhibitors (Fig 8A), we characterised where HERV-K Env acts in the ERK1/2 pathway. 293T cells were transfected as before, treated with each inhibitor individually 18h later and ERK1/2 phosphorylation and expression of the transcription factors were measured at 48h (Fig 8B). First, we used FTI-277 that targets H and K-Ras. As expected, this inhibitor efficiently suppressed the phosphorylation of ERK1/2, with a corresponding decrease in transcription factor expression induced by the transfection of a constitutively active H-Ras (Fig 8C & 8D). However, it did not affect the activation of ERK1/2 mediated by HERV-K Env or JSRV Env. This suggests that either these Envs act downstream of Ras or that they activate another form of Ras (e.g. N-Ras). In contrast, TAK632, a potent pan-Raf inhibitor, completely abolished ERK1/2 signal transduction and transcription factor induction by both JSRV and HERV-K Envs, indicating that the two glycoproteins act upstream of the kinase Raf. U0126, a MEK1/2 inhibitor, also impaired the activation of ERK1/2 mediated by HERV-K and JSRV Envs, as expected. Of note, unlike JSRV Env, the inhibition was only partial for HERV-K Env.
In this paper, we report on the pro-oncogenic properties possessed by the Env protein of the HERV-K family and investigate the mechanism of action. Using the MCF10A cell line, we demonstrated that the stable expression of HERV-K Env, at a high but physiologically relevant level (i.e. similar to the expression level naturally observed in some germ cell derived cell lines), induces clear changes in the expression of EMT-associated genes towards a more mesenchymal phenotype, with the cell morphology altered accordingly. Remarkably, these HERV-K Env-induced changes are accompanied by an increase in cell motility. The modification of the attachment proteins we observed is similar to that obtained after treatment with TGFß, but the transcription factors that are induced are different. A number of different transcription factors have been associated with EMT, and this generic term in fact covers several processes that can occur in different circumstances, either in normal development or in the course of disease [30]. Previous studies had already hinted at oncogenic properties for HERV-K Env, but it had only been shown that this Env protein can alter the phenotype of pre-transformed, cancer-derived cell lines [23–25] whereas we demonstrate here that it can also direct non-malignant cells in the path towards transformation. Given the change of the EMT markers and cell motility observed, it is likely that HERV-K Env expression by a tumour or a pre-tumour could also trigger further changes and favour metastasis.
Using microarrays in a different cell model, we identified a very limited number of genes whose expression is induced following HERV-K Env transfection. We showed that the induction of these genes is due to activation of the ERK1/2 pathway in cells following HERV-K Env expression. Most of the identified genes are transcription factors that have already been associated with transformation and cancer. In fact, the list of the induced genes is remarkably similar to that observed in tumours with a mutated BRAF [46]. BRAF-activating mutations have been reported in several tumours, but are particularly common in melanomas [47]. Specific inhibitors of mutated BRAF have been developed and used to treat patients. They promote a dramatic improvement in patient health for up to 6 months until relapse, due to tumours developing resistance to the treatment [48]. HERV-K Env expression in melanomas has been reported by several independent groups [19, 20, 49–52]. It is possible that HERV-K Env expression is part of the mechanism used by the tumours to escape the treatment against mutated BRAF by re-activating the ERK1/2 pathway. Interestingly, a study recently showed that HERV-K expression in several breast cancer cell lines leads to an increase expression of ERK1/2 [25], which indicates that HERV-K expression could increase ERK1/2 signalling by several independent mechanisms.
The consensus HERV-K Env that we used to demonstrate ERK1/2 activation is more efficient than present-day alleles for most canonical virological properties. However, when tested under the same conditions, three out of the six previously characterised full-length “alleles” of HERV-K Env activated ERK1/2 nearly as efficiently as the consensus, giving credence to a possible role of HERV-K proviruses in tumour development. Of interest, these active alleles have been found to be spontaneously expressed in several human cell lines [52, 53]. In addition, the effect of these alleles could be additive, especially when HERV-K proviruses are expressed due to general LTR activation (eg the reported induction by MITF in melanoma [54]), instead of locus-specific expression. Remarkably, no other retroviral Env protein that we tested activated ERK1/2, except for JSRV Env which is known to be a strong oncogene. Furthermore, the oncogenic activity of JSRV Env has been linked to ERK1/2 activation [36, 38], further supporting a role for HERV-K Env in tumour development. The effects observed with JSRV Env are admittedly stronger than those observed with HERV-K Env. However, it should be noted that the JSRV Env used in this study is that encoded by the infectious strain of the virus. Like HERV-K Env, the endogenous JSRV-18 Env was unable to transform 208F cells in our assays and no transforming effect has ever been reported for any of the endogenous copies [1]. It is not surprising that a gene showing such deleterious effects to the host should be lost very quickly following endogenisation. It is quite remarkable that the HERV-K family, which entered the primate lineage more than 40 million years ago, could have conserved some oncogenic properties for so long, even if these properties are slightly subdued. Indeed we showed that a recent, fully infectious endogenous JSRV Env protein has completely lost the ability to activate the ERK1/2 pathway, unlike some HERV-K Env alleles. It is possible that the last amplification of HERV-K elements in the human genomes is the product of a horizontal transmission of an infectious virus that would have remained active in other primate species, instead of the re-activation of previously dormant proviruses. It would be less unexpected for an exogenous, infectious virus to conserve such oncogenic properties that could play an important role in its propagation.
Finally, we used mutant forms of HERV-K Env to map the domains responsible for ERK1/2 activation. We found that the TM subunit of Env is required, and that the cytoplasmic tail plays an important role, although there is still some ERK1/2 activation when it is deleted. Additionally, the 3 endogenous alleles of HERV-K Env-Rec that most strongly activate ERK1/2 all possess an intact cytoplasmic tail (S2 Fig). However, the K115 allele possesses the same sequence and is much less potent for ERK1/2 activation. It is therefore likely that several domains of HERV-K Env cooperate for its oncogenic properties, as demonstrated for JSRV Env for which both SU and TM are involved in the transformation process [55]. Concerning JSRV Env, several intracellular interacting proteins expected to be involved in transformation have been reported recently [56, 57]. In all cases, the reported interacting domain is the Env cytoplasmic tail. Since JSRV and HERV-K Envs show no homology in this region (S3 Fig), these proteins are unlikely to interact with HERV-K Env. Using motif prediction software, we failed to identify any relevant motif in the sequence of the HERV-K Env cytoplasmic tail, and cannot therefore propose a likely mechanism of action. In the future, the elucidation of the cellular proteins interacting with the HERV-K Env cytoplasmic tail could lead to the development of specific inhibitors able to block its oncogenic properties, and could be of therapeutic interest in tumours where HERV-K Env is expressed, providing alternative treatments to those currently available.
HIV-1 derived particles were produced as described [43], using a modified CSGW [58] expressing HERV-K Env or control proteins and the hygromycin resistance gene. pBabe-H-RasG12V [59] was kindly gifted by A. Puisieux. All Env transient expression vectors are CMV-driven. Plasmids for Ampho, IAPE, RD114, GaLV, FeLV and HERV-K Envs have been described previously [60–64]. phCMV-JSRV Env was constructed by replacing the G protein ORF in phCMV-VSV-G (GenBank accession no. AJ318514) with the JSRV Env ORF (plus the 3’ LTR) present in pCMV3JS21ΔGP [3] (a gift from M. Palmarini). The Env gene from the enJSRV-18 provirus was similarly cloned by PCR on sheep genomic DNA using a forward primer on the ATG and the reverse primer indicated in [65]. phCMV-MMTV Env contains the MMTV Env ORF and the 3’ LTR from the pEnv vector [66] (a gift from S. Ross). The HTLV-1 Env plasmid [67] was a gift from C. Pique. phCMV-HERV-K(HML2) Env-LP was derived from phCMV-HERV-K(HML2) Env by changing aa 103 and 104 into stop codons. pCMV-ß (Clontech) is an expression vector for beta-galactosidase. It was used as a control vector and designated “None” in the figures. It was also used to adjust total DNA content in transfection experiments.
293T (ATCC CRL-3216), HeLa (ATCC CCL-2) and 208F (ECACC 85103116) cells were maintained at 37°C, 8% CO2, in DMEM with 10% heat-inactivated FCS, 100u/mL penicillin and 100μg/mL streptomycin (PS). MCF10A cells (ATCC CRL-10317) were cultured at 37°C, 10% CO2, in DMEM:F12 medium supplemented with 5% horse serum, 5ng/mL EGF (Peprotech), 10μg/mL insulin (Sigma), 1ng/mL cholera toxin (Sigma), 100μg/mL hydrocortisone (Sigma) and PS. Unless specified all reagents were from Life Technology. Lentiviral particles were produced as described [43] using JetPrime (PolyPlus Transfection). Cells were infected with viral supernatants and selected with Hygromycin B (46u/mL, Calbiochem) 3 days later. The populations of resistant cells were thereafter maintained in selection media.
Cells in 12-well plates were transfected with 250ng total DNA (50 or 30ng of Env in 293T and HeLa cells respectively, supplemented with pCMV-ß) using 1.25μL of Fugene6 (Promega). Media were replaced 18 hours post transfection (without FCS for Hela cells to minimize background). When used, inhibitors were added during the medium change (FTI-277 (5μM, Sigma), TAK-632 (5μM, Selleckchem), U0126 (5μM, Cell Signaling)). TNFα was used at 100ng/μL (R&D Systems). Cells were used for protein or RNA extraction 48 hours post transfection. For Western blot analysis, cells were lysed in PBS, 1% NP40 or RIPA (Life Technology) complemented with Halt protease and phosphatase inhibitor cocktail (ThermoScientific). Cell lysates were then subjected to SDS-PAGE as described [43]. Proteins of interest were detected using antibodies from Cell Signalling: p44/42 MAPK, phospho-p44/42 MAPK (Thr202/Tyr204), p38 MAPK, phospho-p38 MAPK (Thr180/Tyr182), pan-Akt, phospho-Akt, IKBα or Sigma (Tubulin). HERV-K(HML2) anti-Env antibody was previously described [61]. MLV ampho Env protein was detected using a goat antiserum directed against Rauscher leukemia virus gp70 (from the National Cancer Institute, Frederick, MD). HERV-K Rec protein was detected using a polyclonal rabbit antiserum given by R. Löwer. HRP-conjugated secondary antibodies (GE Healthcare Or Dako) and ECL Plus Reagent (GE Healthcare) were used for Western blots. Membrane stripping was done using ReBlot Plus Strong (Merck Millipore).
Total RNA were extracted with the RNeasy extraction kit (Qiagen) and treated with DNase I (Ambion). For microarray experiments, we compared duplicates of RNA from 293T transfected with either HERV-K(HML2) Env or the control plasmid, HERV-K(HML2) Env-LP, collected 24 and 48 hours post-transfection. Gene expression analysis was performed on Agilent SurePrint G3 Human GE 8x60K Microarrays (Agilent Technologies, AMADID 39494). Data were extracted using Feature Extraction software (v10.5.1.1; Agilent Technologies) and normalized using an empirical Bayes method. Top-ranked genes were selected for an absolute fold-change ≥2 using a False Detection Rate (FDR) <0.05.
DNase-treated RNAs were reverse-transcribed using the MLV reverse-transcriptase (Applied Biosystems). qPCR was performed using the QuantiFast SYBR Green PCR kit (Qiagen) on the ABI PRISM 7000 system. Efficacy of the PCR reaction was checked for each primer pair. Transcript levels were normalized to RPLO employing the ΔΔCt method.
MCF10A cells were resuspended in DMEM:F12 media without serum or additional additives. 5x104 cells in 500μL were seeded into the top of each transwell (Corning, 24-well inserts, 8μm pore), and 750μL of complete culture medium was added to the bottom of each well as a chemoattractant.
The cells were incubated for 22 hours before non-migratory cells were removed and the membrane fixed with methanol. Membranes were stained with DAPI and the migrated cells counted. Invasion assays were performed similarly except transwells were precoated with 16μg of Matrigel (Corning).
MCF10A cell populations were incubated in DMEM/F12 media supplemented with 5μM CellTracker Green (Invitrogen) for 30 minutes, and then cultured for an additional 60 minutes in complete media before fixation in 4% PFA. Cells were imaged to assess changes in morphology. Average cell size (area) was calculated by measuring the surface area covered by the cells (green stain) relative to the number of cells.
208F cells (seeded at 2x105 cells per 3.5 cm dish the day before) were transfected with 4μg DNA using Fugene 6 (Promega). After 24h, they were reseeded in a 10 cm dish. When the cells reached confluence, medium was replaced by DMEM complemented with PS, 5% FCS and 1μM dexamethasone. Medium was replaced weekly. After 3–4 weeks, cells were stained with Leishmann to allow counting of transformed foci.
|
10.1371/journal.pmed.1002532 | Cost-effectiveness of multidisciplinary care in mild to moderate chronic kidney disease in the United States: A modeling study | Multidisciplinary care (MDC) programs have been proposed as a way to alleviate the cost and morbidity associated with chronic kidney disease (CKD) in the US.
We assessed the cost-effectiveness of a theoretical Medicare-based MDC program for CKD compared to usual CKD care in Medicare beneficiaries with stage 3 and 4 CKD between 45 and 84 years old in the US. The program used nephrologists, advanced practitioners, educators, dieticians, and social workers. From Medicare claims and published literature, we developed a novel deterministic Markov model for CKD progression and calibrated it to long-term risks of mortality and progression to end-stage renal disease. We then used the model to project accrued discounted costs and quality-adjusted life years (QALYs) over patients’ remaining lifetime. We estimated the incremental cost-effectiveness ratio (ICER) of MDC, or the cost of the intervention per QALY gained. MDC added 0.23 (95% CI: 0.08, 0.42) QALYs over usual care, costing $51,285 per QALY gained (net monetary benefit of $23,100 at a threshold of $150,000 per QALY gained; 95% CI: $6,252, $44,323). In all subpopulations analyzed, ICERs ranged from $42,663 to $72,432 per QALY gained. MDC was generally more cost-effective in patients with higher urine albumin excretion. Although ICERs were higher in younger patients, MDC could yield greater improvements in health in younger than older patients. MDC remained cost-effective when we decreased its effectiveness to 25% of the base case or increased the cost 5-fold. The program costed less than $70,000 per QALY in 95% of probabilistic sensitivity analyses and less than $87,500 per QALY in 99% of analyses. Limitations of our study include its theoretical nature and being less generalizable to populations at low risk for progression to ESRD. We did not study the potential impact of MDC on hospitalization (cardiovascular or other).
Our model estimates that a Medicare-funded MDC program could reduce the need for dialysis, prolong life expectancy, and meet conventional cost-effectiveness thresholds in middle-aged to elderly patients with mild to moderate CKD.
| Chronic kidney disease is a major cause of morbidity and mortality in the US.
Multidisciplinary care—when healthcare providers of different expertise collaborate to treat a single disease—has successfully reduced mortality and the incidence of end-stage renal disease in patients with chronic kidney disease.
Understanding the economic impact of multidisciplinary care in chronic kidney disease could help policy makers decide whether such a program improves health outcomes in a cost-effective way.
We developed a novel Markov model that accurately simulates the progression of chronic kidney disease to end-stage renal disease and accounts for heterogeneity in patients with chronic kidney disease.
Multidisciplinary care was cost-effective in most patients with chronic kidney disease and was more cost-effective in patients with higher levels of albuminuria.
Multidisciplinary care remained cost-effective in the model even if it was substantially less effective or more costly than base case estimates.
Multidisciplinary care could improve the health of patients with chronic kidney disease at reasonable value for money.
Policy makers could consider implementing pilot multidisciplinary care programs to formally test their effectiveness and cost-effectiveness.
| Chronic kidney disease (CKD) affects approximately 10% of Medicare beneficiaries in the US but accounts for a disproportionate 20% of expenditures [1]. Patients with end-stage renal disease (ESRD) are more costly, representing 1.6% of Medicare beneficiaries and responsible for 7.2% of costs [1]. At the same time, life expectancy is substantially lower in patients with CKD than in the general population [1–3].
Multidisciplinary care (MDC) has been proposed as a way to mitigate the high costs and mortality associated with CKD. It has led to successful outcomes in other settings, including heart failure [4], intensive care [5], and cancer [6]. In CKD, researchers have investigated a variety of strategies, including nurse and advanced practitioner coordinated models [7–12], use of dieticians and social workers [7–9,12,13], and education programs [14–19]. Several systematic reviews have shown that MDC slows CKD progression [20], delays the onset of dialysis, and decreases mortality [21].
Little is known about the cost-effectiveness of MDC in a US CKD population. Although prior studies have suggested that MDC is cost-effective, these studies were performed in other countries and did not use validated models for CKD progression [22,23]. Developing an accurate model is challenging because CKD progression is associated with mortality, and previous studies did not account for this relationship. Furthermore, these studies did not consider heterogeneity in CKD. Many patients with mild to moderate CKD do not progress to ESRD and may not benefit from an intensive disease management program [24]. Determining the subgroups that benefit the most from MDC may help providers more effectively treat vulnerable patients with CKD.
In this study, we performed a cost-effectiveness analysis of a theoretical Medicare MDC program for US populations of differing CKD severity. We did so after developing a novel CKD progression model that incorporates disease heterogeneity and mortality risk. To account for inefficiencies in a nationally funded and broadly applied MDC program, we also tested if more expensive and less effective programs remained cost-effective. We hypothesized that MDC is more cost-effective in patients with more severe CKD. We also hypothesized that even an inefficiently deployed program would be cost-effective by conventional thresholds.
To model the cost-effectiveness of MDC in CKD, our analysis involved 3 elements: (1) we constructed and calibrated a CKD progression model; (2) we modeled the cost-effectiveness of MDC; and (3) we performed multiple sensitivity analyses. Because the effectiveness of MDC varies for different types of patients, we performed these analyses in US patients of different ages (45–64, 65–74, and 75–84 years old), sexes (female and male), races (white, black, and other), estimated glomerular filtration rates (eGFRs ranging from 20 to 59 ml/min/1.73 m2 in 5-ml/min/1.73 m2 increments), and approximate albuminuria levels (urine albumin to creatinine ratio [UACR] 1, 300, 1,000, and 3,000 mg/g).
For our CKD progression model, we used a deterministic Markov model that simulates the disease course of patients with CKD as they experience progressive CKD, ESRD, and death. The model accounts for population-level heterogeneity, including differences in age, sex, and race as well as eGFR and albuminuria. For instance, patients with lower eGFRs and patients with higher levels of albuminuria are more likely to develop ESRD in our model. Similarly, older patients have higher mortality rates in our model. To ensure that our model accurately simulated CKD progression and mortality, we calibrated it to long-term mortality rates and ESRD incidence rates, which we obtained from published literature. We show results from our calibration procedure in S2 Appendix to demonstrate that our progression model accurately reflects published literature for different subpopulations with CKD.
To estimate the cost-effectiveness of a theoretical Medicare-funded MDC program in the US, we used our CKD progression model to simulate the total lifetime costs and outcomes (as measured by quality-adjusted life years [QALYs]) of medical care for patients under MDC versus under usual care. After discounting total accrued costs and QALYs by an annual rate of 3%, we computed the incremental cost-effectiveness ratio (ICER), the difference in discounted costs divided by the difference in discounted QALYs. Conceptually, the ICER is the number of dollars that Medicare must spend on MDC to gain 1 QALY. We assumed that a cost-effective program would have an ICER of no more than $150,000 per QALY, which corresponds roughly to the 2017 inflation-adjusted ICER for dialysis ($129,000 per QALY in 2009) [25].
In the literature, MDC spans different interventions and has been used in patients of varying disease severities. The aims of MDC programs have ranged from slowing the progression of CKD to ESRD, to preventing cardiovascular events such as myocardial infarction and stroke, to optimizing decision-making at the onset of ESRD. We limited our study to MDC programs that aimed to slow the progression of CKD to ESRD. We did not incorporate the reduction of cardiovascular hospitalizations or prevention of acute kidney injury into our model because it is not clear from the literature whether MDC is effective in preventing these outcomes. It is important to note that our study assessed reductions in all-cause mortality and thus implicitly assessed reductions in cardiovascular mortality. However, given the absence of data on the effect of MDC on intermediate cardiovascular endpoints, we were unable to incorporate cardiovascular or other hospitalizations into our cost-effectiveness estimates directly.
Additionally, the literature often conflates MDC programs aimed at slowing the progression of CKD with MDC programs aimed at optimizing dialysis planning through reducing the use of tunneled dialysis catheters and improving the use of home dialysis. However, these programs are operationally distinct and are meant for different populations: the former are typically used in patients with stage 3 and 4 CKD, while the latter are reserved for patients on the cusp of needing dialysis, with late stage 4 or early stage 5 CKD. Our interest was in earlier interventions focused on slowing the progression of CKD to ESRD in patients with stage 3 and 4 CKD.
These MDC programs typically comprise periodic visits with nephrologists, advanced practitioners, educators, dieticians, and social workers, which ramp up in frequency as CKD becomes more severe. We assumed that MDC increased the use of medications and laboratory tests routinely used to manage anemia and derangements in bone-mineral metabolism, common in patients with CKD.
Using a recently published systematic review, we constructed a model that estimated the effectiveness of MDC in reducing mortality and in preventing ESRD [21]. This review, and the published literature in general, provided an estimate of the average effectiveness of MDC across many severities of CKD, without accounting for population-level heterogeneity. To incorporate the clinical intuition that many patients with earlier stages of CKD do not progress to ESRD and thus may not benefit from an MDC program, we assumed that MDC was less effective in patients with less severe CKD.
Because a Medicare-funded MDC program would likely have a large range of effectiveness and cost, we performed a wide array of sensitivity analyses varying our assumptions. We did so by repeating our analyses under different scenarios, where MDC was down to only 25% as effective as the base case or up to 5 times more expensive. We also simulated scenarios where MDC did not have any effect on CKD progression. For each scenario, we performed a probabilistic sensitivity analysis, which allowed us to produce cost-effectiveness acceptability curves and 95% confidence intervals.
Below, we describe the technical specifications of our modeling technique.
We used Matlab version R2016b (MathWorks, Natlick, MA) for modeling. We used Stata version 14.1 (StataCorp, College Station, TX) and SAS version 9.4 (SAS Institute, Cary, NC) to estimate costs and target probabilities.
The Stanford University Institutional Review Board approved of this study and the use of the USRDS database (IRB-17804).
All work using the USRDS was conducted under a data use agreement between Dr. Tara Chang and the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). An NIDDK officer reviewed the manuscript and approved it for submission.
Our study conformed to the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) guidelines (S1 Text) [56].
We found that MDC improved health by 0.23 (95% CI: 0.08, 0.42) QALYs per person over usual care, from 1.78 to 2.01 QALYs (Tables 6–9). This translated to hazard ratios of 0.77 (95% CI: 0.63, 0.92) for death and 0.62 (95% CI: 0.43, 0.86) for progression to ESRD (Tables 10–13). The improvement in outcomes was more pronounced for younger patients, with 45–64 year olds seeing a higher increase in QALYs (95% CI: 0.31–0.76) than 75–84 year olds (95% CI: 0.09–0.31). Across different races, sexes, and eGFRs, the gain in health was similar. Patients with higher levels of albuminuria generally gained fewer QALYs than those with lower levels of albuminuria. For example, patients with a UACR of approximately 3,000 mg/g gained fewer QALYs (95% CI: 0.09–0.33) than patients with no albuminuria (UACR approximately 1 mg/g) (95% CI: 0.22–0.76). However, relative gains in QALYs were similar across different levels of albuminuria (a 9% to 33% increase versus a 10% to 19% increase, respectively).
On average, MDC cost $12,001 (95% CI: $5,098, $19,358) more than usual care per patient over the lifetime, an increase from $68,571 to $80,572 (Tables 6–9). The corresponding ICER was $51,285 per QALY gained, which was cost-effective at a threshold of $150,000 per QALY (net monetary benefit of $23,100; 95% CI: $6,252, $44,323). The ICERs ranged from $42,663 to $72,432 per QALY gained in all subgroups. In general, MDC was more expensive in patients with lower levels of albuminuria. For instance, patients with a UACR of 1 mg/g had ICERs of $55,315 to $57,958 per QALY gained versus $48,323 to $50,916 per QALY gained in patients with a UACR of 1,000 mg/g. Similarly, younger patients tended to have higher ICERs, though these higher relative expenses brought forth greater improvements in health.
MDC remained cost-effective at the WTP threshold of $150,000 per QALY in most sensitivity analyses where we varied its effectiveness (Tables 14 and S9–S11). Even when the program was only 25% as effective as the base case, MDC was cost-effective on average in the entire population, with an ICER of $127,927 per QALY gained, though the 95% confidence interval for net monetary benefit did cross $0 (net monetary benefit of $1,076; 95% CI: −$2,194, $4,088). This corresponded to hazard ratios of 0.95 for death and 0.91 for progression to ESRD. Notably, MDC was not cost-effective when operating at 25% effectiveness in patients with an eGFR of 59 ml/min/1.73 m2 and UACR of 1 mg/g (ICER $151,869 per QALY) (S9 Table). When we assumed that MDC did not attenuate progression to ESRD and was only 25% as effective as the base case in preventing death (hazard ratios of 0.93–0.97 for death and 1.00–1.02 for progression to ESRD), the program was not cost-effective in all subgroups.
In all scenarios that varied the cost, we found that point estimates for the ICERs were less than $150,000 per QALY gained (Fig 2). MDC remained cost-effective even when we increased the monthly cost 5-fold, although the upper end of the 95% confidence interval exceeded our threshold for cost-effectiveness. Additionally, MDC remained cost-effective when it increased the use of medications and laboratory tests in 100% of the population. In this case, the upper end of the 95% confidence interval for patients without albuminuria exceeded $150,000 per QALY gained. For patients with higher levels of albuminuria, the 95% confidence interval remained less than $150,000 per QALY gained. In all cost sensitivity analyses, we found that MDC was more expensive in patients without albuminuria versus those with UACR of 300 or 1,000 mg/g.
In the base case, we found that MDC was cost-effective in over 99% of probabilistic sensitivity analyses at a threshold of $87,500 per QALY and in over 95% of probabilistic sensitivity analyses at a threshold of $70,000 per QALY (Fig 3A–3D). In all subgroups, we found that MDC cost less than $150,000 per QALY in over 99% of probabilistic sensitivity analyses. Although findings were similar across eGFR levels, MDC tended to have a lower probability of cost-effectiveness for patients with less albuminuria.
When we decreased the effectiveness of MDC to 50% of the base case, we found that the program was cost-effective with 98% probability at a threshold of $150,000 per QALY (Fig 3E). However, the probability dropped to 74% under the scenario where the program was 25% as effective as the base case. The probability of cost-effectiveness dropped further when we assumed MDC only affected mortality and not progression to ESRD.
In most cost sensitivity analyses, the probability that MDC was cost-effective was greater than 95% at a threshold of $150,000 per QALY gained (Fig 3F). However, once we increased the cost of MDC 5-fold (relative to the base case), the probability declined to 89%.
Using data from literature, we developed and calibrated a deterministic Markov model that accurately models disease progression of patients with stage 3 and 4 CKD and accounts for population heterogeneity including age, sex, race, severity of kidney disease, and albuminuria. From this model, we found that a Medicare-funded MDC program in non-dialysis-requiring (eGFR 20 to 59 ml/min/1.73 m2) CKD is cost-effective in middle-aged to elderly patients. Although cost-effective in all subgroups, the program is more cost-effective in patients with more advanced CKD, particularly those with higher levels of albuminuria. Provision of MDC to younger patients was more expensive, but younger patients gained the most in health outcomes, as measured by QALYs. Notably, MDC remained cost-effective even if it was 5 times more expensive or one-quarter as effective as our base case.
Treating CKD requires effective management across multiple dimensions. Blood pressure control, especially the use of ACE inhibitors and angiotensin receptor blockers, is a mainstay of care [57,58]. Reducing cardiovascular risk factors and managing diet, fluid, and electrolyte balance are also important adjuncts [59,60]. Medication and dietary non-adherence are prevalent and can limit the success of CKD treatment [61,62].
MDC in CKD could help bridge the gap between successful scientific advances and lagging population health. Since CKD is a heterogeneous disease, determining subgroups that benefit most is important. Younger patients and those with less albuminuria gained the most QALYs. These findings are likely due to longer baseline life expectancy, since we found similar relative gains in health in all subpopulations. This probably explains why MDC is more expensive in younger patients and patients with less albuminuria, since a longer lifespan leads to additional healthcare spending, including on MDC. Importantly, we found that MDC represents excellent value potential for patients with higher levels of albuminuria, where slowing kidney progression can delay the onset of dialysis. Our study could help providers and program developers identify the subgroups of patients most likely to benefit from MDC.
Surprisingly, MDC was cost-effective, albeit more expensive, in patients without albuminuria, suggesting that MDC may be worthwhile in this population. The added expense is likely due to poor generalizability of MDC to patients without albuminuria who are at low risk for developing ESRD [33]. Even though MDC was cost-effective in this group on average, many such patients probably would not benefit from an intensive MDC program, especially one that aims to slow progression to ESRD. In patients without albuminuria, benefits from MDC probably reflect reduced mortality from cardiovascular disease rather than slowed CKD progression [63]. This finding was corroborated by our sensitivity analysis in which we assumed that MDC did not attenuate progression to ESRD. Here, MDC remained cost-effective, except under the most pessimistic scenario. Future studies of MDC could focus on disentangling the components of MDC that reduce cardiovascular complications and mortality from those that slow progression to ESRD, which would increase its applicability to patients without albuminuria.
Few studies have attempted to assess the economic impact of MDC in CKD, and none to our knowledge has used a CKD progression model incorporating disease heterogeneity. Several simulation studies have suggested that more optimally managing CKD could lead to up to $20,000 of savings per patient (€17,700 per patient) [22] and $60 billion annually [13]. Because these studies aimed to quantify only the savings associated with slowing progression to ESRD, they did not consider the cost of the intervention used to improve CKD management. We found that the cost of an MDC program would likely exceed potential savings from preventing the onset of ESRD, but that these excess costs would be relatively modest. Other studies using data from MDC trials also reported that MDC could lead to cost savings [10,17,41]. However, these studies did not incorporate the cost of anemia and bone-mineral metabolism management, which can be expensive, and none accrued costs over patients’ lifetimes.
Our results indicate that a US Medicare-funded MDC program would likely have significant costs. Thus, it is unlikely that healthcare providers would independently deploy such a program without reimbursement, since MDC could cost a provider seeing 100 patients with CKD $800,000 to $2,000,000 annually, assuming Medicare reimburses at marginal cost. This could explain why MDC has not had widespread adoption, especially as Medicare has begun rewarding providers for demonstrating cost savings [64–66]. While we did not show a reduction in Medicare expenditures, we found that modest investment in CKD care management could yield substantial gains in health. Additionally, our findings in patients 45–64 years of age could be extended to non-Medicare US healthcare payers because the vast majority of patients do not qualify for Medicare until they turn 65. Beneficiaries of these healthcare payers potentially have the most to gain, since MDC is especially beneficial for younger patients.
Other countries could also see improvements in health with small investments in MDC programs. The global prevalence of CKD is reported to be 8%–16% worldwide and continues to grow [67–69]. MDC in CKD has been successfully tested in non-US populations, including in Canada [7,15,16,42,46,70], the United Kingdom [43], France [44], Italy [70], the Netherlands [11], South Korea [14], and Taiwan [17,19,41,45,47,48], and could reduce the growing need for renal replacement at reasonable cost. In developing countries with poor access to renal replacement, MDC could be an inexpensive alternative to providing dialysis, which requires additional investment in infrastructure and capital [71,72].
Although Wang et al. reported that MDC reduces mortality and ESRD in CKD [21], we believe that their effectiveness estimates were inflated. Many of the included studies were not randomized controlled trials, and probably reflected populations likely to benefit. Additionally, providers actively investigating MDC are probably among the best at care management. Implementing a national program may be less efficient, especially if adherence to the program is modest or poor. It seems unlikely that a large-scale MDC program would be as effective as our base case. However, when we tested programs with much smaller effect sizes, MDC remained cost-effective, particularly in more severe kidney disease. Testing MDC programs that were less effective than our base case also allowed us to capture scenarios where patient adherence is poor. Nevertheless, Medicare would probably benefit from pursuing a cautious approach. Our most pessimistic scenario was not cost-effective, especially in patients with mild to moderate, non-proteinuric CKD. Implementing a pilot program for patients vulnerable to progressing to ESRD could increase the certainty that a nationally implemented program would be cost-effective.
Our study had some important limitations. First, the literature likely overestimates the effectiveness of MDC. We attempted to address this issue by using conservative estimates for our base case and by extensively testing changes in cost and effectiveness. We found that substantially less effective programs remained cost-effective. Second, our simulations relied on a calibrated model based on reported CKD progression and mortality rates. Results from published literature have limited generalizability, and although we accounted for population-level heterogeneity, our model cannot fully incorporate differences in patient factors. Third, although our model was novel in that it accounted for aspects of population-level heterogeneity, including age, sex, race, and severity of kidney disease, our model was unable to capture other known determinants of CKD progression such as socioeconomic status [73–75]. Fourth, our analysis relied on prior studies that did not generally stratify effectiveness by CKD stage. We accounted for this by assuming that MDC was less effective in milder stages of CKD and by varying its effectiveness in different CKD stages. Finally, we limited our investigation to patients with mild to moderate CKD and only studied the effect of MDC on CKD progression and all-cause mortality. Given the absence of data on the effect of MDC on intermediate endpoints, we were unable to incorporate cardiovascular or other hospitalizations into our cost-effectiveness estimates.
Strengths of our study include our development and calibration of a novel CKD progression model, which was able to reliably reproduce long-term rates of mortality and progression to ESRD in many different subpopulations. Using this model, we were able to detect variation in the cost-effectiveness of MDC that incorporates heterogeneity in the population. Our model is also flexible enough to allow investigators to test other interventions in patients with CKD. Additionally, we tested a wide swath of effectiveness and cost scenarios, and our results were robust to these changes except in the most pessimistic scenarios. Finally, our probabilistic sensitivity analysis allowed us to simultaneously test variation in all our parameters while accounting for correlations. By fitting our parameters to probability distributions, we were able to ensure that the joint distribution of our model fit parameters reported in literature.
In conclusion, our model estimates suggest that a Medicare-funded MDC program, even if implemented with modest efficiency, is likely to be cost-effective in middle-aged to elderly patients with mild to moderate CKD. Reimbursing providers for intensive disease management could be a relatively inexpensive way to improve the health of patients with CKD.
|
10.1371/journal.ppat.1007259 | A pathogen-derived effector modulates host glucose metabolism by arginine GlcNAcylation of HIF-1α protein | The essential role of pathogens in host metabolism is widely recognized, yet the mechanisms by which they affect host physiology remain to be fully defined. Here, we found that NleB, an enteropathogenic Escherichia coli (EPEC) type III secretion system effector known to possess N-acetylglucosamine (GlcNAc) transferase activity, GlcNAcylates HIF-1α, a master regulator of cellular O2 homeostasis. We determined that NleB-mediated GlcNAcylation at a conserved arginine 18 (Arg18) at the N-terminus of HIF-1α enhanced HIF-1α transcriptional activity, thereby inducing HIF-1α downstream gene expression to alter host glucose metabolism. The arginine transferase activity of NleB was required for its enhancement of HIF-1α transactivity and the subsequent effect on glucose metabolism in a mouse model of EPEC infection. In addition, HIF-1α acted as a mediator to transact NleB-mediated induction of glucose metabolism-associated gene expression under hypoxia. Thus, our results further show a causal link between pathogen infection and host glucose metabolism, and we propose a new mechanism by which this occurs.
| Accumulating evidence shows that pathogens can affect host metabolism, resulting in human diseases such as obesity and type 2 diabetes. However, how pathogens influence their hosts is still not clear, and this results in a lack of effective and specific clinical treatments. Further investigations into the causes of pathogen disturbance of host metabolism are urgently needed. In this study, we show that a protein molecule, NleB, secreted by enteropathogenic bacteria (EPEC) can get into host cells and modify the function of a master regulator of cellular O2 homeostasis, HIF-1α, thereby altering host glucose metabolism. We show that HIF-1α acts as a mediator to transact NleB-mediated induction of glucose metabolism-associated gene expression under hypoxia. Our results reveal a causal link between pathogen infection and host glucose metabolism, which may provide a new explanation for the causes of human diseases related to metabolic disturbance.
| Over the past decade, it has become clear that pathogens play an important role in host metabolism, but the mechanisms by which they affect host metabolism remain poorly defined [1–10]. Although pathogenic organisms likely rely on effectors to communicate with their host, the identity and function of pathogen-encoded effectors remain largely unknown [2,8,9]. Identifying pathogen effectors that affect host metabolism and defining the pathways that effectors use to enact their changes to host metabolism can help us better understand the relationship between pathogen infection and host physiology, as well as provide insights into the mechanisms underlying human diseases related to metabolic disturbance, such as obesity and type 2 diabetes (T2DM).
Human pathogenic Escherichia coli, enteropathogenic E. coli (EPEC), are attaching/effacing (A/E) pathogens that are the causative agent of infantile diarrhea. EPEC, enterohemorrhagic E.coli (EHEC), and the mouse pathogen Citrobacter rodentium all use a type III secretion system (T3SS) to transfer effectors (virulence proteins) into host intestinal epithelial cells to modulate various cell functions [11] [12]. The EPEC T3SS effector NleB, a protein known to inhibit host nuclear factor kappa-light-chain-enhancer of activated B cells (NF-kB) signaling [12], has recently been found to possess N-acetyglucosamine (GlcNAc) transferase activity that specifically modifies conserved arginine residues in death domain-containing host proteins, including TRADD, FADD, RIPK1, and TNFR1 [12–15]. Additionally, GAPDH is another target that becomes glycosylated by NleB at Arg197 and Arg200 [16], though this is debatable [13]. Due to this unprecedented function, NleB and its homologous effectors may modify additional host targets beyond those already identified [16,17]. Defining additional targets and demonstrating the consequences and underlying mechanisms of NleB-modified host targets will not only help us understand the correlation between pathogen and host physiology, but also provide an opportunity to discover therapeutic strategies to combat human diseases related to these pathogens.
Oxygen (O2) is a key factor for balancing demand against substrate availability in the glucose metabolic response, affecting both the rate and pathway of substrate utilization for energy production [18]. Based on O2 availability, HIF-1α, which is a master regulator of cellular O2 homeostasis, regulates glycolytic and oxidative glucose metabolism gene expression in the glucose metabolism pathway [18–20] [21–25]. Other host factors are already known to modulate HIF-1α function and thus also influence cellular glucose homeostasis [26,27], but other HIF-1α modulators may exist as well [20].
The intestines maintain low O2 conditions due to counter-current blood flow [28]. Multiple lines of evidence indicate that pathogens greatly influence host glucose metabolism, but the major underlying mechanisms remain largely unknown [10,20]. We found here that the T3SS effector NleB from EPEC modifies arginine residues in host HIF-1α through GlcNAcylation, thereby enhancing HIF-1 signaling. Moreover, the arginine GlcNAc transferase activity of NleB is required for enhancing host glucose metabolism in C. rodentium-infected mice. These results provide a causal, mechanistic link to explain how pathogens modulate host glucose metabolism.
While studying HIF-1α function, we noted that some arginine-to-lysine (R/K) mutations affected HIF-1α activity. However, none of the arginine-modifying factors we screened (such as arginine methyltransferases [PRMTs]) accounted for the effect displayed in HIF-1α R/K mutants. Two recent studies have shown that the EPEC T3SS effector NleB modifies host proteins through arginine GlcNAcylation [13,14], prompting us to investigate whether bacterial NleB can modify host HIF-1α by arginine GlcNAcylation and thereby affect host glucose metabolism [20] [1,2,18–20]. We first tested whether the wild-type NleB protein could GlcNAcylate HIF-1α by co-expressing GFP-tagged wild-type NleB (GFP-NleB) together with Myc-tagged HIF-1α (Myc-HIF-1α) in HEK293T cells. A clear band was detected by anti-arginine (glcnac) antibody (ab195033, Abcam) after the lysates were co-immunoprecipitated (co-IP) with anti-Myc antibody-conjugated agarose beads [17,29], but this band was not present when Myc-HIF-1α was co-expressed with a GFP-tagged catalytic motif (DXD) mutant NleB (GFP-NleB-DXD, Asp221Ala/Asp223Ala double mutant) (Fig 1A). These observations suggest that NleB might GlcNAcylate HIF-1α at arginine residues [29], and that the arginine GlcNAc transferase activity of NleB is required for this GlcNAcylation.
To confirm that NleB indeed GlcNAcylated HIF-1α at arginine residues, we made an R/K HIF-1α mutant (HIF-1α-R-free, in which all 35 arginine residues were mutated to lysine residues) and examined the ability of NleB to GlcNAcylate it. After co-expression and co-IP, the anti-arginine (glcnac) antibody could not detect any GlcNAcylated residues in HIF-1α-R-free, indicating that NleB could indeed GlcNAcylate arginine(s) in HIF-1α (Fig 1B).
To determine whether NleB could GlcNAcylate endogenous HIF-1α, we took advantage of the RCC4 cell line, an VHL-deficient kidney cancer cell line in which HIF-α (HIF-1α and HIF-2α) is highly expressed under normoxia [30]. When GFP-NleB was overexpressed, an clear band was detected by anti-arginine (glcnac) antibody after the lysates were co-immunoprecipitated with anti-HIF-1α antibody, but this band was not present when GFP-NleB-DXD was overexpressed (Fig 1C). This data suggests that NleB GlcNAcylates endogenous HIF-1α.
To determine whether NleB could GlcNAcylate endogenous HIF-1α under hypoxia, we used the HCT116 cell line, in which endogenous HIF-1α is induced by hypoxia. Similar to what was exhibited in the RCC4 cell line, when GFP-NleB was overexpressed, an clear band was detected by anti-arginine (glcnac) antibody after the lysates were co-immunoprecipitated with anti-HIF-1α antibody, but this band was not present when GFP-NleB-DXD was overexpressed (Fig 1D). This data suggests that NleB GlcNAcylates endogenous HIF-1α under hypoxia.
Subsequently, to determine whether EPEC T3SS-delieved NleB could GlcNAcylate endogenous HIF-1α similar to NleB overexpression, we used the wild-type EPEC strain (E2348/69) and the mutant EPEC strains with different genetic backgrounds after they were complemented with various genes (S1 Table; S1 Fig; Fig 1E). In EPEC, NleB is encoded directly upstream from NleE, another type III effector [12,31]. In addition, it has been reported previously that both NleB and NleE can suppress NF-kB activation [12,31]. To exclude the potential effects of NleE on HIF-1α, we used a mutant EPEC strain with double deletion of nleE and nleB (ΔnleBE) for subsequent assays, similar to that used by Li et al. [13]. HIF-1α GlcNAcylation was detected only in RCC4 cells infected with the wild-type EPEC (E2348/69) (S1 Fig) and in the nleE- and nleB-deleted mutant EPEC strain complemented with a plasmid expressing wild-type NleB (ΔnleBE+pNleB), but not in uninfected cells or in cells infected with the escN-deleted EPEC mutant strain (which is defective in its T3SS delivery system and is unable to deliver proteins into host cells) complemented with a plasmid expressing wild-type NleB (ΔescN+pNleB), the control ΔnleBE+vector, or ΔnleBE+pNleB-DXD (S1 Fig; Fig 1E).
This observation suggests that NleB delivered by bacteria efficiently GlcNAcylates endogenous HIF-1α in host cells, and that the T3SS machinery is required for NleB to GlcNAcylate endogenous HIF-1α.
To identify which arginine residue(s) in HIF-1α were GlcNAcylated by NleB, we conducted a series of fine mapping assays using regional R/K HIF-1α mutants. We first narrowed the region down to residues 1–311 (Fig 2A) and then to 1–50 (Fig 2B and 2C). Of the 5 arginine residues, we found that they are evolutionarily conserved between HIF-1α and HIF-2α (Fig 2D). GlcNAcylation of only the R18K mutant was greatly reduced compared to the other 4 single R/K mutants, including the R17 residue next to R18 in the wild-type HIF-1α (Fig 2E and 2F). In addition, mass spectrometry analysis also suggested that Arg18 (R18) was GlcNAcylated by NleB (S2 Table). These data suggest that NleB GlcNAcylates the HIF-1α protein at multiple arginine sites with Arg18 (R18) as the key site.
To determine whether NleB GlcNAcylates HIF-1α through direct protein–protein interaction, we assessed the interaction between NleB and HIF-1α. Initially, we co-transfected GFP-HIF-1α and RFP-NleB into HeLa cells or HCT116 cells and found that HIF-1α co-localized with NleB (Fig 3A). Subsequently, we co-transfected GFP-NleB and Myc-HIF-1α into HEK293T cells. Myc-HIF-1α pulled-down GFP-NleB when anti-Myc-conjugated agarose beads were used for co-IP (Fig 3B). Confirming this finding, Flag-NleB also reciprocally pulled-down Myc-HIF-1α (Fig 3C), and endogenous HIF-1α could pull-down transfected GFP-NleB in HCT116 cells (Fig 3D).
To further determine which HIF-1α region was essential to bind to NleB, we performed domain mapping and found that the middle region (330–399 amino acids) had a strong ability to bind to NleB (Fig 3E). Interestingly, the R/K mutant, HIF-1α-R-free, still interacted with NleB even though it was not glycosylated by NleB (S2A Fig), and the transferase-deficient NleB-DXD also interacted with wild-type HIF-1α (S2B Fig).
Moreover, glutathione S-transferase (GST)-pull-down assays using GST-tagged NleB and His-tagged HIF-1α (330–399 aa) expressed in E. coli showed that GST-NleB could pull-down His-HIF-1α-330-399 aa (Fig 3F).
These results suggest that NleB interacts with endogenous HIF-1α directly, and that NleB might GlcNAcylate HIF-1α through direct interaction.
To determine the biological consequences of NleB-mediated GlcNAcylation of HIF-1α, we examined whether NleB affected HIF-1α transcriptional activity. Initially, we took advantage of luciferase reporter assays using well-defined luciferase reporters to monitor HIF-1α activity [32–37]. NleB expression enhanced p2.1-luciferase reporter and hypoxia response element (HRE)-reporter activity activated by HIF-1α overexpression in HCT116 cells under normoxia [34,35] (Fig 4A and 4B). Under hypoxia, NleB expression significantly enhanced p2.1-luciferase reporter and HRE-reporter activity (Fig 4C and 4D). The expressions of transfected HA-HIF-1α and GFP-NleB were confirmed by western blot assays (S3A, S3B, S3C and S3D Fig).
Because we were particularly interested in whether and how the pathogen correlated with host glucose metabolism, we next selectively examined whether NleB affected glucose metabolism-associated genes known to be downstream of HIF-1α [19,20,25,38–42], including LDHA, PDK1, PGK1, SLC2A1 (GLUT1), PKM2, and HK1 [20]. Consistent with the above reporter assays, NleB expression caused an increase in the mRNA levels of these genes in HCT116 cells under hypoxia as revealed by semi-quantitative RT-PCR assays (Fig 4E, 4F, 4G and 4H). The expressions of transfected GFP-NleB were confirmed by western blot assays (S3E Fig).
Interestingly, NleE also enhanced HIF-1α transcriptional activity as observed by promoter assays in both HCT116 cells and HeLa cells (S4 Fig), further reinforcing the usefulness of the nleE- and nleB-deleted mutant EPEC strain for analyzing NleB function in this study.
Next, we examined whether EPEC-delivered NleB can enhance HIF-1α transcriptional activity. When HIF-1α was overexpressed in HCT116 cells under normoxia, the wild-type EPEC E2348/69 strain enhanced p2.1-luciferase reporter and hypoxia response element (HRE)-reporter activity compared with the nleE- and nleB-deleted mutant EPEC strain complemented with an empty control vector (ΔnleBE+vector) (Fig 4I and 4J). Similarly, under hypoxia, infection with the wild-type EPEC E2348/69 strain enhanced p2.1-luciferase reporter and hypoxia response element (HRE)-reporter activity compared with infection with the control EPEC strain ΔnleBE+vector (Fig 4K and 4L). Of note, the promoter activity in cells without infection was higher than that in infected cells (Fig 4I, 4J, 4K and 4L). This phenomenon was probably due to the unhealthy status of the infected cells, which were not suitable for efficient expression of transfected vectors.
Additionally, we found that NleB also enhanced the promoter activity of BNIP3, which is a HIF-α downstream target responsible for hypoxia-induced cell death [43–45] [46] (S5 Fig), indicating that NleB might impact hypoxia-induced cell death by modulating HIF-α [47].
Consistently, under hypoxia, the infection of the wild-type EPEC E2348/69 strain induced the expressions of PDK1, PGK1, SLC2A1 (GLUT1), and PKM2, four well-defined HIF-1α targets, but did not induce the expression of SOD2, a well-defined HIF-2α target (Fig 4M).
To further confirm this induction, we examined the expressions of SLC2A1 (GLUT1) and PDK1 in RCC4 cells under normoxia. As shown in Fig 4N and 4O, only infection with the wild-type EPEC E2348/69 or the mutant strain ΔnleBE+pNleB resulted in significant induction of SLC2A1 (GLUT1) and PDK1, but in cells that were infected with the mutant strain ΔescN+pNleB, the control strain ΔnleBE+vector, or the mutant strain ΔnleBE+pNleB-DXD, there was no significant effect on the induction of SLC2A1 (GLUT1) and PDK1 (Fig 4N and 4O).
These results suggest that NleB enhances HIF-1α transcriptional activity, that NleB delivered by bacteria efficiently enhances HIF-1α transcriptional activity, and that the T3SS machinery is required for NleB to enhance HIF-1α transcriptional activity.
To determine whether the effect of NleB on hypoxia signaling is mediated by HIF-1α, we initially knocked-down HIF-1α by transient transfection of HIF-1α shRNA in HCT116 cells (S6 Fig). As shown in Fig 5A and 5B, when HIF-1α shRNAs were co-transfected, the activities of the p2.1-luciferase reporter and hypoxia response element (HRE) reporter induced under hypoxia were reduced significantly (Fig 5A and 5B), indicating the efficiency of HIF-1α shRNAs (HIF-1α shRNA-1 and HIF-1α shRNA-2) in knocking-down HIF-1α. When HIF-1α was knocked-down by HIF-1α shRNA-2, the enhancement of expression of SLC2A1 (GLUT1) and PGK1 by overexpression of NleB under hypoxia was diminished (Fig 5C and 5D).
To further confirm that the enhancement of hypoxia-inducible gene expression by NleB in HCT116 cells was indeed mediated by HIF-1α, we specifically inhibited HIF-1α with 25μM PX-478 [48,49]. PX-478 prevented enhanced SLC2A1 and PGK1 expression by overexpression of NleB under hypoxia (Fig 5E and 5F).
Furthermore, we examined the effect of NleB delivered by EPEC on hypoxia-inducible gene expression in RCC4 cells using lentivirus to knock-down HIF-1α (because RCC4 cells are difficult to transfect using the transfection reagents tested so far). As shown in Fig 6A, infection with HIF-1α shRNA lentivirus (HIF-1α-shRNA-1 and HIF-1α-shRNA-2) specifically prevented HIF-1α expression, but not HIF-2α expression, in RCC4 cells, which was further confirmed by detecting LDHA protein (a specific downstream target of HIF-1α [32]) (Fig 6A). When HIF-1α was knocked-down by adding HIF-1α-shRNA-2 lentivirus, the enhancement of expression of SLC2A1 (GLUT1) and PDK1 by infection with the wild-type EPEC E2348/69 strain was diminished (Fig 6B and 6C). Moreover, PX-478 prevented the enhanced expressions of SLC2A1 (GLUT1), PDK1, and PKM2 induced by infection with the mutant EPEC strain ΔnleBE+pNleB (Fig 6D, 6E and 6F).
Taken together, these results suggest that HIF-1α mediates NleB’s ability to enhance glucose metabolism-associated gene expression. We also confirmed these findings in HeLa cells (S7 Fig), suggesting that NleB-mediated enhancement of HIF-1α transcriptional activity was not dependent on host cell type.
Interestingly, NleB overexpression did not significantly enhance transactivity of the HIF-1α-R18K mutant (S8 Fig), reinforcing the idea that Arg18 is the key GlcNAcylated site in HIF-1α.
Given that proline hydroxylation of HIF-α by PHD proteins (PHD1, PHD2, or PHD3) heavily regulates HIF-α function [50,51], we asked whether GlcNAcylation of HIF-1α by NleB could impact HIF-1α hydroxylation. In HEK293T cells, GlcNAcylation by NleB still occurred in the hydroxylation site-dead HIF-1α mutant (DM), in which the two functional proline residues (Pro402/564) were mutated into alanine (Pro402/564Ala) (S9A Fig). Moreover, HIF-1α hydroxylation by PHD2 was unaffected by NleB expression as revealed by probing the co-IP lysates with an anti-HIF-OH antibody (S9B Fig). In addition, during EPEC infection, NleB delivered by EPEC also did not alter HIF-1α hydroxylation (S9C Fig). As expected, NleB expression in HCT116 cells also enhanced HIF-1α-DM activity under normoxia (S9D and S9E Fig).
It has been proposed that in cancer cells, OGT alters HIF-1α stability via regulation of α-ketoglutarate levels [52]. To determine whether NleB operates by a similar mechanism on HIF-1α, we measured α-ketoglutarate levels after EPEC infection. We found that NleB did not alter α-ketoglutarate levels in HCT116 cells (S10 Fig).
These results suggest that NleB enhances HIF-1α transcriptional activity, resulting in upregulation of glucose metabolism-associated genes downstream of HIF-1α.
Since HIF-2α is structurally similar to HIF-1α and also performs its function under hypoxic conditions, we assessed whether NleB had a functional effect on HIF-2α in the context of hypoxia. We found that expression of wild-type NleB, but not NleB-DXD, was able to GlcNAcylate HIF-2α (S11A Fig), and NleB could also interact with HIF-2α (S11B Fig). To our surprise however, NleB had no detectable effect on the transcriptional activity of HIF-2α as revealed by promoter assays in both HCT116 and HeLa cells (S11C to S11F Fig). Actually, we had already noticed that infection with the wild-type EPEC E2348/69 strain did not enhance expression of SOD2 (an HIF-2α-specific downstream gene) in HCT116 cells under hypoxia (Fig 4M).
Further confirming this finding, the expression of HIF-2α-specific downstream genes (PAI1, POU5F1, SOD2, and CITED2) did not change in RCC4 or 780-O (a pVHL-deficient kidney cancer cell line containing only HIF-2α) cells after infection with the EPEC strains with or without NleB [32,33], and bacteria-delivered NleB also did not change the expression of HIF-2α-specific downstream genes (S11G and S11H Fig).
Therefore, NleB may not affect HIF-2α transactivity even though it can still GlcNAcylate HIF-2α, indicating a divergent influence of NleB-mediated arginine GlcNAcylation on hypoxia signaling pathways.
The role of HIF-1α in cellular glucose metabolism has been well defined [20,25,39,53]. Under hypoxia, HIF-1α directly transactivates the expression of glucose metabolism-associated genes [20]. Based on the above observations, NleB might enhance HIF-1α activity to upregulate the expression of glucose metabolism-associated genes, including LDHA, PDK1, PGK1, SLC2A1 (GLUT1), PKM2, and HK1. Cellular glucose uptake is a major step in glucose metabolism and is mediated by a cell membrane glucose transport system that includes glucose transporter 1 (encoded by GLUT1). To further delineate the biological consequence of NleB-mediated enhancement of host glucose metabolism-associated gene expression, we examined cellular glucose uptake. In RCC4 cells (which contain both HIF-1α and HIF-2α) [30], ΔnleBE+pNleB infection caused higher cellular uptake of 2-NBDG (a fluorescence glucose analog) as quantified by flow cytometry (FACS) compared with cells infected with the control (nleBE+vector or ΔnleBE+pNleB-DXD strains) (p = 0.0114) (Fig 7A).
To further determine whether NleB-enhanced cellular glucose uptake is dependent on HIF-1α, we knocked-down HIF-1α in RCC4 cells by infection with HIF-1α shRNA-2 lentivirus. As shown in Fig 7B, when HIF-1α was knocked-down, NleB-enhanced cellular glucose uptake by infection with the wild-type EPEC E2348/69 strain was diminished (p = 0.1184) (Fig 7B). Additionally, in 780-O cells, which only contain HIF-2α but not HIF-1α [54], infection with the wild-type EPEC E2348/69 strain, or the control EPEC strain ΔnleBE+vector exhibited similar glucose uptake abilities (p = 0.1364) (Fig 5C).
Moreover, we examined cellular ATP levels in RCC4 cells after infection with the wild-type EPEC E2348/69 strain, the control EPEC strain ΔnleBE+vector, the mutant EPEC strain ΔnleBE+pNleB, the mutant EPEC strain ΔnleBE+pNleB-DXD, or without infection. As shown in Fig 7D, infection with the wild-type EPEC (EPEC E2348/69) strain enhanced cellular ATP levels compared to infection with the control EPEC strain (ΔnleBE+vector); infection with the mutant EPEC ΔnleBE+pNleB strain enhanced cellular ATP levels compared to infection with the mutant EPEC strain ΔnleBE+pNleB-DXD (Fig 7D), reinforcing the role of NleB in cellular glucose metabolism.
Subsequently, we used a metabolomics approach to further address whether NleB indeed enhances cellular glycolytic capacity under hypoxia. We examined the effect of NleB on intermediates from HCT116 cells using liquid chromatography-mass spectrometry (LC-MS). The metabolic profile of HCT116 cells infected by the strain ΔnleBE+pNleB demonstrated a general increase in glycolytic and pentose phosphate pathway (PPP) intermediates (Fig 7E) and a decrease in TCA cycle intermediates compared to infection with the strain ΔnleBE+vector (Fig 7F).
These results suggest that the T3SS-delivered effector NleB redirects cellular metabolism in favor of glycolysis, and that HIF-1α, but not HIF-2α, mediates the function of NleB in glucose metabolism.
The above results indicate the effect of bacterial NleB on host glucose metabolism in vitro. To determine whether the effect on host glucose metabolism is relevant and can occur in vivo, we first examined the expressions of glucose metabolism-associated genes known to be downstream of HIF-1α in mouse colon after infection with wild-type C. rodentium (WT) or its mutant strains, including an nleB-deleted mutant (Δnleb), complemented with wild-type nleB (Δnleb+pNleBc), GlcNAc transferase-deficient nleB (Δnleb+pNleBc-DXD), or without infection. The mRNA levels of Slc2a1 (Glut1) and Pkm2 were enhanced significantly after infection with wild-type C. rodentium or the mutant strain Δnleb+pNleBc compared to infections with the mutant strain Δnleb and the mutant strain Δnleb+pNleBc-DXD (Fig 8A). However, when the mice were treated with PX-478, the enhancement of Slc2a1 (Glut1) and Pkm2 by NleB in mouse colon was diminished (Fig 8B). Moreover, the protein levels of Ldha, Pkm2, and Vegf (another downstream target of HIF-α irrelevant to glucose metabolism) were also higher in mouse colons infected with C. rodentium WT compared to those infected with the nleB-deleted mutant, Δnleb (Fig 8C). By immunofluorescent staining, Slc2a1 (Glut1) was found to be more highly expressed in mouse colons infected with C. rodentium WT compared to those infected with Δnleb (Fig 8D).
To validate the effect of PX-478 on the suppression of HIF-1α expression in mouse colon, we examined the protein levels of HIF-1α by western blotting (S12A Fig). Treatment with PX-478 also inhibited the expressions of glucose metabolism-associated genes known to be downstream of HIF-1α in mouse colons (S12B Fig). Notably, PX-478 did not alter the colonization of C. rodentium in feces and mouse colon (S12C and S12D Fig).
Furthermore, we examined blood glucose levels in mice after infection with the C. rodentium WT or its mutant strains, including Δnleb, complemented with wildtype nleB (Δnleb+pNleBc) or GlcNAc transferase-deficient nleB (Δnleb+pNleBc-DXD). At timepoints either before or after fasting (6 h), mice infected with C. rodentium WT had significantly lower blood glucose levels compared to mice infected with the Δnleb mutant strain (S13A and S13B Fig), and we confirmed that GlcNAc transferase activity was required for nleB to reduce mouse blood glucose levels (S13A and S13B Fig). Notably, the blood glucose levels in mice infected with the Δnleb+pNleBc strain were lower than those in mice infected with C. rodentium WT (S13A and S13B Fig), implying that the Δnleb pNleBc strain could deliver more pNleBc into host cells than C. rodentium WT due to overexpression of pNleBc, resulting in greater enhancement of HIF-1α activity. In addition, this effect also depended on Hif-1α activity, as the inhibitory role of nleB on mouse blood glucose levels diminished when the mice were treated with PX-478 (S13A and S13B Fig).
Given that the liver is virtually the only major organ that produces and supplies blood glucose to maintain blood glucose levels, we tested if gastrointestinal infection with C. rodentium leads to glucose metabolism-associated gene expression alterations in the liver in addition to their actual infection site—the colon. We went on to examine slc2a1 and pkm2 expression levels in mouse livers after infection with the different C. rodentium strains. Consistent with the role of NleB in the colon and on blood glucose levels, NleB enhanced glucose metabolism-associated gene expression, and GlcNAc transferase activity was required for this enhancement (S13E and S13F Fig). The reduction of Hif-1α by PX-478 injection was confirmed in mouse livers by western blot (S13G Fig). However, NleB had no effect on the proliferation of pancreatic islets and did not alter blood insulin levels in mice (S14A and S14B Fig). Moreover, NleB also did not change homeostasis model assessment of insulin resistance (HOMA-IR) significantly in mice (S14C Fig). It appears that NleB affects glucose metabolism independent of host insulin action.
These results suggest that NleB enhances host glucose metabolism in vivo and that NleB requires its GlcNAc transferase activity to function efficiently.
Accumulating evidence has revealed that pathogens affect energy homeostasis, glucose metabolism, and metabolic inflammation, but the causal link between pathogen infection and host metabolism remains to be fully understood [10,55,56]. In this study, our finding that bacterial NleB, a virulence protein (effector) delivered by the T3SS in gut A/E pathogens such as EPEC or C. rodentium, modulates host glucose metabolism by GlcNAcylating arginine residues in HIF-1α [13] provides a causal link between pathogen infection and host glucose metabolism; furthermore, this finding could open new avenues for exploring the underlying causes by which pathogen infection affects host metabolism. Given that the regulation of host glucose metabolism is complicated and that HIF-1α mainly participates in glucose metabolism under hypoxia [20], future studies determining how pathogens modulate host glucose metabolism beyond HIF-1 signaling would be of great interest.
It has been noticed that NleB overexpression either by direct transfection of expression vectors or by infection with the mutant strain ΔnleBE+pNleB might GlcNAcylate targets non-specifically [57]. Therefore, we must seriously consider whether HIF-1α is a specific target modified by NleB. In this study, even though we could not always detect the GlcNAcylation of endogenous HIF-1α by infection of the wild-type EPEC strain, we always observed that NleB delivered by the wild-type strain enhanced HIF-1α transcriptional activity. In addition, knockdown of HIF-1α by HIF-1α shRNA diminished the enhancement of glucose uptake by NleB delivered by the wild-type EPEC strain. Furthermore, knockdown of HIF-1α either by HIF-1α shRNA or by the inhibitor PX-478 diminished the enhancement effect of NleB on HIF-1α transcriptional activity. Thus, it appears that HIF-1α is a native target of NleB. Because of the relatively low level of NleB delivered by the wild-type EPEC strain compared to that delivered by the mutant strain ΔnleBE+pNleB in which NleB is overexpressed, it was often hard to detect GlcNAcylation of endogenous HIF-1α by western blotting. In addition, in the mouse infection model, we noticed that infection with ΔnleB+pNleBc resulted in lower levels of blood glucose compared to infection with C. rodentium WT. The culture status of the wild-type EPEC strain and/or host cells might affect the efficiency of NleB GlcNAcylation of HIF-1α.
Remarkably, the HIF-1α-R18K mutant was still GlcNAcylated by NleB even though it exhibited the lowest GlcNAcylation efficiency compared to all other mutants, indicating that arginine residues other than arginine 18 in HIF-1α might still be GlcNAcylated by NleB. However, NleB did not significantly enhance the transactivity of the HIF-1α-R18K mutant. Therefore, R18 at the N-terminus of HIF-1α appears to be the key site for GlcNAcylation by NleB. Of note, R18 is located in the DNA binding domain (DB), but not the transactivation domain (TAD), of HIF-1α. So GlcNAcylation of HIF-1α protein by NleB might mainly enhance the DNA binding ability of HIF-1α rather than enhancing transactivity directly, resulting in enhanced hypoxia-inducible glucose metabolism-associated gene expression.
However, due to technical limitations, we could not provide direct evidence that endogenous HIF-1α was indeed GlcNAcylated by NleB after infection with the WT and nleB mutant. Therefore, we are not sure whether HIF-1α is modified by NleB under physiological levels of the effector. Hopefully, more advanced techniques will be developed to specifically address this concern and clarify it in the near future.
Pathogens also affect host glucose metabolism via various mediators encoded in pathogens and host cells [58–63], raising the question of whether the host, the pathogen, or both benefit from the interaction. On one hand, modulation of host glucose metabolism by the pathogen might benefit pathogen infection [63–65]. On the other hand, this modulation might also contribute to the development of host diseases related to glucose metabolic disturbance. Therefore, further investigation into the relationship between pathogen infection and host glucose metabolism may provide some clues for developing strategies for both treatment of pathogen infection and the prevention of host diseases related to disturbances in glucose metabolism. Interestingly, normal glucose uptake in the brain and heart requires an endothelial cell-specific HIF-1α-dependent function, which is consistent with our observation that HIF-1α enhancement by NleB increases glucose uptake [25].
As a virulence protein, NleB is known to influence the host immune response, such as through its effects on FADD, TRADD, and RIPK1 [11,13,14,16]. Since HIF-1α is also involved in host immune response to bacterial infection [66–69], we initially also asked whether NleB affected the host immune response by modifying HIF-1α. However, we could not activate an NF-kb reporter and induce TNF-α expression by overexpressing HIF-1α in our cellular system, which should have worked based on previous reports [66,67]. Therefore, we still cannot definitively answer whether NleB affects host immune response through modification of HIF-1α.
Intriguingly, even though NleB GlcNAcylated arginine(s) on HIF-2α, NleB did not enhance HIF-2α transcriptional activity like it did for HIF-1α. As two master regulators of hypoxia signaling, HIF-1α and HIF-2α share the common function of regulating similar downstream genes, but HIF-1α and HIF-2α also have divergent functions regulating their own specific downstream targets [39]. Thus far, since HIF-1α, but not HIF-2α, has been identified as a regulator of glucose metabolism [20], the fact that NleB specifically enhanced HIF-1α transactivity is consistent with its role in host glucose metabolism. Nevertheless, the consequences of NleB GlcNAcylating arginine(s) in HIF-2α are of interest for future studies.
Due to embryonic lethality of HIF-1α-null mice [70,71], we cannot provide genetic evidence to show that NleB-enhanced host glucose metabolism is indeed mediated by HIF-1α. In this study, to obtain in vivo data to understand the physiological role of the type III effector NleB, we used the HIF-1α inhibitor PX-478 in a mouse model. Even though there are non-specific inhibitory roles of PX-478, we found that PX-478 had the same effect as that of HIF-1α shRNA in the cell culture system. In addition, PX-478 effectively blocked HIF-1α protein levels in mouse colon and liver. Therefore, PX-478 might be suitable for specifically suppressing HIF-1α function in mouse models.
Based on the observations in this study, we propose a working model for the role of bacteria-delivered NleB on host glucose metabolism via its effect on HIF-1α (Fig 9). The host gut maintains low O2 conditions (hypoxia) due to counter-current blood flow [28]. When bacteria such as EPEC infect the intestines, the T3SS delivers NleB into host intestinal cells. NleB then binds and GlcNAcylates HIF-1α at arginine residues, enhancing HIF-1α transcriptional activity and increasing expression of glucose metabolism-associated genes downstream of HIF-1α. This modulates processes related to glucose uptake (GLUT1) and glycolysis (HK1, PGK1, PKM2, LDHA, and PDK1).
All procedures involving the use of mice were approved by the ethical board of the Animal Care and Use Committee of the Institute of Hydrobiology, Chinese Academy of Sciences (protocol number IHB-2017001). All protocols conform to the Guide for the Care and Use of Laboratory Animals of the U.S. National Institutes of Health. All efforts were made to minimize the number of animals and their suffering.
HEK293T, HeLa, HCT116, and 786-O cells were originally obtained from the American Type Culture Collection (ATCC). RCC4 cells were kindly provided by Peter J. Ratcliffe. HEK293T, HeLa, and RCC4 cells were grown in Dulbecco’s modified Eagle’s medium (DMEM) (HyClone) supplemented with 10% fetal bovine serum (FBS). 786-O cells were grown in RPMI 1640 (HyClone) supplemented with 10% FBS. HCT116 cells were grown in Mc-Coy5A (HyClone) supplemented with 10% FBS. Cells were cultivated in a humidified incubator containing 5% CO2 at 37°C. Cells were cultivated under hypoxic conditions (2% O2) by using an incubator with O2 control filled with 5% CO2 and balanced with N2 (NBS Galaxy 48R). All cell lines were verified to be free of Mycoplama contamination before use.
Cells were seeded in 24-well plates and transfected with the indicated plasmids using VigoFect (Vigorous Biotech, Beijing, China). pRL-SV40 Renilla was as an internal control. Luciferase activity was measured 18–24 h after transfection using the Dual-luciferase Reporter Assay System (Promega). Data were normalized to Renilla luciferase. Luciferase data are reported as mean ± s.e.m. of three independent experiments performed in triplicate.
The EPEC E2348/69 ∆nleB/E SC3909 strain (∆IE2:: kan and nleBE IE6:: tet) and its derivatives were kindly provided by Feng Shao, and were used for cell culture infection (S1 Table). A single bacterial colony was inoculated into 0.5 mL of LB medium and statically cultured overnight at 37°C. Bacterial cultures were then diluted 1:40 in DMEM supplemented with 1 mM isopropyl-β-D-thiogalactoside (IPTG) and cultured for an additional 4 h at 37°C in the presence of 5% CO2. Infection was performed at a multiplicity of infection of 200:1 in the presence of 1 mM IPTG for 2 h. Cells were washed four times with PBS. To assay NleB-induced modification, HEK 293T cells were transfected with pCMV-Myc-HIF-1α or pCMV-Myc-HIF-1α plasmids 24 h before infection.
The p2.1 reporter was purchased from ATCC and is commonly used for monitoring HIF-1α transcriptional activity [34]. Hypoxia response element (HRE) reporter and SOD2 promoter luciferase reporter were provided by Navdeep Chandel and Xin-Hua Feng, respectively. The HRE-reporter contains three hypoxia response element (HRE) repeats and is commonly used for monitoring HIF-1α transcriptional activity [35]. The wild-type EPEC NleB gene was originally obtained from Feng Shao, and its enzymatically dead mutant (D221/223A) was PCR amplified and subcloned into the pCS2-EGFP vector and pGEX-4T vector. Wild-type human HIF-1α was subcloned into the pCMV-Myc vector and pCMV-HA vector (Clontech). Human HIF-1α-330-399aa was cloned into the pET32a vector. All R/K mutants of HIF-1α were subcloned into the pCMV-Myc vector. Wild-type human HIF-2α was subcloned into the pCMV-Myc vector and the pCMV-HA vector. Human HIF-1α domains were subcloned into the pCMV-Myc vector.
The antibodies used were as follows: anti-c-Myc antibody (9E10, 1:1000 for IB analysis, Santa Cruz), anti-HA antibody (1:5000 for IB analysis, Covance), anti-β-actin antibody (AC004, 1:20,000 for IB analysis, ABclonal), anti-HIF-1α antibody (A6265, 1:1000 for IB analysis, ABclonal), anti-GFP antibody (AE012, 1:1000 for IB analysis, ABclonal), anti-HIF-OH antibody (3434S, 1:1000 for IB analysis, Cell Signaling), anti-Arginine (glcnac) (ab195013, 1:1000 for IB analysis, Abcam), anti-His (H15, 1:500 for IB analysis, Santa Cruz), anti-Ldha antibody (A1146, 1:1000 for IB analysis, ABclonal), anti-Pkm2 antibody (GB11392, 1:1000 for IB analysis, Servicebio), anti-Vegf (GB11034, 1:1000 for IB analysis, Servicebio), anti-Slc2a1 (Glut1) antibody (GB11215, 1:100 for IF analysis, Servicebio), anti-Ki67 antibody (GB113030-2, 1:100 for IF analysis, Servicebio), anti-Insulin antibody (GB13121, 1:100 for IF analysis, Servicebio).
PX-478 (S7612) was purchased from Selleck. 6-[N-(7-Nitrobenz-2-oxa-1,3-diazol-4-yl) amino]-2-deoxy-D-glucose (2-NBDG), a glucose analog, was purchased from Invitrogen. Insulin (Mouse) ELISA Kit (KA3812) was purchased from Abnova. Alpha-Ketoglutarate Assay Kit (K677-100) was purchased from Biovision.
Coimmunoprecipitation and western blot analysis were performed as previously described [33]. Anti-Myc and anti-Flag antibody-conjugated agarose beads were purchased from Sigma. Protein A/G-Sepharose beads were purchased from GE Company. A Fuji Film LAS4000 mini-luminescent image analyzer was used to photograph the blots. Multi Gauge V3.0 was used to quantify protein levels based on the band density.
His-tagged HIF-1α-330aa-399aa and GST-tagged NleB (DXD) were expressed in E. coli BL21-Gold (DE3). GST resin (Novagen) was used for protein purification. The gels were stained by Coomassie blue or transferred to polyvinylidene difluoride (PVDF) membranes for western blot assays.
Total RNA was extracted using TRIzol reagent (Invitrogen). cDNA was synthesized using a first-strand cDNA synthesis kit (Fermentas). The following primers were used for internal control 18S rRNA: 5’-TCAACTTCGATGGTAGTCGCCGT-3’ and 5’-TCCTTGGATGTGGTAGCCGTTCT-3’. Other primers were synthesized as described in previous reports [32,33]. Semi-quantitative RT-PCR data are reported as mean ± s.e.m. of three independent experiments performed in triplicate.
To determine GlcNAcylation site(s) in HIF-1α by NleB, HA-HIF-1α and GFP-NleB were co-expressed in 293T cells, and HA-HIF-1α was purified using anti-HA antibody-conjugated agarose beads. Then, the purified HIF-1α was digested by trypsin and analyzed by online nanoflow LC-MS/MS using the Ultimate 3000 nano-LC system (Dionex, Sunnyvale, CA) connected to an LTQ-Orbitrap Elite (Thermo Scientific) mass spectrometer. Samples were injected onto an analytical C18-nanocapillary LC column (C18 resin with 3 μm particle size, 15 cm length × 75 μm inner diameter, Acclaim PepMap RSLC, Thermo Scientific) and eluted at a flow rate of 300 nL/min with a 130 min gradient from 5% solvent B (90% ACN/0.1% formic acid, v/v) to 50% solvent B. The peptides were then directly ionized and sprayed into an Orbitrap Elite mass spectrometer by a nanospray ion source. The mass spectrometer was operated in data-dependent mode with an automatic switch between MS and MS/MS acquisition. Full MS spectra from m/z 350 to 1800 were acquired with a resolution of 60,000 at m/z 400 in profile mode. Following every survey scan, up to 15 of the most intense precursor ions were picked for MS/MS fragmentation by high-energy collisional dissociation (HCD) with a normalized collision energy of 30%. The dynamic exclusion duration was set to 120 s with a repeat count of one and ±10 ppm exclusion window.
All acquired raw data were processed with pFind software (version 3.1.2) (Chi et al., 2018) and searched against the NCBI Human Protein Database (https://www.ncbi.nlm.nih.gov/protein). Two missed cleavages were allowed for trypsin. The precursor and fragment ion mass tolerances were set to 20 ppm. Carbamidomethyl (C) was set as a fixed modification, whereas N-acetylhexosamine addition to arginine (Arg-GlcNAc), deamidation (NQ), and oxidation (M) were selected as variable modifications. The estimated false discovery rate (FDR) of peptide identification was less than 1%.
HCT116 cells were cultured under hypoxia for 12 h, and then infected with the bacterial strain ΔnleBE+vector or ΔnleBE+pNleB for 4 h. Cells were counted, and approximately 107 cells per sample were spun down at 300 g for 3 min. Cells were washed twice with PBS and then snap-frozen on dry ice and stored at −80°C until analysis by Metabolon. After the addition of 1000 μL of extract solvent (acetonitrile-methanol-water, 2:2:1, containing internal standard), the samples were vortexed for 30 s, homogenized at 45 Hz for 4 min, and sonicated for 5 min in an ice-water bath. The homogenization and sonication cycle were repeated 3 times, followed by incubation at −20°C for 1 h and centrifugation at 12,000 rpm (4°C) for 15 min. The resulting supernatants were transferred to LC-MS vials and stored at −80°C until UHPLC-QE Orbitrap/MS analysis. The quality control (QC) sample was prepared by mixing an equal aliquot of the supernatants from all of the samples. LC-MS/MS analyses were performed using an UHPLC system (1290, Agilent Technologies) with a UPLC HSS T3 column (2.1 mm × 100 mm, 1.8 μm) coupled to Q Exactive (Orbitrap MS, Thermo). The mobile phase A was 0.1% formic acid in water for positive, and 5 mmol/L ammonium acetate in water for negative, and the mobile phase B was acetonitrile. The elution gradient was set as follows: 0 min, 1% B; 1 min, 1% B; 8 min, 99% B; 10 min, 99% B; 10.1 min, 1% B; 12 min, 1% B. The flow rate was 0.5 mL/min. The injection volume was 3 μL. The QE mass spectrometer was used for acquiring MS/MS spectra on an information-dependent basis (IDA) during LC/MS experiments. In this mode, the acquisition software (Xcalibur 4.0.27, Thermo) continuously evaluates the full scan survey MS data. ESI source conditions were set as follows: sheath gas flow rate as 45 Arb, Aux gas flow rate as 15Arb, capillary temperature 320°C, full ms resolution as 70,000, MS/MS resolution as 17,500, collision energy as 20/40/60 eV in NCE mode, spray voltage as 3.8 kV (positive) or −3.1 kV (negative), respectively.
RCC4 and 786-O cells were infected with EPEC E2348/69 and it derivatives for 2–4 h. RCC4 or 786-O cells were grown under normal conditions with or without 100 μM 2-NBDG (Invitrogen) for 1.5 h. Fluorescence was measured with a fluorescence-activated cell sorting (FACS) analyzer.
Control short hairpin RNA (luciferase shRNA) and HIF-1α short hairpin RNAs (HIF-1α shRNA-1 and HIF-1α shRNA-2) were cloned into the lentivirus vector lentiLox3.7. The shRNA target sequences were as follows: control shRNA (luciferase shRNA), 5’-GTTGGCACCAGCAGCGCAC-3’; HIF-1α shRNA-1, 5’-GAGCTTGCTCATCAGTTGC-3’; HIF-1α shRNA-2, 5’-GGGTTGAAACTCAAGCAAC-3’.
For knocking-down HIF-1α in HCT116 cells, control shRNA vector, HIF-1α shRNA-1 vector, or HIF-1α shRNA-2 vector were transiently transfected into HCT116 cells respectively. For knocking-down HIF-1α in RCC4 cells, due to the extreme low efficiency for transfection in RCC4 cells, lentiviruses were produced. Lentivirus of HIF-1α shRNAs and control (luciferase) shRNA were generated by transfecting HEK293T cells with transducing vector packaging vectors VSVG, RSV-REV, and pMDL g/p RRE together with control shRNA vector, HIF-1α shRNA-1 vector, or HIF-1α shRNA-2 vector. After transfection for 48 h, virus particles in the medium were harvested, filtered, and transduced into target cells.
RCC4 cells and 786-O cells were transduced with lentivirus encoding control shRNA or HIF-1α shRNAs in the absence or presence of bacterial strains for 3 h, and ATP concentrations were measured by an ATP assay kit (Beyotime) following the manufacturer’s instructions.
Male C57BL/6 mice, 5–6 weeks old, were individually maintained in high-efficiency particulate air (HEPA)-filtered cages with autoclaved food and water. Mice were randomized into each experimental group without investigator blinding. All mice were male and of the same size distribution (17–19 g).
Deletion of the gene encoding NleBc in C. rodentium strain DBS100 (ATCC51459; ATCC) and its derivatives were provided by Feng Shao (S1 Table). For oral inoculation, C. rodentium wild-type strain and its derivatives were prepared by shaking the bacterial culture overnight at 37°C in LB broth. Mice were orally inoculated using a gavage needle with 200 μl bacterial suspension in PBS (~4 × 109 CFU or ~2 × 109 CFU).
Mice were weighed every 2 days and feces collected every 2 or 4 days for enumeration of CFU (~2 × 109 CFU), and the number of viable bacteria per gram of feces was determined by plating serial dilutions onto LB agar containing the appropriate antibiotics. Eight days after inoculation, colons were removed aseptically, weighed, and diluted in PBS. The serial dilutions were plated to determine CFU counts. Colonization data were analyzed using Student’s t test (GraphPad Prism 5.0). P < 0.05 was considered significant.
For western blot assays of mouse colon, the distal portion of the colon from the caecum to the rectum was removed from infected mice. For immunofluorescent staining of mouse colon, colon sections were incubated with anti-Slc2a1(glut1) antibody diluted in PBS (1:100).
For PX-478 treatment, PX-478 powder (Selleck) was dissolved in PBS (20 μg/μl, stock solution). The stock solution was diluted to 1.2 μg/μL (working solution), and a total of 500 μL of the working solution (600 μg/each mouse) was injected intraperitoneally (30 μg/g; final injected dosage for each mouse was based on a 20 g body weight). Two days after PX-478 injection, mice were orally inoculated using a gavage needle with 200 μL bacterial suspension in PBS (~4 × 109 CFU). Pre- and post-fasting (6 h) blood glucose was measured.
Two days after inoculation of bacteria (~4 × 109 CFU), blood glucose was obtained by tail bleeding before fasting, and glucose was measured using an automatic glucose monitor (Ultra One Touch) (before fasting). Then, mice were fasted for 6 h and blood glucose was measured again after fasting. Two independent experiments were performed using 3 or 4 mice per group.
All independent experiments carried out in this study and indicated in the figure legends were biological replicates. All results are presented as means + SEM. Unless otherwise noted, p-values were calculated using unpaired t tests (GraphPad Prism 5, GraphPad Software Inc.).
|
10.1371/journal.pgen.1002086 | STAT Is an Essential Activator of the Zygotic Genome in the Early
Drosophila Embryo | In many organisms, transcription of the zygotic genome begins during the
maternal-to-zygotic transition (MZT), which is characterized by a dramatic
increase in global transcriptional activities and coincides with embryonic stem
cell differentiation. In Drosophila, it has been shown that
maternal morphogen gradients and ubiquitously distributed general transcription
factors may cooperate to upregulate zygotic genes that are essential for pattern
formation in the early embryo. Here, we show that Drosophila
STAT (STAT92E) functions as a general transcription factor that, together with
the transcription factor Zelda, induces transcription of a large number of
early-transcribed zygotic genes during the MZT. STAT92E is present in the early
embryo as a maternal product and is active around the MZT. DNA–binding
motifs for STAT and Zelda are highly enriched in promoters of early zygotic
genes but not in housekeeping genes. Loss of Stat92E in the
early embryo, similarly to loss of zelda, preferentially
down-regulates early zygotic genes important for pattern formation. We further
show that STAT92E and Zelda synergistically regulate transcription. We conclude
that STAT92E, in conjunction with Zelda, plays an important role in
transcription of the zygotic genome at the onset of embryonic development.
| In the initial phase of the early embryo, transcription is inactive and
development is supported by maternally derived gene products. During a time
window termed the maternal-to-zygotic transition (MZT), the maternal gene
products are degraded and the zygotically expressed genes required for
embryogenesis initiate their transcription. How the dramatic upregulation of
zygotic genes during the MZT is achieved is not completely understood, although
it has been shown that the transcription factor Zelda plays a critical role. In
this manuscript, we show that Drosophila STAT (STAT92E)
functions as a general transcription factor that, together with Zelda, induces
transcription of a large number of early-transcribed zygotic genes during the
MZT. We further show that STAT92E and Zelda synergistically regulate
transcription. Thus, multiple transcription factors, such as STAT92E and Zelda,
cooperate to control transcription of the zygotic genome at the onset of
embryonic development.
| Embryonic pattern formation is a complex and progressive process. In many
multicellular organisms, the initial period of embryogenesis relies on gene products
inherited from the mother. In Drosophila, maternally derived
morphogen proteins form broad gradients along the major body axes to define body
polarities [1]–[3]. Zygotic transcription
begins during the maternal-to-zygotic transition (MZT), which is characterized by a
decline in maternal mRNA levels and a dramatic increase in a large number of zygotic
transcripts [4], [5]. Many of the zygotic genes transcribed the earliest,
exhibit region-specific patterns. For instance, the “gap genes”, such as
zygotic hunchback (hb), Krüppel (Kr),
knirps (kni), and tailless (tll) are transcribed
zygotically in broad and mostly non-overlapping domains along the anteroposterior
(A/P) body axis. The boundaries of these zygotic genes are determined by morphogen
gradients that are set up by maternal gene products, such as Bicoid (Bcd) and
maternal Hb [2], [3]. Additional zygotic genes, mostly transcription factors,
are induced in more refined embryonic regions as a result of cooperation between the
maternal morphogens and gap gene products. The combinatorial input of different
transcription factors at different positional coordinates results in expression of
thousands of zygotic genes in an increasingly refined pattern, leading to cell fate
determination and differentiation [1]–[3], [6].
To date, only a few transcription factors have been implicated in transcription of
the zygotic genome during the MZT. For example, the maternal morphogens Bcd and
Dorsal activate target genes along the anteroposterior (A/P) and dorsoventral (D/V)
axis, respectively [7], [8]. The dramatic increase in gene expression that occurs
during the MZT raises the possibility that additional unidentified transcription
factors are involved in the rapid initiation and maintenance of the heightened
levels of zygotic gene transcription that characterize the MZT. It has been proposed
that the few known regionally localized transcription factors, such as Bcd and
Dorsal, act in conjunction with ubiquitously present factors to induce and maintain
expression of a large number of zygotic genes in cell type-specific patterns. This
idea is supported by the identification of a ubiquitous factor encoded by
zelda (zld; a.k.a. vielfaltig
or vlf) [9], and further by the demonstration that combining Dorsal
with Zelda- or STAT-binding sites supports transcription in a broad domain in the
embryo [10].
To identify additional ubiquitous transcription factors that are important for
transcription of the zygotic genome during the MZT, we first conducted in
silico analyses, taking advantage of the large amount of information
available in public databases on transcriptional regulation of zygotic genes
expressed during early embryogenesis in Drosophila. This approach
led to the identification of STAT92E, in addition to Zelda, as a plausible
transcription factor important for the upregulation of multiple genes during the
MZT. Global expression profiling studies indicate that loss of STAT92E, similarly to
loss of Zelda, preferentially causes down-regulation of zygotic genes essential for
early embryogenesis. We further demonstrate that STAT92E is indeed involved in
transcription of the developmentally important genes dpp, tailless
(tll), and Kr during early embryogenesis. Our
results suggest that STAT92E is essential for upregulation of a multitude of
zygotically transcribed genes during the MZT, and thus is important for transition
of the early embryo from a totipotent embryonic stem cell state to a state of
cellular differentiation.
To identify general transcription factors that are required for transcription of
a large number of zygotic genes at early embryonic stages, or during the MZT, we
performed a meta-analysis to search for candidate transcription factors required
for activation of multiple zygotic genes. To this end, we first selected a list
of developmentally important zygotic genes transcribed during the MZT (referred
to as “zygotic genes”), whose expression patterns altogether cover
the entire embryo, and whose transcriptional activation has previously been
studied. We analyzed a total of 21 early zygotic genes, including the gap genes:
hunchback (hb), huckebein (hkb),
Giant (Gt), Krüppel (Kr), knirps (kni), and tailless
(tll); the pair-rule genes: even skipped (eve), fushi
tarazu (ftz), hairy (h), odd paired (opa), paired (prd), sloppy paired 1
(slp1), and runt (run); the segmental polarity and
other genes: engrailed (en) and Sex lethal
(Sxl), as well as genes expressed along the D/V axis:
decapentaplegic (dpp), zerknüllt (zen), rhomboid (rho), short
gastrulation (sog), snail (sna), and twist
(twi).
As a second step, for each of these genes, we searched Flybase (http://flybase.org) and PubMed (http://www.ncbi.nlm.nih.gov), and compiled a list of all
currently known or potential transcriptional activators or signaling pathways
involved in their transcriptional induction (Table S1).
We used the RedFly database (http://redfly.ccr.buffalo.edu) [11] to obtain a list of
experimentally verified transcription factor binding sites for each target gene,
and the FlyEnhancer program (http://genomeenhancer.org/fly) [12] to search for the
presence of particular transcription factor binding sites in the promoter region
(defined as 4 kb upstream of the transcriptional start site) of all the target
genes. Based on these search results, we assigned activation scores to the
putative or known transcriptional activators to reflect their importance in the
expression of a particular zygotic gene (Table S1). These scores were added to obtain
a cumulative score for each activator (Figure 1A; Table S2).
The connections between activators and their target genes are represented in an
activation map (Figure
1B).
The top seven activators identified, in descending order of cumulative
interaction score, were Zelda (Zld), Bicoid (Bcd), STAT92E, Torso, Caudal (Cad),
Dorsal, and Twist (Twi) (Figure
1A; Table S2). Zelda has previously been shown to be a key transcription
activator of the early zygotic genome [9], validating our bioinformatic
approach. Both Bcd and Cad are maternal-effect gene products that form gradients
along the A/P axis in the early embryo [7], [13], [14]; Torso signaling is
activated only at the anterior and posterior poles, and the specific
transcriptional activators that it regulates remain unidentified [15]–[17]; Dorsal and Twi are active only in the ventral region
of the embryo [18]. On the other hand, STAT92E is ubiquitously
distributed in the early embryo as a maternal product [19] and is activated early [20], and thus has
the potential to act more universally. STAT92E is the transcriptional activator
mediating the JAK/STAT (Hop/STAT92E) pathway [19], [21], [22], and also participates in Torso
signaling [23]–[25]. Thus, we decided to investigate whether STAT92E acts
as a general transcriptional regulator during early embryogenesis, similar to
Zelda.
To test whether STAT92E is important for transcription of early “zygotic
genes”, we first assessed the occurrence of consensus STAT92E binding
sites (TTCnnnGAA) in the promoter region, defined as 4 kb genomic sequence
upstream of the transcription start site, of the 21 zygotic genes in this study.
The Drosophila genome is slightly AT-rich, with 57.4% AT
and 42.6% GC base pairs [12]. Thus the probability
for A or T to occur at any position is 0.287, and for G or C is 0.213, and the
probability (p) for random occurrence of one STAT binding site
(with 6 fixed nucleotides) at any position is 3.08x10−4 (0.2874x0.2132),
and its frequency of occurrence within the 4 kb upstream regulatory regions of
21 genes (n = 84,000 bp) at random is 25.9
(np; expected value). However, when we searched for STAT
binding sites within the 4 kb upstream region of the 21 zygotic genes, we found
43 in total (observed value) (Figure 1C). Assuming the actual occurrence of STAT-binding sites
exhibits Binomial distribution with a probability of 3.08x10−4, the
standard deviation (σ) should be 5.1. The difference between the observed
(43) and expected (25.9) values is 17.1, which is beyond three standard
deviations (Z = 3.29;
p = 0.001).
In contrast, when we searched for STAT-binding sites within a 4 kb window
upstream of the transcription start site of 21 housekeeping genes (defined as
ubiquitously expressed, both maternally and zygotically, with generally cellular
metabolic or structural functions), including rp49, GAPDH,
Actin5C, and those encoding ribosomal proteins and RNA polymerases,
we found a total of 13 STAT-binding sites (Figure 1D), which is significantly lower than
the expected 25.9 sites (Z = 2.48;
p = 0.013). (A total of 78 housekeeping genes and the
numbers of STAT-binding sites in their upstream regions are listed in Table S3.)
Moreover, many of the STAT-binding sites in the upstream regions of the 21
zygotic genes are clustered (defined by two sites occurring within 500 bp),
which is characteristic of functional transcription factor binding sequences
[12],
[19], [25], [26] (Figure 1C), whereas in the
promoter regions of the 21 housekeeping genes, the STAT-binding sites occur as
single sites (Figure 1D;
Table
S3).
It has been shown that Zelda-binding sites (the TAGteam motif) are enriched in
the promoter regions of “zygotic genes” [9], [27]. We examined the
distribution of Zelda-binding sites in the promoter regions of the 21 zygotic
and housekeeping genes, respectively. Consistent with the previous report [9], [27] and
similar to STAT-binding sites, we found that Zelda-binding sites are similarly
enriched in the promoters of the zygotic and very infrequently in the
housekeeping genes (Figure 1C,
1D). Since the enhancers for many of the early zygotic genes are not
localized in the upstream promoter regions, we also searched for STAT and
Zelda-binding sites in the promoter-distal enhancers for these 21 zygotic genes,
and found that promoter-distal enhancers are not enriched for STAT-binding sites
(Z = 0.63; p = 0.736), but are
significantly enriched for Zelda-binding sites (Z = 3.13;
p = 0.0017) (Figure S1). Such a result suggests that
STAT92E might differ from Zelda and might not be important for regulating
promoter-distal enhancers, which usually control spatial expression patterns.
Nonetheless, our studies indicate that DNA-binding sites for both STAT and Zelda
are enriched in the upstream promoter regions of the 21 zygotic genes that are
highly transcribed during the MZT, but are underrepresented in the housekeeping
genes that are ubiquitously transcribed. This observation is consistent with the
finding that Zelda is required specifically for expression of “zygotic
genes” at the MZT [9], raising the possibility that STAT may play a similar
role.
To determine whether STAT92E functions as a general transcriptional activator of
the zygotically expressed genes in the early embryo, we determined the
expression profiles of early stage embryos (corresponding to nuclear division
cycle 8–14, a time window for the MZT) of wild-type control and of those
lacking the maternal Stat92E gene products (referred to as
Stat92Emat–; see Methods) at the same
stage.
We found that in Stat92Emat– embryos, 657 genes
were down regulated and 558 genes up-regulated by at least 1.5 fold, compared
with wild-type control (Figure
2A). In Stat92Emat– embryos, genes
exhibiting >1.5 fold change in expression constituted 8.9% of all
genes (n = 13,615) on the Gene Chip, while the majority
(91.1%) of the genes exhibited no significant changes (Figure S2).
Consistent with the idea that STAT92E is preferentially required for expression
of “zygotic genes”, the vast majority (78.2%) of the
down-regulated genes in Stat92Emat– embryos
were “zygotic genes” (Figure 2B, left; Table S4). In contrast, the up-regulated
genes contained more maternally expressed than zygotically expressed genes
(Figure 2B, right; Table S5).
This observation is reminiscent of gene expression profiles of
zld mutant embryos at the same stage, in which more
“zygotic genes” than maternal genes are down-regulated [9]. By comparing
the two sets of genes, we found that >50% of the “zygotic
genes” that were down-regulated in
zldmat– embryos (67/120) were also
down-regulated in Stat92Emat–embryos,
suggesting that these genes might be co-regulated by STAT and Zelda (Table
S4).
Consistent with the observed difference in the abundance of STAT-binding sites
present in their promoter regions, the 21 zygotic genes (except for
hb) were all significantly down-regulated, with a 4.3 fold
down-regulation on average, whereas the 21 housekeeping genes showed no
significant changes in expression, with the exception of DNase II (Figure 2C), in
Stat92Emat– embryos. Similar to
Stat92Emat– embryos, in
zldmat– embryos, many of these 21
zygotic genes were also significantly down-regulated, whereas the housekeeping
genes were not significantly changed [9], suggesting that STAT92E and
Zelda may both be important for transcription of early zygotic genes. Expression
profiling experiments indicate that STAT92E and Zelda do not transcriptionally
regulate each other (Liang et al., 2008; this study). We further performed
qRT-PCR experiments and found that Zelda mRNA levels were indeed not
significantly changed in Stat92E loss-of-function or
hop gain-of-function mutants (Figure S3),
suggesting that STAT92E does not indirectly control zygotic gene activation by
affecting Zelda levels.
Finally, we tested expanded sets of zygotic and housekeeping genes to include
>40 genes in each set (Table S6) using the Gene Set Enrichment
Analysis (GSEA) software (http://www.broadinstitute.org/gsea/index.jsp), which is a
computational method that determines whether an a priori
defined set of genes shows statistically significant, concordant differences
between two biological states (e.g., mutant versus wild-type) [28].
Indeed, by subjecting our microarray data to GSEA analysis, we found that the
“zygotic genes” were highly significantly down regulated
(p = 0.00), whereas the housekeeping genes were
insignificantly changed (p = 0.44), in
Stat92Emat– embryos when compared with
wild-type control (Figure
2D). Thus, similar to Zelda, STAT92E is preferentially required for
transcription of “zygotic genes”.
To validate our gene profiling results from the microarray studies, we
investigated the effects of over-activation and loss of STAT92E on transcript
levels of a number of early “zygotic genes”. We chose to examine
expression levels of dpp, Kr, tll, and eve,
four early zygotic genes whose promoter regions contain STAT-binding sites and
whose expression domains span broad and distinct regions of the early embryo
(see below).
We first examined mRNA levels of dpp, Kr,
tll, and eve in the early embryo
(1–2 h after egg laying) using semi-quantitative reverse-transcription
polymerase chain reaction (RT-PCR) in Stat92E gain- or
loss-of-function genetic backgrounds. We found that in hopGOF
embryos, in which STAT92E is overactivated [29]–[31], mRNA of these
four genes were all expressed at significantly higher levels relative to
wild-type; whereas in Stat92Emat– embryos,
these four genes were expressed at approximately 50% of the wild-type
levels (Figure 3A, 3B).
Moreover, reducing the dosage of zelda by half in
Stat92Emat– embryos caused further
reductions in the transcript levels of dpp,
Kr, tll, and eve
(zelda+/–;
Stat92Emat– in Figure 3A, 3B). We examined
zelda+/–;
Stat92Emat– embryos only, because it was
technically not possible to examine embryos lacking both Zelda and Stat92E. We
further confirmed the expression results by quantitative real-time PCR (Figure 3C). These results were
consistent with the microarray data, which suggested that Stat92E and Zelda may
co-regulate transcription of many “zygotic genes”.
We next investigated whether STAT92E binds to the putative STAT-binding sites in
the respective promoter regions of dpp, Kr,
and tll using chromatin immunoprecipitation (ChIP) experiments
with early embryo extracts using anti-STAT92E antisera. Binding of STAT92E to
the eve enhancer and of Zelda to the TAGteam sequences enriched
in “zygotic genes” have been previously shown [9], [19], [21]. Using primers flanking the
putative STAT-binding sites in these promoter regions, we detected STAT92E
binding to the promoter regions dpp, Kr, and
tll (Figure
3D). The results from RT-PCR and ChIP studies were consistent with
the bioinformatic and gene profiling studies shown above, suggesting that
STAT92E, likely together with Zelda, regulates the transcription of early
“zygotic genes” in vivo.
Having shown that STAT92E regulates expression levels of early “zygotic
genes”, and that STAT92E binds to the consensus STAT-binding sites present
in the promoter regions of dpp, Kr, and tll,
we next investigated whether these consensus STAT-binding sites are indeed
essential for mediating STAT92E transcriptional activation, and whether STAT92E
and Zelda cooperate to regulate “zygotic genes”, as it has
previously been shown that Zelda is essential for expression of dpp, Kr,
tll, and eve, among others, in the early embryo
[9]. We
carried out reporter gene assays in Drosophila S2 cells (Figure 4A).
We first tested whether activated STAT92E binds to the promoter regions of
dpp, Kr, tll, and
eve in S2 cells as it does in early embryos (see Figure 3C). We transfected a
V5-tagged STAT92E into S2 cells and performed ChIP assays. STAT92E activation in
S2 cells was achieved by co-expressing Hop, which phosphorylates and activates
STAT92E when over-expressed (Figure
4B). By immunoprecipitation with anti-V5 antibody, we found that
co-transfection with Hop leads to an enrichment of STAT92E binding to the
endogenous dpp promoter (Figure 4C, lane 3). Activation of JAK/STAT
signaling thus induces a stronger association of STAT92E with the
dpp promoter, consistent with the idea that STAT92E
directly regulates dpp expression. However, the same ChIP
experiments failed to detect association of STAT92E with the Kr,
tll, or eve promoter in S2 cells, in contrast to
the ChIP results in early embryos (see Figure 3C), suggesting that the epigenetic
states of these promoter sequences may be different in S2 cells than in early
embryos. We thus focused on the dpp promoter for reporter gene
analysis. To this end, we isolated a 1.3 Kb dpp promoter
fragment (Figure 4A; Figure S4),
which contains the two clustered STAT92E binding sites we had tested in ChIP
experiments (see Figure 3C,
Figure 4C).
To test whether the STAT-binding sites in the dpp promoter are
important for JAK/STAT-induced dpp expression, we made reporter
genes by fusing a wild-type dpp promoter fragment (WT), or a
mutant version with both STAT-binding sites mutated (DM), with an enhanced
yellow fluorescent protein (EYFP), and transfected S2 cells (Figure 4A). In order to
activate reporter gene expression, we first treated the cells with H2O2/vanadate
(pervanadate), which causes rapid and efficient STAT92E phosphorylation [32], [33] (Figure S5A)
and is more efficient than transient transfection of hop in
activating STAT. We found that, indeed, EYFP was expressed 1.5 hours after
pervanadate treatment in S2 cells transfected with the wild-type (WT), but not
the double mutant (DM) construct (Figure S5B), indicating that these
STAT92E-binding sites are important for phosphorylated STAT92E-induced reporter
gene expression.
To more accurately quantify transcription from the dpp promoter
with or without the two STAT-binding sites, we replaced EYFP with luciferase in
the reporter constructs to obtain dppWT-luc and
dppDM-luc, respectively. In addition, we
used Hop and STAT92E co-transfection, instead of pervanadate, to ensure specific
activation of STAT92E. In the presence of co-transfected Hop and STAT92E, we
detected an increase in luciferase activity in S2 cells tranfected with
dppWT-luc to more than 20 fold when measured
72 hours after transgene expression, and this increase was abolished when
dppDM-luc was used in the assay, which
showed much less pronounced increase (Figure 4D). These results further
substantiate our finding that STAT92E-mediated activation of
dpp requires the two STAT92E binding sites.
It has previously been shown that transcription of dpp is
significantly down-regulated in the absence of Zelda [9], and that Zelda-binding sites
are present in the dpp promoter region (Figure 1C; Figure 4A; also see [9]). To test whether Zelda binds
to the putative site in the dpp reporter gene, we carried out
ChIP assays in S2 cells after transfecting a Zelda-Flag plasmid. Indeed, we
detected Zelda binding to the dpp promoter region using an
anti-Flag antibody and ChIP assay (Figure 4E).
We next investigated the role of Zelda in dpp transcription
using dppWT-luc and a mutant promoter fragment with
the Zelda-binding site and the two STAT-binding sites mutated (designated as
dpp™-luc as it bears triple
mutations; Figure 4A). To
evaluate whether Zelda and STAT cooperate in regulating dpp
transcription, we co-transfected S2 cells with STAT92E (together with Hop to
achieve STAT activation) or Zelda, or both STAT92E (with Hop) and Zelda, in the
presence of dppWT-luc or
dpp™-luc, and carried out luciferase
assays. When assayed at 72 h after induction of transgene expression, we found
that STAT activation alone induced dppWT-luc
transcription by 22 fold, and Zelda alone caused upregulation of
dppWT-luc by 48 fold, whereas in the
presence of both Zelda and activated STAT, dppWT-luc
was up-regulated by 230 fold (Figure 4F). Mutating STAT and Zelda binding sites prevented the
dramatic increase in transcription as measured by luciferase activity (Figure 4F). These results
suggest that Zelda and STAT have synergistic effects on
dppWT-luc transcription. Interestingly, an
increase in luciferase activity was observed even when binding sites for STAT or
Zelda, or both, were mutated, albeit to a much less pronounced level than with
the wild-type promoter (Figure 4D,
4F), suggesting that there might be other cryptic binding sites
present in the promoter, or that other molecules were activated by
over-expressed JAK or Zelda.
The apparent synergy between STAT92E and Zelda could explain the results from the
gene profiling experiments. Microarray results show that embryos without STAT92E
(in which Zelda presumably remains active) exhibit a 3.1 fold decrease in
dpp expression (Figure 2B), and that Zld
mutant embryos (in which presumably STAT92E is still active) have reduced
dpp expression by 5.7 fold [9]. These data suggest that in
the early embryo either Zelda or STAT activation could induce
dpp transcription to a limited extent, whereas the presence
of both Zelda and STAT activation synergistically promote dpp
transcription.
Having shown that STAT92E, possibly acting synergistically with Zelda, is
important for expression levels of many early “zygotic genes”, we
next investigated whether loss of STAT92E also affects the spatial expression
patterns of the early “zygotic genes”. We examined the expression of
dpp, Kr, and tll in the
early embryo, by in situ hybridization, while the effects of
Stat92E mutation on eve expression have
previously been documented [19], [21]. These genes are expressed in distinct spatial
domains that altogether cover nearly the entire early embryo (see below).
The dpp expression domain spans nearly the entire A/P axis in
the dorsal regions of the early embryo [34]-[37] (Figure 5A). It has been shown that
dpp transcription in the ventral region is repressed by
Dorsal, a Rel family transcription factor [38], and that general
transcription factors, such as Zelda and STAT, are responsible for
dpp expression in the dorsal region ([9]; this study). By employing
in situ hybridization, we found that compared to wild type,
the overall level of dpp mRNA is much reduced in
Stat92Emat– embryos, especially in the posterior pole
region (Figure 5B).
Moreover, we found that JAK/STAT signaling also regulates dpp
expression during late embryogenesis (Figure S6). These results are consistent with
previous findings in other developmental contexts [39], [40] as well as with the above
microarray results and mRNA measurements (Figure 2, Figure 3A–3C).
Kr is expressed in the central region of the early embryo [41] (Figure 5C). Other than the
maternal morphogens Bcd and Hb, it is not known whether additional factors
contribute to Kr transcriptional activation. We found that in
Stat92Emat– embryos, although the overall expression
pattern of Kr mRNA was little changed, its levels were reduced
(Figure 5D), consistent
with the microarray and qPCR results.
tll is expressed in two domains along the A/P axis-the anterior
and posterior pole regions [42] (Figure
5E). The Torso pathway controls tll expression by
antagonizing its repressors [17], [43]; the identity of transcriptional activators of
tll remains obscure, although STAT92E has been speculated
to contribute to tll expression [25]. We have previously shown that
STAT92E is essential for the expansion of tll expression
domains caused by Torso, over-activation, but not for the extent of
tll spatial expression domains under normal conditions
[25]. In
addition, we have previously shown that there are two consensus STAT binding
sites in the tll promoter region that are particularly
important for Torso overactionvation-induced ectopic tll
expression [25].
In light of our finding that STAT92E is important for the expression levels of
dpp, Kr, and tll, we
reexamined the role of STAT92E in endogenous tll expression in
Stat92Emat– and wild-type control
embryos by in situ hybridization done under identical
conditions. We found that, similar to dpp and
Kr mRNA, while the spatial patterns of tll
expression were not dramatically changed as previously shown [25], the overall
levels of tll mRNA were significantly reduced in
Stat92Emat– embryos (Figure 5F).
Taken together, the above results indicate that loss of STAT92E led to much
reduced expression levels of dpp, Kr, and
tll, without affecting their spatial expression domains.
Similarly, it has been shown that loss of STAT92E results in reductions, but not
complete loss of, eve stripe 3 and 5, without affecting the
overall spatial expression pattern of eve
[19], [21]. Thus, STAT92E
is likely required for regulating the expression levels of early “zygotic
genes”, but not for controlling their spatial patterns.
Finally, we investigated the biological consequences of reducing expression
levels, without altering spatial domains, of multiple zygotically expressed
early genes, as with loss of STAT92E. The correct expression of the early
zygotic genes during the MZT is essential for formation of different tissues and
body parts at the correct positions, i.e., pattern formation [1]–[3]. Pattern formation
in Drosophila can be conveniently visualized by examining the
exoskeleton (cuticle) morphology of the larva or late embryo [1]–[3].
In the wild-type cuticle (Figure
6A), anteroposterior (A/P) polarity is defined by the head skeleton
and three thoracic segments in the anterior, followed by the abdominal segments,
and the posterior and terminal structure, consisting of the 8th
abdominal segment and the Filzkörper (Figure 6A; Arrow). Dorsoventral (D/V)
polarity can easily be seen by the positions of the eight abdominal denticle
belts, which form in the ventral region, while bare cuticle marks the dorsal
region (Figure 6A). Removal
of STAT92E from the early embryo resulted in heterogeneous defects, mostly
notably along the A/P axis as seen in the larval cuticles, which were missing
part or all of A3, A4, A5, and A8 to various degrees (Figure 6B; also see [19], [25]). Thus, loss of STAT92E, which
significantly reduces multiple early “zygotic genes” but does not
completely eliminate their expression (see Figure 5), leads to heterogeneous patterning
defects, consistent with defects in multiple pathways.
To understand the role of STAT92E in individual signaling pathways important for
pattern formation, we investigated whether loss of STAT92E could further
compromise pattern formation in sensitized genetic backgrounds. To this end, we
examined cuticles of Stat92Emat– embryos that
were also heterozygous for tll, Kr, or
dpp, and indeed found patterning defects (see below).
The gap gene tll is essential for the development of terminal
structures [17],
[42], and
tll mutant homozygous embryos do not have A8 and the
Filzkörper (Figure 6C).
tll heterozygous flies, in contrast, are perfectly viable
and normal, with cuticles indistinguishable from wild-type controls, according
to our own observation. In the absence of STAT92E, however, we found that
tll+/– embryos were missing the
terminal structures (A8 and Filzkörper) (Figure 6D). This suggests that without
STAT92E, a half dose of tll+ is no longer
sufficient for development, consistent with the idea that STAT92E is partially
required for tll transcriptional output.
Kr is required for development of the thoracic and anterior
segments, and these segments are missing in
Kr–/– embryos (Figure 6E; also see [44]).
Kr+/– embryos are mostly normal
but have subtle anterior defects (Figure S7; also see [44]). In the absence of
STAT92E, however, we found that Kr+/–
embryos were missing a large area of the thoracic and anterior regions (Figure 6F), suggesting a
haploinsufficiency in the absence of STAT92E, similar to what we observed for
tll.
The dorsally expressed dpp specifies dorsal cell fates and is
crucial for the dorsoventral polarity of the embryo, which is reflected in the
cuticle by the presence of naked cuticles in the dorsal region and eight
abdominal denticle belts in the ventral region (Figure 6A) [37]. Notably, although
dpp expression was significantly reduced in
Stat92Emat– embryos (Figure 2, Figure 3A–3C, Figure 5B), they did not exhibit gross D/V
polarity defects (Figure
6B), suggesting that the residual dpp transcripts
present in Stat92Emat– embryos are sufficient
for specifying dorsal cell fates, or that the reduction in dpp
expression is compensated for by a reduction in a dpp
antagonist that is also regulated by STAT92E. Despite the fact that
dpp is haploinsufficient for viability,
dpp heterozygous embryos exhibit normal D/V polarity, with
clearly discernable ventral denticle belts and bare dorsal cuticles (Figure S7),
suggesting that a half dose of dpp+ suffices
for D/V patterning (also see [37]. Embryos homozygous for dpp,
nonetheless, are completely “ventralized,” having denticle belts
that extend into the dorsal region to surround the entire D/V axis (Figure 6G; also see [36], [37]). The
combination of Stat92Emat– and
dpp heterozygosity caused partial ventralization of the
embryo; in 13% of Stat92Emat–;
dpp+/– embryos
(n = 11/86), the posterior-most denticle belt extended
significantly dorsally to cover approximately 80% of the circumference
(Figure 6H, arrow).
Similar ventralization defects were never observed in
Stat92Emat– and
dpp+/– embryos (n>500). Thus,
in the absence of STAT92E, a half dose of dpp is no longer
sufficient for dorsoventral patterning, consistent with the notion that STAT92E
normally regulates dpp expression levels.
In summary, loss of STAT92E caused heterogeneous patterning defects, as revealed
by varying cuticle defects, consistent with an insufficiency of multiple
pathways. A further reduction in the dosage of genes in different pathways, such
as tll, Kr, and dpp,
uncovered the role of STAT92E in regulation of specific early zygotic genes
important for pattern formation.
We have undertaken a bioinformatics approach to investigating the mechanisms
controlling transcription of the zygotic genome that occurs during the MZT, and have
identified STAT92E as an important general transcription factor essential for
up-regulation of a large number of early “zygotic genes”. We have
further investigated the role of STAT92E in controlling transcription of a few
representative early zygotic genes, such as dpp,
Kr, and tll, that are important for pattern
formation and/or cell fate specification in the early embryo. Our studies suggest
that STAT92E cooperate with Zelda to control transcription of many “zygotic
genes” expressed during the MZT. While STAT mainly regulates transcription
levels, but not spatial patterns, of dpp, tll, and
Kr, and possibly also other “zygotic genes”, Zelda
is essential for both levels and expression patterns of these genes [9].
The transcriptional network that controls the onset of zygotic gene expression during
the MZT has remained incompletely understood. It has been proposed that
transcription of the zygotic genome depends on the combined input from maternally
derived morphogens and general transcription factors. The former are distributed in
broad gradients in the early embryo and directly control positional information
(e.g., Bicoid, Caudal, and Dorsal), whereas the latter are presumably uniformly
distributed regulators that augment the upregulation of a large number of
“zygotic genes”. Other than Zelda, which plays a key role as a general
regulator of early zygotic expression [9], the identities of these general
transcriptional activators have remained largely elusive. It has been shown that
combining Dorsal with Zelda- or STAT-binding sites supports transcription in a broad
domain in the embryo [10]. The demonstration of STAT92E as another general
transcription factor sheds light on the components and mechanisms of the controlling
network in the early embryo. Moreover, we have found that STAT92E and Zelda may
cooperate to synergistically regulate “zygotic genes”. Our results thus
validate the bioinformatics approach as useful in identifying ubiquitously expressed
transcription factors that may play redundant roles with other factors and thus
might otherwise be difficult to identify.
Our conclusion that STAT92E is important for the levels but not the spatial domains
of target gene expression in the early embryo is consistent with several previous
reports. It has been shown that in Stat92E or hop
mutant embryos, expression of eve stripes 3 and 5 are significantly
reduced but not completely abolished [19], [21]. In addition, JAK/STAT activation is required for the
maintenance of high levels, but not initiation, of Sxl expression
during the MZT [45], [46]. Moreover, it has previously been shown that STAT92E is
particularly important for TorsoGOF-induced ectopic tll
expression but not essential for the spatial domains of tll
expression in wild-type embryos under normal conditions [25]. On the other hand, Zelda may be
important for both levels and spatial patterns of gene expression. This idea is
consistent with our finding that Zelda-binding sites are enriched in both promoter
and promoter-distal enhancers regions, whereas STAT-binding sites are enriched in
promoter regions only. It has been reported that pausing of RNA polymerase II is
prominently detected at promoters of highly regulated genes, but not in those of
housekeeping genes [47]. In light of our results that STAT and Zelda sites are
highly enriched in the early zygotic gene promoters, we suggest that these
transcription factors might contribute to chromatin remodeling that favors RNA
polymerase II pausing at these promoters.
Finally, the MZT marks the transition from a totipotent state to that of
differentiation of the early embryo. As a general transcription factor at this
transition, STAT, together with additional factors (such as Zelda [9]), is important
for embryonic stem cell differentiation. Further investigation is required to
understand the molecular mechanism by which STAT and Zelda [9] cooperate in controlling zygotic
transcription in the early Drosophila embryo. Moreover, it would be
interesting to investigate whether STAT plays similar roles in embryonic stem cell
differentiation in other animals.
All crosses were carried out at 25°C on standard cornmeal/agar medium unless
otherwise specified. Fly stocks of hopTum-l,
Stat92E6346, and
dppH46 were from the Bloomington
Drosophila Stock Center (Bloomington, IN). To generate
Stat92Emat– embryos, hsp70-flp;
FRT82B Stat92E6346/TM3 females were
crossed to hsp70-Flp; FRT82B [ovoD1,
w+]/TM3 males. Their 3rd instar larval
progeny were heat-shocked at 37°C for 2 hrs daily for 3–4 days, and
resulting adult females of the genotype hsp70-flp; FRT82B
Stat92E6346/FRT82B [ovoD1,
w+] were used to produce embryos that lack
maternal Stat92E gene products, as described in the dominant
female-sterile “germline clone” technique [48].
The following rules were used for assigning a score to known or putative
activators of each of the “zygotic genes”. We placed top importance
on genetically demonstrated activation during early embryogenesis, with such an
activator receiving an activation score of 10. For instance, Torso was assigned
a score of 10 as an activator of tll transcription based on the
reports that tll is not expressed in torso
loss-of-function mutants and is overexpressed in torso
gain-of-function mutants [17], [49]. Activators identified by biochemical/promoter
studies in early embryos or by genetic studies at other developmental stages
were assigned a score of 5. Lower scores were assigned to other less stringent
evidence of interaction, such as unconfirmed genetic screen results (5), in
vitro biochemical assays (2), or bioinformatics studies (1) (Table
S1).
Databases and programs used in this study:
Flybase (http://flybase.org); PubMed (http://www.ncbi.nlm.nih.gov); RedFly (http://redfly.ccr.buffalo.edu/); FlyEnhancer (http://genomeenhancer.org/fly).
The dpp promoter used in this study was a 1.3 kb genomic DNA
fragment including the upstream regulatory sequences and the non-coding exon 1
of the of dpp transcript A (Figure S2).
This genomic region has previously been shown to be the core promoter of
dpp
[38]. Standard
cloning was used to generate transcription fusions between the
dpp promoter and cDNAs of reporter genes, such as enhanced
yellow fluorescent protein (EYFP) and luciferase. Mutagenesis of two STAT92E
binding sites within the dpp promoter was done by PCR, and was
verified by sequencing. V5-Hop and V5-STAT92E are gifts from S.X. Hou [50].
Cuticle preparations were performed according to a standard protocol with minor
modifications. Embryos were dechorionated with 50% Clorox, washed
extensively with 0.1% Triton, mounted in Hoyer's, and photographed
using dark-field optics. In situ hybridization for detecting
dpp, Kr, and tll mRNA was
performed according to a standard protocol using digoxigenin-incorporated
antisense RNA probes made from dpp, Kr, and
tll cDNA, respectively, according to the supplier's
protocol. A standard protocol was used for antibody staining of embryos, and a
biotinylated secondary antibody and the Vectastain ABC kit (Vector Laboratories,
Inc.) were used according to the manufacturer's instructions. Stained
embryos were mounted in DAPI-containing mounting medium for accurate staging,
when necessary. Mounted embryos were photographed using Normaski optics on a
Zeiss Axioscope and images were analyzed using Photoshop or ImageJ software.
Total RNA was isolated from embryos (from flies raised at 25°C) collected at
1–2 h after egg laying (corresponding to nuclear division cycles
8–14) using trizol (Invitrogen) or the RNeasy Kit (QIAGEN) according to
the manufacturer's instructions. RNA quality was assessed using the Agilent
2100 Bioanalyzer and the RNA 6000 Nano kit (Agilent Technologies Inc., Palo
Alto, CA).
For RT-PCR analysis, first strand complementary DNA (cDNA) was generated from 5
µg of purified total RNA using Superscript III reverse transcriptase
(Invitrogen) and oligo(dT)12–18 in 50 µl total reaction volume. The
cDNA (at 1∶100 dilution) was used as template for either semi-quantitative
PCR reactions or real time PCR analysis using SYBR green based detection on a
BioRad iCycler. Reactions were carried out in triplicate, and melting curves
were examined to ensure single products. Results were quantified using the
“delta-delta Ct” method to normalize to rp49
transcript levels and to control genotypes. Data shown are averages and standard
deviations from at least three independent experiments. The following primer
pairs were used.
rp49: TCCTACCAGCTTCAAGATGAC, CACGTTGTGCACCAGGAACT.
dpp: AATCAATCTTCGTGGAGGAGCCGA, TTGGTGTCCAACAGCAGATAGCTC.
eve: TGCACGGATACCGAACCTACAACA, GTTCTGGAACCACACCTTGATCGT.
Kr: CAAGACGCACAAACGCGAACCTTA, TTGACGGTTTGCAGCCAGAAGTTG.
tll: AATACAACAGCGTGCGTCTTTCGC, ACATTGGTTCCTGTGCGTCTTGTC.
For microarray analysis, 200 ng of total RNA was used to prepare biotin-labeled
RNA using Ambion MessageAmp Premier RNA Amplification Kit (Applied Biosystems,
Foster City, CA). Briefly, the first strand of cDNA was synthesized using
ArrayScript reverse transcriptase and an oligo(dT) primer bearing a T7 promoter.
Then DNA polymerase I was used (in the presence of E. coli
RNase H and DNA ligase) to convert single-stranded cDNA into a double-stranded
DNA (dsDNA). The dsDNA was then used as a template for in vitro transcription in
a reaction containing biotin-labeled UTP and T7 RNA Polymerase to generate
biotin-labeled antisense RNA (aRNA). Twenty µg of labeled aRNA was
fragmented and fifteen µg of the fragmented aRNA was hybridized to
Affymetrix Drosophila Genome 2.0 Array Chips according to the
manufacterer's Manual (Affymetrix, Santa Clara, CA). Array Chips were
stained with streptavidin-phycoerythrin, followed by an antibody solution
(anti-streptavidin) and a second streptavidin-phycoerythrin solution, performed
by a GeneChip Fluidics Station 450.
The Array Chips were then scanned with the Affymetrix GeneChip Scanner 3000. The
microarray image data were converted to numerical data with Genespring software
(Agilent Technologies Inc., Palo Alto, CA) and normalized using the recommended
defaults. The signals from 11 perfect matched oligonucleotides for a specific
gene and 11 mis-matched oligonucleotides were used to make comparisons of
signals. Genes were identified as present when the present (P) assignment was
significant (p<0.05).
The Gene Set Enrichment Analysis (GSEA) online software (http://www.broadinstitute.org/gsea) was used to determine
whether the predetermined gene sets (e.g., zygotic versus housekeeping; see
Figure
S6) show statistically significant, concordant differences between
wild-type and Stat92Emat– embryos.
Primary antibodies used in this study include mouse anti-V5 (Invitrogen;
1∶500 for Western blots), Rabbit anti-V5 (QED; 1∶200 for
immunoprecipitation), goat anti-STAT92E (Santa Cruz; Cat# sc-15708;
affinity-purified against the N-terminus of STAT92E; 1∶200), rabbit
anti-Kr (1∶5000; a kind gift from C. Rushlow), and anti-phospho-STAT92E
(Cell Signaling Technology; 1∶1000). Common secondary antibodies were used
in whole-mount immunostaining or Western blots.
Drosophila Schneider L2 (S2) cells were cultured at 25°C in
Drosophila Serum-Free Medium (SFM; Invitrogen) supplemented
with 10% Fetal Bovine Serum (FBS; Invitrogen) and 0.5x
Antibiotic-Antimycotic (Invitrogen). Cells were cultured at
2.5×106/ml prior to transfection. Transfections were
performed with FuGene 6 (Roche) according to the manufacturer's
instructions. Cu2SO4 (Sigma) was added to the medium at a
final concentration of 0.5 mM 16 hours after transfection, and cells were
harvested 48 hours after induction. To stimulate JAK/STAT activation in S2
cells, 2 mM H2O2 and 1 mM sodium vanadate (pervanadate)
were added to the medium and cells were harvested at desired times after
treatment. Treated S2 cells were harvested in cell lysis buffer (from Cell
Signaling Tech.) for Western blotting or ChIP experiments.
ChIP experiments were carried out according to standard protocols with the
following modifications. 200 µl of early embryos (1–2 h AEL) or
1×107 S2 cells were treated with 1% formaldehyde at room
temperature for 20 min (embryos) or 10 min (cells) to crosslink protein with
chromatin/genomic DNA. Embryos or cells were homogenized and lysed in 300
µl of RIPA lysis buffer with 2 mM EDTA and protease inhibitors on ice. The
lysate was sonicated to shear the genomic DNA to lengths between 500 and 1000
bp. An aliquot (50 µl) of sonicated sample was saved as the input control.
5 µg goat anti-STAT92E (Santa Cruz, CA) or rabbit anti-V5 antibodies were
added to 200 µl experimental samples in RIPA buffer with 2 mM EDTA and
protease inhibitors, and the mixture was incubated overnight at 4°C with
rotation. Protein G beads (Sigma), pre-blocked with sonicated salmon sperm DNA
(Stratagene), were added to precipitate the antibody-bound chromatin and the
precipitate was washed extensively. After reversing the crosslink, DNA was
recovered by using a Qiagen PCR purification kit and quantified by PCR. The
following forward and reverse primers (flanking two STAT-binding sites in the
respective promoter regions) were used for PCR reactions.
dpp: AATTCCGGATAGCGCCTGG, AAAGATGGCACACGCTGGG.
Kr:
CATGCGTTTGCATACTGGAG,
CTATTCGAATCGCCCTTGTC.
tll:
AGTGCTTTGAGGTCGGAATG,
AAGAAACCGTGGTGTCCTTG.
Stat92E:
TGACTGCCCGCTTTTATACC,
CAAACGGCGGTCAATAGTTT.
|
10.1371/journal.ppat.1002931 | Identification of Alternatively Translated Tetherin Isoforms with Differing Antiviral and Signaling Activities | Tetherin (BST-2/CD317/HM1.24) is an IFN induced transmembrane protein that restricts release of a broad range of enveloped viruses. Important features required for Tetherin activity and regulation reside within the cytoplasmic domain. Here we demonstrate that two isoforms, derived by alternative translation initiation from highly conserved methionine residues in the cytoplasmic domain, are produced in both cultured human cell lines and primary cells. These two isoforms have distinct biological properties. The short isoform (s-Tetherin), which lacks 12 residues present in the long isoform (l-Tetherin), is significantly more resistant to HIV-1 Vpu-mediated downregulation and consequently more effectively restricts HIV-1 viral budding in the presence of Vpu. s-Tetherin Vpu resistance can be accounted for by the loss of serine-threonine and tyrosine motifs present in the long isoform. By contrast, the l-Tetherin isoform was found to be an activator of nuclear factor-kappa B (NF-κB) signaling whereas s-Tetherin does not activate NF-κB. Activation of NF-κB requires a tyrosine-based motif found within the cytoplasmic tail of the longer species and may entail formation of l-Tetherin homodimers since co-expression of s-Tetherin impairs the ability of the longer isoform to activate NF-κB. These results demonstrate a novel mechanism for control of Tetherin antiviral and signaling function and provide insight into Tetherin function both in the presence and absence of infection.
| Regulation of innate immunity is critical to maintain a balance between control of a perceived threat and immunopathology. The interferon induced cellular factor Tetherin has been shown to restrict budding of a broad range of enveloped viruses including the human immunodeficiency virus. Though Tetherin appears to be a bona fide viral restriction factor, additional cellular functions have been observed including an involvement in actin cytoskeleton organization in polarized cells, regulating interferon secretion and signaling through nuclear factor-kappa B (NF-κB). Our studies present a mechanism by which Tetherin function is regulated at the translational level through the production of alternatively translated isoforms. The short isoform of Tetherin was observed to be significantly more resistant to HIV-1 Vpu. In contrast, the longer isoform can induce NF-κB activity, a function lacking in the short isoform. Critical NF-κB signaling residues include a dual tyrosine motif, which is only present in the long isoform. Identification of these isoforms helps to illuminate how Tetherin functions, not only as a restriction factor, but also as a signaling molecule. These data highlight a previously unappreciated level of regulation and furthers our understanding of additional Tetherin functions.
| Tetherin (BST-2/CD317/HM1.24) is an interferon (IFN) induced, type II transmembrane glycoprotein which has been shown to function as an intrinsic antiviral factor that restricts release of a broad range of enveloped viruses, including members of the arenavirus [1], [2], filovirus [2], [3], gamma-herpesvirus [4], [5], paramyxovirus [1], retrovirus [6]–[8] and rhabdovirus [9], [10] families. Selective pressure from Tetherin on the aforementioned families of viruses is evident because many of these Tetherin sensitive viruses encode proteins that counter Tetherin function. To date, several virally encoded Tetherin antagonists have been identified, a few of which function through different mechanisms. These include HIV-1 Vpu [11], [12], KSHV K5 [4], SIV Nef [13], HIV-2 and SIV Env [14], [15], and ebolavirus GP [3].
Tetherin has an unusual topology that has been shown to be critical for function as a restriction factor. Tetherin is a lipid raft resident protein anchored to the cellular membrane by a N-terminal transmembrane domain and a C-terminal GPI anchor [16]. Localization to sites of enveloped virus budding and anchoring at both ends is needed for Tetherin function as a cellular restriction factor through a direct “tethering” mechanism which bridges the viral and host membranes [17]. The N-terminus also includes a cytoplasmic tail, which contains a tyrosine-based motif that mediates trafficking between lipid rafts at the plasma membrane and intracellular compartments, including the trans-Golgi Network (TGN), via adaptor complexes involved in clathrin-mediated endocytosis [18], [19]. Additionally, the Tetherin cytoplasmic tail has a serine-threonine-serine stretch and lysine residues that were identified as targets for HIV-1 Vpu-mediated tetherin downregulation through the recruitment of cellular ubiquitin ligase complexes [20]–[22].
A number of studies have examined the ability of Tetherin to function as a cellular restriction factor; however, there is evidence that Tetherin has additional roles that may be important both in the presence and absence of viral infection. Tetherin was originally described as a marker on differentiated B cells and was suggested to be involved in B cell development [23], [24]. It is apparent now that Tetherin is constitutively expressed on various other cells of the immune system including macrophages, plasmacytoid dendritic cells (pDCs) and T lymphocytes [25], [26]. Tetherin is also preferentially expressed on the surface of various transformed and metastatic cell types [27], [28], though whether Tetherin is actively contributing to the transformed state of the cell is unknown. Additionally, Tetherin has been proposed to act as a scaffold protein in polarized cells [29] and contribute to monocyte adhesion to endothelial cells [30]. Apart from directly restricting viral egress, Tetherin has been shown to have immunomodulatory properties. Tetherin regulates cytokine secretion in murine [25] and human pDCs. In human pDCs, Tetherin acts as a ligand for immunoglobulin-like transcript 7, which results in modulation of Toll-like receptor-mediated IFN and proinflammatory cytokine secretion [31]. Additionally, a cDNA screen identified BST-2 (aka Tetherin) as one of several uncharacterized cellular factors that could activate NF-κB [32]. However, this potential role for Tetherin in signaling has not been further explored.
Here we identify a previously undescribed short isoform of human Tetherin generated by alternative translation initiation at an in-frame codon 33 nucleotides downstream of the canonical translation start site which results in the production of two Tetherin species. The long (l-) and short (s-) Tetherin isoforms are generated in both cultured and primary human cells. Importantly, the two isoforms have distinct biological properties. s-Tetherin, which lacks 12 residues present in the long isoform, is significantly more resistant to HIV-1 Vpu-mediated downregulation and consequently more effectively restricts HIV-1 release in the presence of Vpu than l-Tetherin. In contrast, the ebolavirus glycoprotein (GP) effectively counters both Tetherin isoforms. We also find that Tetherin is a significant activator of NF-κB. However, the Tetherin isoforms have dramatically different signaling potential, with NF-κB activation requiring a tyrosine motif found in the l-Tetherin isoform but absent in s-Tetherin. These observations provide insight into how the expression of alternatively translated Tetherin isoforms regulates multiple cellular functions and begins to shed more light on a poorly characterized signaling property.
Sequence analysis of the cytoplasmic tail of Tetherin revealed minimal amino acid conservation across species. Residues that were highly conserved included two cytoplasmic methionine residues and a dual tyrosine motif (Figure 1A), previously found to be involved in Tetherin trafficking between the plasma membrane and intracellular compartments [18], [19]. Methionine codons can function as translation start sites when they are flanked by a Kozak consensus sequence, which influences ribosomal binding at the site of translation initiation in eukaryotic cells. Analysis of the human tetherin mRNA sequence revealed that the Kozak sequences flanking the first AUG (M1) appears to be a potentially “leaky” signal due to a non-consensus pyrimidine at the −3 position (Figure 1B) [33]. The downstream methionine (M13) appears to be a stronger translation start site with compensatory cytosine residues at positions −1 and −2, which could enhance the strength in the absence of the purine in the −3 position [34]. Although neither of these methionine residues can be defined as strong Kozak consensus start sites, both sequences display conservation at the important +1 through +4 positions. Analysis of the mRNA sequences of Tetherin from other species revealed a similar potentially “leaky” Kozak signal for the upstream methionine (Figure S1), suggesting that production of multiple isoforms may be a conserved Tetherin characteristic.
To address whether alternatively initiated Tetherin isoforms could be produced, mutations that individually abrogated each of the AUG codons were generated in the wt cDNA (Figure 2A). N-linked glycosylation of Tetherin produces a heterogeneous population of proteins that is not well resolved by electrophoresis [[35] and Figure S2], complicating unambiguous identification of the observed Tetherin bands. Therefore, to better differentiate the potential isoforms that are predicted to differ by 12 residues (∼1.5 kDa), lysates from transiently expressing cells were treated with Peptide: N-Glycosidase F (PNGase) to remove N-linked glycosylation before analysis by Western blot. Analysis of deglycosylated lysates from HT1080 cells expressing the wt Tetherin cDNA reveals two major bands of between 15–20 kDa (Figs. S2 and 2B); the predicted sizes of proteins initiated at the M1 and M13 translation initiation sites is approximately 18 and 16.5 kDa respectively. Transient expression of the wt tetherin cDNA produces more of the larger form compared to the smaller species. Expression of the M13I mutant in HT1080 cells produces a single major product that aligns with the larger protein species seen in wt cDNA transfected cells (Figure 2B). The M1A mutant, that abolishes the upstream potential initiation site, generates a protein that migrates similarly to the smaller species produced by wt cDNA. Co-transfection of plasmids expressing the M1 and M13 mutants recapitulates the pattern seen with wt Tetherin (Figure 2B). Overall, these data demonstrate that transient expression of the wt cDNA produces two alternatively initiated species that we refer to as the l-Tetherin (long) and s-Tetherin (short) isoforms.
To address whether expression of the s-Tetherin isoform was due to translational leakage, an optimized translational initiation sequence was generated by converting the sequence at the −3 position of the upstream M1 ATG in the wt Tetherin cDNA from a pyrimidine (T) to a purine (A). Upon introduction into HT1080 cells of a vector with a strong Kozak sequence at the M1 AUG, expression of l-Tetherin was greatly enhanced while expression of s-Tetherin was almost undetectable (Figure 2B). This mutational analysis, coupled with the data presented above, demonstrates that Tetherin utilizes a non-canonical Kozak sequence to produce alternatively initiated proteins with different cytoplasmic tails.
To determine whether both isoforms of Tetherin are endogenously expressed, non-stimulated and interferon (IFN)-stimulated 293T, HT1080, HeLa, primary human CD4 T-cell lysates were analyzed as described above and compared to 293T cells transiently expressing the M13I and M1A mutants. As predicted from previous results, no Tetherin was seen upon analysis of deglycosylated proteins from unstimulated 293T and HT1080 cells. Interferon treatment of 293T and HT1080 cells induces Tetherin expression and upon PNGase treatment, collapses a diffuse banding pattern to two bands that migrate similarly to the two isoforms produced by transfection of l-Tetherin and s-Tetherin encoding plasmids (Figure 2C, 1st panel and Figure S3). This demonstrates that IFN stimulation produces both alternatively initiated forms of Tetherin. Analysis of lysates from HeLa cells, which constitutively express relatively high levels of Tetherin, revealed the presence of the doublet in the absence of IFN (Figure 2C, 2nd panel). Both bands appear to be enhanced after IFN stimulation (Figure S3). Similarly to HeLa cells, analysis of untreated primary CD4 T-cells revealed two Tetherin species that were enhanced upon IFN stimulation (Figure 2C, 3rd panel). In all cell types tested for endogenous expression of Tetherin, roughly equal expression of both isoforms was observed after 48 h of IFN stimulation. Expression of the M13I and M1A mutants in 293T cells produced the l- and s-isoforms seen above in HT1080 cells, however in this case smaller molecular mass species are seen (denoted by an asterisk on the 4th panel of Figure 2C and throughout the manuscript). These species are only observed in transiently expressing 293T, and not in transfected HT1080 or interferon induced cells. Though the derivation of these species is unclear, they may represent a precursor or proteolyzed product. Finally, analysis of cells transiently expressing rhesus or murine Tetherin cDNA or IFN treated murine J774 cells demonstrated a doublet indicative of production of two isoforms (Figure S4A and B) indicating that production of long and short isoforms of Tetherin is a highly conserved feature. In contrast, the rhesus cell line FRhK-4 appears to produce only a single Tetherin isoform (Figure S4B). Taken as a whole, the presence of these isoforms in numerous cell types, under various conditions and in multiple species is suggestive of important biological roles for the newly identified l-and s-Tetherin isoforms.
Tetherin exists as a disulfide-linked dimer that is believed to represent the functional form necessary for viral restriction. Structural studies on the ectodomain suggest that Tetherin dimers may also form higher order aggregates, which may be functionally important [36], [37]. To address whether the isoforms interact to form heterodimers, lysates from HT1080 cells expressing wt, l- or s-Tetherin were separated under non-reducing conditions and analyzed by Western blot. Expression of either the l- or s-Tetherin isoform produced single bands of a molecular mass indicative of homodimer formation (Figure S5A). In comparison, the wt cDNA produced a broader expression profile that overlapped with the l- and s-profiles suggesting both homo- and heterodimer formation (Figure S5A). To directly assess the interaction of the l- and s- isoforms, a C-terminally FLAG-tagged l-Tetherin and N-terminally AU1 tagged s-Tetherin were co-expressed and immunoprecipitated using antibodies to the epitope tags (Text S1). As anticipated for efficient heterodimer formation, immunoprecipitation with either tag effectively co-precipitated the other isoform (Figure S5B). As discussed below, the ability to form homo- and heterodimers may have significant effects upon the biological activities of Tetherin.
The cytoplasmic tail of Tetherin is dispensable for function as a viral restriction factor [17]. However, residues in the cytoplasmic tail have been shown to confer sensitivity to viral antagonists. For example, a serine-threonine-serine sequence which is unique to l-Tetherin has been found to be important for efficient Vpu antagonism [22]. Using a transient HIV-1 virus-like particle (VLP) budding assay, the l- and s- isoforms were analyzed for their ability to restrict viral release. Both isoforms were effective viral restriction factors and displayed strong inhibition of HIV-1 VLP release (Figure 3A, top panel). The antiviral activity of wt Tetherin, which expresses both Tetherin isoforms, is antagonized by co-expression of HIV-1 Vpu with VLP release correlating with the level of Vpu expression. In contrast, antiviral activity of the l-Tetherin isoform is exquisitely sensitive to Vpu with significant rescue of viral budding seen at the lowest levels of Vpu expression analyzed (Figure 3A, top panel). Additionally, a decrease in l-Tetherin protein levels was readily observed at higher levels of Vpu expression (Figure 3A, third panel). Conversely, the s- isoform appears to be resistant to antagonism by Vpu, with very little VLP budding rescue observed at the highest levels of Vpu tested (Figure 3A, top panel). Moreover, turnover of the s- isoform also appears to be resistant to Vpu as there was no obvious decrease in s-Tetherin expression even at the highest levels of Vpu (Figure 3A, third panel).
To determine whether the differential sensitivity to viral antagonists was a common property of the Tetherin isoforms, sensitivity to a different viral factor, the ebolavirus glycoprotein, GP, was analyzed. GP is hypothesized to rescue virus budding in Tetherin-expressing cells through a non-cytoplasmic tail-associated mechanism [38] and might therefore be predicted to interact similarly with both Tetherin isoforms. In contrast to the results with Vpu, within the range of GP levels tested, GP-mediated rescue of budding was similar for wt and the two Tetherin isoforms. (Figure 3B, top panel). Additionally, total cellular Tetherin protein levels appeared unchanged for wt Tetherin and each individual isoform even at the highest levels of ebolavirus GP co-expression (Figure 3B, third panel).
To determine whether the differential particle release by the isoforms in response to Vpu was due to altered Tetherin cell surface levels, flow cytometric analysis of live cells stained for Tetherin was performed. In Vpu expressing cells, surface wt Tetherin expression followed the same trend as the total cell Tetherin protein levels (compare Figure 3A third panel and Figure 4). Surface expression of l-Tetherin was downregulated by Vpu to a greater extent than the s-Tetherin isoform, while wt Tetherin appeared to have an intermediate level of downregulation (Figure 4). Analysis of lysates from the cells prepared in parallel with those used for flow cytometry confirmed the enhanced degradation of l-Tetherin compared to s-Tetherin observed above in both the reduced monomers and stable dimers. Additionally, preferential degradation of the l-isoform was seen in lysates of cells expressing wt Tetherin and high levels of Vpu (Figure 4; arrow). In comparison, there is no difference in the effect of ebolavirus GP on surface or total levels for the l- and s- isoforms though a decrease in surface staining was observed across all Tetherin constructs tested. The minimal 2-fold decrease in the mean fluorescence intensity of Tetherin surface staining observed in ebolavirus GP expressing cells is consistent with steric shielding of epitopes previously documented for this glycoprotein [39], [40] and appears similar to the slight decrease noted recently by Lopez and colleagues [38]. Overall, these observations suggest that the Tetherin isoforms can have different sensitivities to viral antagonists, which may have important consequences during natural infection.
To further explore the basis for the relative Vpu resistance of s-Tetherin compared to l-Tetherin, a series of mutations were engineered into the 12 residue region unique to l-Tetherin (Figure 5A). Two motifs within this region have been proposed to play a role in Vpu mediated antagonism; a dual tyrosine motif (amino acids 6 and 8, Figure 5A) involved in endocytic cycling of Tetherin [41] and a stretch of serine and threonine residues (amino acids 3 to 5, Figure 5A) that can be ubiquitinated [22], [42]. Analysis of HIV-1 VLP release from 293T cells transiently expressing a tyrosine mutant AxA produced in l-tetherin expressing cDNA shows decreased release in response to Vpu as compared to parental l-Tetherin (lane l-AxA, Figure 5B). However, the observed decreased Vpu antagonism of AxA is not as pronounced as that seen with s-Tetherin. Altering the serine/threonine residues also impairs Vpu mediated particle release (lane l-STS, top panel Figure 5B). However, as with the tyrosine mutations the effect seen is intermediate between l- and s-Tetherin. A mutant that combines the tyrosine and serine/threonine changes appears to recapitulate the high level of resistance to Vpu antagonism observed with s-Tetherin (lane l-SY, top panel Figure 5B). Thus it appears that both these regions are involved in Vpu mediated antagonism of Tetherin function.
Although the tyrosine and serine/threonine mutants appear to similarly affect response to Vpu antagonism and particle release, the STS mutant Tetherin protein demonstrates resistance to Vpu-mediated degradation comparable to that of s-Tetherin (third panel, Figure 5). In contrast, the AxA mutant protein levels decrease upon Vpu expression comparably to parental l-Tetherin (third panel, Figure 5). The combined tyrosine and serine/threonine mutant (third panel, lane l-SY) is Vpu resistant like s-Tetherin and the STS mutant. Overall, these findings support an important role for the STS region in Vpu mediated degradation of Tetherin.
Tetherin has been primarily characterized as a cell intrinsic viral restriction factor. However, Tetherin has been proposed to have additional activities including a role in induction of the proinflammatory response regulator NF-κB [32]. To confirm that Tetherin was able to activate NF-κB, a luciferase reporter assay was employed in 293T cells transiently expressing Tetherin. Activation of NF-κB by Tetherin displays a bell shaped curve with decreased stimulation at high levels of Tetherin expression (Figure S6), suggesting an optimal range of expression for signaling. Expression of wt Tetherin led to NF-κB activation approximately 20–30 fold over control cells transfected with the expression vector alone (Figure 6A). Moreover, Tetherin activation of NF-κB is specific since there was no effect upon AP1-regulated signaling (Figure 6B). TRAF6, a ubiquitin ligase involved in regulating various signal transduction pathways including NF-κB, upregulated both NF-κB and AP1. Supporting the specificity of the observed effect for NF-κB, activation by Tetherin was effectively blocked in a dose dependent manner by expression of a dominant-negative form of the NF-κB activating kinase IKKβ (DN-IKKβ, Figure 6C).
We next assessed whether the Tetherin isoforms had differential abilities to activate NF-κB. Unlike wt Tetherin, the truncated s-Tetherin isoform displayed no NF-κB activation (Figure 6A). Conversely, compared to wild type, the l-Tetherin isoform was, on average, a more potent activator and increased NF-κB activity approximately 40-fold over basal levels. As was seen for wt Tetherin, NF-κB activation by l-Tetherin was sensitive to inhibition by DN-IKKβ (Figure 6C). These findings demonstrate that sequences within the 12 amino acid region absent in s-Tetherin are required for the signaling events downstream of Tetherin that activate NF-κB.
As described above, Tetherin can assemble into hetero and homodimers of the l- and s- isoforms. Given the observations that s-Tetherin does not activate NF-κB, l-Tetherin is more potent than wt, and wt Tetherin has an intermediate signaling phenotype to that of l- and s-isoforms, we investigated whether the short isoform could modulate signaling activity of the longer species. To address this question, varying ratios of the two isoforms were assessed for their ability to activate NF-κB. As seen in Figure 7, l-Tetherin mediated NF-κB activation is strongly diminished by the presence of s-Tetherin. The ability of s-Tetherin to sharply reduce signaling at a 1∶1 and 1∶3 ratio of l- to s-Tetherin expressing plasmids is supportive of a model in which homo-oligomers of the longer species are needed to signal. Overall, this observation suggests an inhibitory role for s-Tetherin in regulating NF-κB signaling by l-Tetherin.
Within the 12 amino acids unique to l-Tetherin are two highly conserved tyrosine residues at positions 6 and 8 (Figure 1). To address whether these residues contributed the ability of the longer isoform to activate NF-κB, alanine substitution mutations were generated in the context of l-Tetherin and tested by transient expression in 293T cells. Compared to the parent l-Tetherin, the ability of the tyrosine mutant (AxA) to activate NF-κB was significantly reduced (Figure 6A). Although impaired, a low but reproducible level of activity was observed for the dual tyrosine mutant. These data suggest that the mutated residues form part of an important signaling motif within the cytoplasmic tail region needed to activate NF-κB that is present only in the longer isoform of Tetherin.
Here we identify a previously undescribed isoform of human Tetherin generated by translation at an in-frame initiation codon 33 nucleotides downstream of the canonical translation start site, which results in the loss of a 12-amino-acid N-terminal sequence. Alternative translation initiation in eukaryotes can occur by 3 different mechanisms: internal ribosome entry, reinitiation, or leaky ribosome scanning. Our analysis indicates that the tetherin mRNA contains a “leaky” Kozak sequence around the first AUG codon (M1). Based on the scanning model of translation, this would allow the ribosome to skip over the first start codon a fraction of the time, leading to initiation at the next in frame AUG (M13). Selective disruption of each AUG and introduction of a strong Kozak at the upstream AUG confirmed that alternate sites of translation initiation accounted for each of the observed products. Conservation of an upstream leaky Kozak and a second cytoplasmic methionine in most mammalian tetherin sequences, along with our analysis of expression from rhesus and murine cDNAs, suggests that multiple isoforms exist in other species and indicate an important biologic role(s) for the two isoforms.
For many messages with alternatively translated isoforms the resultant proteins display distinct functionalities [43]–[46]. This is clearly the case for Tetherin where we identified unique biologic properties for the two isoforms. Although both isoforms exhibit antiviral activity, the shorter Tetherin isoform, derived by initiation at M13, is highly resistant to the antagonistic effects of HIV-1 Vpu and therefore may play a more important role in restricting virion budding in HIV-1 infected cells. Analysis of mutants in the 12 residues unique to l-Tetherin suggests that the combined loss of two tyrosine and three serine/threonine residues accounts for the Vpu resistance of s-Tetherin. By contrast, l-Tetherin seems to be hypersensitive to Vpu antagonism. However, this longer isoform possesses the ability to induce the immune response regulatory transcription factor NF-κB while s-Tetherin does not have this capability. Expression of the wt Tetherin cDNA, which produces both protein species yielded phenotypes for Vpu sensitivity and NF-κB signaling that were intermediate between those seen upon expression of the individual isoforms. This likely reflects that fact that when the isoforms are co-expressed both hetero and homodimers form, and that the unique properties observed (e.g. Vpu resistance and NF-κB signaling) require short-short or long-long homodimers respectively.
In several other systems where alternative translation initiation produces multiple isoforms there are observed differences in the ratios of the isoforms in tissues [44], [45] or under various physiologic conditions [47] suggesting regulation of alternative initiation. For example, the levels of glucocorticoid receptor (GR) isoforms vary significantly in different tissues, indicating that cells can regulate this ratio with important consequences for GR gene regulation [44]. Similarly, LPS regulates differential expression of alternatively translated forms of the CCAAT/enhancer-binding trans-activator proteins C/EBPα and C/EBPβ, producing isoforms with unique properties [47]. In our experiments, the ratios of l- and s-Tetherin differed in transient expression versus endogenous expression. Lysates harvested 48 h post-transient transfection appeared to produce more of the longer species. In contrast, 48 h stimulation with IFN produced roughly equal amounts of both isoforms. Interestingly, tissue specific expression of ovine Tetherin isoforms has been reported [6]; however, the mechanism for generation of these isoforms was not investigated. Our analysis of human Tetherin did not query a large number of cell types or conditions, therefore it will be interesting to more fully address whether there are instances in which the ratio of l- and s-Tetherin is regulated. Conversely, the increased sensitivity of l-Tetherin to Vpu compared to s-Tetherin suggests that viral antagonists might alter the isoform ratio. Because both NF-κB signaling and Vpu sensitivity appear to be affected by the formation of homodimers, this suggests a model whereby subtle changes in isoform stoichiometry, either by differential expression or susceptibility to viral antagonists of one isoform, could rapidly and dramatically affect Tetherin function.
The importance of the shorter Tetherin isoform with increased antiviral activity or resistance to viral antagonists is strongly supported by recent evidence. While this manuscript was in preparation, a polymorphic allele that abrogated expression from the first methionine residue, and thus produced only short isoforms, was described in NZW mice. In these mice, expression of shorter Tetherin isoforms strongly correlated with decreased Friend retrovirus replication and pathogenesis [48]. Tetherin sequence variants have also been identified in rhesus macaques, African green monkeys and mice [49], [50]. Interestingly, none of the identified macaque polymorphisms account for the single species we observed by Western blot in the macaque FRhK-4 cell line. Further analysis of genomic sequence of tetherin from this rhesus cell line is required to determine the form being expressed.
We find that the ability of s-Tetherin to retain budding HIV-1 virions and to resist Vpu more effectively than l-Tetherin appears to be due to loss of tyrosine and serine threonine motifs found in the cytoplasmic tail of the long isoform. Mutation in either motif rendered Tetherin partially resistant to Vpu. Combining these two sets of mutations recapitulated loss of Vpu sensitivity seen for s-Tetherin suggesting these represent important differences between l- and s-Tetherin. The tyrosine motif in l-Tetherin is a non-canonical trafficking signal (YXYXXV) that engages the AP-1 and AP-2 clathrin adaptors [18], [19]; mutations within this sequence alter cellular trafficking and distribution [51], [52]. It has recently been demonstrated that Vpu impairs recycling of Tetherin to the cell surface and that this is an key mechanism for antagonizing Tetherin function [41]. It will be interesting to determine if s-Tetherin effectively recycles in the presence of Vpu or if endocytosis from the cell surface is affected. Also, within the region unique to l-Tetherin is a serine-threonine-serine stretch that has been shown to be important for Vpu mediated Tetherin degradation [22] although more recent studies have questioned this conclusion [22], [42]. Our finding that mutations in this motif confer partial resistance to Vpu antagonism and also affect Vpu induced degradation support an import role for this region in interactions with Vpu. Interestingly, particle retention and Tetherin degradation appear to be separable features because the STS mutant appeared to be resistant to Vpu induced degradation while the YxY mutant remained sensitive to Vpu degradation, yet both mutations partially impair Vpu induced virion release. Our data is consistent with a model in which both decreased protein turnover and altered trafficking account for the ability of s-Tetherin to effectively retain virions in Vpu-expressing cells.
The observation that the long isoform activates NF-κB while the short isoform shows no activity suggests that sequences within the 12 residues unique to l-Tetherin mediate signaling. Moreover, data demonstrating higher NF-κB induction by expression of the l-Tetherin isoform compared to the wt cDNA, coupled with an inhibitory affect of s-Tetherin on NF-κB activation, support a model in which homodimers of the longer isoform are responsible for signaling. Though Tetherin does not possess canonical tyrosine-based motifs involved in activating signal transduction pathways, there is evidence that dimerization of CLEC-2 allows approximation of a non-canonical tyrosine-based motif and permits signaling via Syk kinase [53]. Furthermore, phosphoproteomic analysis has shown that Tetherin tyrosine residues at positions 6 and 8 can be phosphorylated [54]. Moreover, mutation of both these tyrosine residues greatly diminished NF-κB induction, demonstrating that they are critical for this activity. Because the dual tyrosine residues have been shown to be important in trafficking it will be crucial to assess whether differential localization may also play a role in activating signal transduction. The ability of a dominant negative form of the kinase IKKβ to inhibit Tetherin-induced NF-κB activity indicates that activation likely involves the canonical pathway, however the mechanism by which Tetherin couples to this pathway remains to be elucidated. Similarly to our findings with Tetherin, it has recently been found that the cytoplasmic tail domain of HIV-1 envelope activates NF-κB, but not AP1 [55]. In this case activation occurs via the canonical NF-κB pathway, with HIV-1 envelope directly engaging TGF-β-activated kinase 1 (TAK1). Whether Tetherin utilizes a similar mechanism to specifically activate NF-κB remains to be explored.
It has been demonstrated that Tetherin is a ligand for ILT7 on human dendritic cells and that binding initiates signaling via the ILT7–FcεRIγ complex, inhibiting production of interferon and proinflammatory cytokines by dendritic cells [31]. It will be interesting to investigate if ILT7, or an unknown ligand can modulate NF-κB activation by Tetherin. The viral restriction factor TRIM5α has recently been shown to function as a unique pattern recognition receptor recognizing the hexagonal lattice pattern of entering retroviral capsids to activate NF-κB and MAP kinase pathways [56]. In a similar vein, we speculate that Tetherin could recognize the “pattern” caused by budding virions to activate NF-κB.
293T, HT1080 and HeLa cells were maintained in high glucose DMEM+10% Cosmic Calf Serum (Hyclone). Primary CD4 T cells (obtained from the University of Pennsylvania Center for AIDS Research Immunology Core, ND307) were maintained in RPMI+10% FBS (Invitrogen). All cells were maintained at 37°C with 5% CO2.
Human tetherin cDNA in pCMV-SPORT6 was acquired from Open Biosystems. Tetherin mutants were generated using site-directed mutagenesis by PCR. Primers for site-directed mutagenesis were designed using QuikChange Primer Design Program (Agilent Technologies). The HIV-1 gag-pol encoding construct, psPAX2 was obtained from Addgene (plasmid 12260). HIV-1 vpu and ebolavirus GP cloned into pCAGGS were previously described [3], [57]. The NF-κB reporter plasmid, pBIIX-Luciferase was provided by Dr. Michael May (University of Pennsylvania). The TRAF6 coding region was cloned from a cDNA (Open Biosystems) into pKMyc (Addgene plasmid 19400) to generate an N-terminally FLAG-tagged version using an XbaI restriction site adjacent to the FLAG-tag. The dominant negative FLAG epitope tagged IKKβ (K44M) plasmid was obtained from Addgene (plasmid 11104). The AP1 reporter (Biomyx) and MEKK1-delta plasmids were obtained from Dr. Sunny Shin (University of Pennsylvania).
HT1080 cells in a 6 well were transiently transfected with Tetherin expression constructs using Lipofectamine2000 (Invitrogen) according to manufacturers instructions. For IFN induction experiments, HT1080, 293T, HeLa and primary CD4 T cells were incubated +/− 1000 U of recombinant type I IFN (PBL Interferon Source) for 48 h. Two days post-transfection/IFN treatment, cells were washed with PBS and lysed using RIPA buffer (10 mM Tris-HCl pH 8.0, 5 mM EDTA, 140 mM NaCl, 1% sodium deoxycholate, 0.1% SDS, 1% NP40). Lysates were passaged through a 25G needle before being cleared by centrifugation at 17,900× g for 15 min at 4°C. Cleared lysates were treated with either treated or not with PNGase (New England BioLabs) for 2 h, then reduced using DTT containing loading buffer. Samples were separated on a 15% Criterion gel (BioRad) and transferred to PVDF membrane. After blocking in 5% milk TBST (Tris-buffered saline+0.1% Tween), Western blots were analyzed for Tetherin expression using rabbit anti-BST2 sera (Dr. Klaus Strebel, #11721 National Institutes of Health AIDS Research and Reference Reagent Program).
2.5×105 293T cells seeded on a 24 well plate were co-transfected with the indicated tetherin construct (12.5 ng), an HIV-1 Gag-Pol expression vector pSPAX (50 ng) and increasing amounts of (25, 50 or 100 ng) of pCAGSS-vpu or pCAGGS ebolavirus GP using Lipofectamine2000. Cell lysates and supernatants were harvested 48 h post-transfection. Cell lysates were harvested in Triton X-100 buffer (50 mM Trs-HCl pH 8.0, 5 mM EDTA, 150 mM NaCl, 1% Triton X-100) with Complete (Roche) protease inhibitor cocktail. Cell lysates were cleared by centrifugation at 17,900× g for 3 min at 4°C. Supernatants containing VLPs were cleared are 1700× g for 2 min at 4°C. VLPs were subsequently purified by pelleting through a 20% sucrose cushion at 40,000 in a TLA120.1 rotor (Beckman) for 30 min. VLPs were resuspended in PBS on ice for 3 h. Cleared cell lysates and resuspended VLPs were separated on 15% or 4–15% Criterion gels (BioRad) respectively before being transferred to PVDF membrane. Membranes were blocked in 5% milk in TBST for 40 min prior to incubation with primary antibody. HIV-1 Gag p24 was detected in Western blots for both VLPs and cell lysates using a monoclonal anti-p24 antibody (24-3, National Institutes of Health AIDS Research and Reference Reagent Program). The GAPDH loading control was analyzed using GAPDH antibody (Calbiochem, CB1001). HIV-1 Vpu was detected using HIV-1 pNL4-3 Vpu antiserum (969, National Institutes of Health AIDS Research and Reference Reagent Program). Ebolavirus GP was detected using rabbit antiserum against the GP1 portion of the protein [58].
293T cells (2.5×105) seeded on a 24 well plate were transiently co-transfected with 25 ng Tetherin constructs (wt, l-Tetherin, s-Tetherin or l-AxA) and 100 ng HIV-1 vpu or ebolavirus GP. Two days post transfection, cells were washed once with 1×PBS while on the plate. Whole cells were removed using cold 1×PBS and the cells were split into equal aliquots used for flow cytometry and western blot analysis. Cells used for Western blot analysis were pelleted by centrifugation then lysed in Triton X-100 buffer. Lysates were cleared and PNGase treated prior to SDS/PAGE Western blot analysis as described above. Cells to be used for flow cytometry were resuspended in cold FACS buffer (1×PBS, 1%BSA+0.05% sodium azide) and pelleted at 2150× g. Cells were resuspended in FACS buffer+PE conjugated anti-BST2 antibody (Biolegend). After a 1 h incubation on ice, cells were washed three times with cold FACS buffer and analyzed on a FACS Calibur (BD Biosciences Immunocytometry Systems). Data analysis performed using FlowJo 9.3.1 (Tree Star).
293T cells (2.5×105) were seeded on a 24 well plate and the following day were transfected with the indicated tetherin constructs (50 ng) and pBIIX-Luciferase reporter (250 ng) using Lipofectamine2000. In experiments where the long and short isoforms were co-expressed the total amount of tetherin expression plasmid was kept constant at 50 ng. At 30 h post-transfection, cell lysates were harvested in Triton X-100 lysis buffer. Lysates were transferred to a black flat bottom 96-well plate. Luciferase Assay System substrate (Promega E1501) was added to the lysates according to manufacturers directions. Samples were analyzed in a Luminoskan Ascent microplate luminometer (Thermo Scientific). Statistical analysis was performed using PRISM (GraphPad).
|
10.1371/journal.pgen.1003341 | Both the Caspase CSP-1 and a Caspase-Independent Pathway Promote Programmed Cell Death in Parallel to the Canonical Pathway for Apoptosis in Caenorhabditis elegans | Caspases are cysteine proteases that can drive apoptosis in metazoans and have critical functions in the elimination of cells during development, the maintenance of tissue homeostasis, and responses to cellular damage. Although a growing body of research suggests that programmed cell death can occur in the absence of caspases, mammalian studies of caspase-independent apoptosis are confounded by the existence of at least seven caspase homologs that can function redundantly to promote cell death. Caspase-independent programmed cell death is also thought to occur in the invertebrate nematode Caenorhabditis elegans. The C. elegans genome contains four caspase genes (ced-3, csp-1, csp-2, and csp-3), of which only ced-3 has been demonstrated to promote apoptosis. Here, we show that CSP-1 is a pro-apoptotic caspase that promotes programmed cell death in a subset of cells fated to die during C. elegans embryogenesis. csp-1 is expressed robustly in late pachytene nuclei of the germline and is required maternally for its role in embryonic programmed cell deaths. Unlike CED-3, CSP-1 is not regulated by the APAF-1 homolog CED-4 or the BCL-2 homolog CED-9, revealing that csp-1 functions independently of the canonical genetic pathway for apoptosis. Previously we demonstrated that embryos lacking all four caspases can eliminate cells through an extrusion mechanism and that these cells are apoptotic. Extruded cells differ from cells that normally undergo programmed cell death not only by being extruded but also by not being engulfed by neighboring cells. In this study, we identify in csp-3; csp-1; csp-2 ced-3 quadruple mutants apoptotic cell corpses that fully resemble wild-type cell corpses: these caspase-deficient cell corpses are morphologically apoptotic, are not extruded, and are internalized by engulfing cells. We conclude that both caspase-dependent and caspase-independent pathways promote apoptotic programmed cell death and the phagocytosis of cell corpses in parallel to the canonical apoptosis pathway involving CED-3 activation.
| Caspases are cysteine proteases that in many cases drive apoptosis, an evolutionarily conserved and highly stereotyped form of cellular suicide with functions in animal development and tissue maintenance. The dysregulation of apoptosis can contribute to diseases as diverse as cancer, autoimmunity, and neurodegeneration. Caspases are often thought to be required for, or even to define, apoptosis. Although there is evidence that apoptosis can occur in the absence of caspase activity, caspase-independence can be difficult to prove, as most animals have multiple caspases. The nematode Caenorhabditis elegans has four caspases, CED-3, CSP-1, CSP-2, and CSP-3. CED-3 has a well-established role in apoptosis, but less is known about the functions of the CSP caspases. In this study, we show that CSP-1 promotes apoptosis in the developing C. elegans embryo and that CSP-1 is regulated differently than its homolog CED-3. Furthermore, we show that apoptosis and the engulfment of dying cells can occur in mutants lacking all four caspases, proving that neither apoptosis nor cell-corpse engulfment require caspase function and that caspase-independent activities can contribute to apoptosis of some cells during animal development.
| The elimination of unnecessary or dangerous cells is fundamental to development, tissue homeostasis and disease mitigation in multicellular organisms. The primary mechanism of cell elimination is apoptosis, a form of cell suicide that was originally defined by evolutionarily conserved morphological characteristics that include chromatin condensation, shrinkage of the cytoplasmic volume and membrane blebbing [1] and by biochemical features like phosphatidylserine exposure and DNA fragmentation [2], [3]. Apoptosis serves as a highly controlled mechanism for the removal and degradation of damaged or unnecessary cells, and blocking apoptosis can lead to catastrophic forms of cell death, such as necrosis, which can cause dangerous inflammatory responses [4]. The discovery of the CED-3 caspase as a cell-autonomous executioner of programmed cell death in the nematode Caenorhabditis elegans led to the paradigm that the caspase family of cysteine proteases drives apoptosis through the cleavage of substrate proteins at specific peptide sequences [5], [6]. Indeed, caspases have evolutionarily conserved roles in apoptosis throughout metazoa [7].
Despite the compelling causal link between caspases and apoptosis, a growing body of evidence indicates that apoptosis can occur in the absence of caspases [4]. For example, mouse cells lacking Apaf-1, an activator of the apical caspase Caspase-9, which in turn activates effector caspases, can undergo apoptosis in response to pro-apoptotic stimuli [8]. In the presence of caspase inhibitors, TNF (tumor necrosis factor) can induce a form of cell death termed necroptosis, which exhibits characteristics of both necrosis and apoptosis [4], [9]. The mitochondrial flavoprotein AIF (apoptosis-inducing factor) is thought to promote apoptotic cell death in mammals even in the presence of caspase inhibitors [10]. Furthermore, cell death with aspects of apoptotic morphology occurs in non-metazoans, including unicellular eukaryotes and prokaryotes, that lack clear caspase homologs [11], [12]. Thus, it is possible that apoptosis, as defined morphologically and biochemically, can occur in the absence of caspases.
A standard approach to assaying the caspase-dependence of apoptotic stimuli in tissue cell culture is through the pharmacological inhibition of caspases. However, it is difficult to prove that caspase activity is completely blocked in such experiments, and it is possible for caspase inhibitors to trigger non-apoptotic forms of cell death [13]. Studies of caspase-independent apoptosis in metazoans are also complicated by the existence of multiple caspases with potentially redundant functions in promoting cell death. The human genome, for example, encodes at least 10 caspase homologs, seven of which (caspases-2, -3, -6, -7, -8, -9 and -10) have demonstrated roles in apoptosis [14]. The genome of Drosophila melanogaster encodes seven caspase homologs (dcp-1, dronc, drice, dredd, decay, damm and strica) [7], several of which are essential for organismal viability. The C. elegans genome encodes three caspase homologs (csp-1, csp-2 and csp-3) in addition to ced-3 [15]. Therefore, the use of mutant animals or cell lines deleted for one or two caspases might not eliminate all caspases expressed within a specific cell. Furthermore, since caspases have different substrate specificities [16], the use of a chemical substrate-competitive caspase inhibitor might not completely block all caspase activity. Ideally, experiments that test whether apoptosis can occur in the absence of caspases should be performed using mutant animals or cells that are genetically deleted of all caspase homologs. In this regard, C. elegans is an excellent animal for studies of caspase-independent programmed cell death, because: (1) there are several examples of ced-3-independent programmed cell death in C. elegans [17]–[19]; (2) mutants of ced-3, csp-1, csp-2 and csp-3 are viable [18]–[23]; and, (3) it is relatively easy to generate multiply mutant C. elegans strains.
The ced-3 caspase gene is required for most programmed cell deaths that occur during C. elegans development [5], [20]. However, a small number of cells die in animals carrying null mutations of ced-3. The male-specific linker cell, which facilitates the connection of the vas deferens to the cloaca and then dies, undergoes a non-apoptotic cell death that bears morphological features (e.g., nuclear membrane crenellation) not seen with other C. elegans programmed cell deaths and that occurs in ced-3 mutants as well as in animals doubly mutant for ced-3 and csp-1, csp-2 or csp-3 [18], [20], [24]. We recently showed that a subset of cells fated to die in the C. elegans embryo are eliminated from ced-3 mutants via a caspase-independent shedding mechanism [19]. Interestingly, the shed cells appear apoptotic, exhibiting chromatin condensation, TUNEL-reactive DNA degradation and phosphatidylserine exposure despite the absence of all four caspases. Unlike other apoptotic programmed cell deaths of C. elegans, the shed cells do not undergo phagocytosis by engulfing cells; instead, they are extruded from the developing embryo. By contrast, a small number of apoptotic cell corpses are visible in the heads of ced-3 larvae [17]. Like other programmed cell deaths of C. elegans, these ced-3-independent cell corpses have a refractile appearance when viewed with Nomarski optics and are not extruded from the animal, suggesting that a ced-3-independent cell-killing activity contributes to these typical programmed cell deaths. The other caspase homologs, csp-1, csp-2 and csp-3, are obvious candidates for driving this ced-3-independent cell-killing activity. However, it has recently been reported that csp-2 and csp-3 inhibit apoptosis in the germline and soma, respectively [22], [23].
To date, the C. elegans caspase homolog csp-1 has no known function in vivo, including in apoptosis. An isoform of CSP-1 can cleave and possibly activate the CED-3 pro-protein in vitro [15]. We tested whether csp-1 can promote or inhibit programmed cell death and whether it is regulated by the canonical programmed cell death pathway that activates ced-3. We found that csp-1 encodes a pro-apoptotic caspase activity that promotes programmed cell death independently of the CED-3 caspase, CED-4 (the Apaf-1 homolog that activates CED-3), and CED-9 (a Bcl-2 family protein that negatively regulates CED-3 activation via inhibition of CED-4). Furthermore, we tested whether csp-1, csp-2 and csp-3 contribute to the ced-3-independent cell-killing activity that generates cell corpses in the heads of ced-3 mutant larvae and found that these apoptotic cell deaths can occur in the complete absence of caspases. Thus, during C. elegans development programmed cell death followed by cell-corpse engulfment is achieved through three redundant pathways: (1) a ced-3-dependent pathway; (2) a csp-1-dependent pathway, which is not regulated by the canonical apoptosis pathway that controls ced-3; and, (3) a caspase-independent pathway.
The C. elegans genes csp-1, csp-2 and csp-3 are paralogs of the pro-apoptotic ced-3 caspase gene [15], which is required for most programmed cell deaths in the worm [5], [20]. Given the conserved role of caspases in apoptosis, we tested csp-1, csp-2 and csp-3 for roles – both pro- and anti-apoptotic – in programmed cell death. We used mutations of csp-1 (n4967 and n5133) and csp-2 (n4871) that completely remove the genomic sequences encoding their respective predicted caspase active sites (SACRG in the CSP-1 protein, and VCCRG in the CSP-2 protein) and therefore eliminate any potential caspase activity encoded by these genes (ref. [19]; Figure 1A). csp-3 lacks a caspase active site (ref. [15], [22]; Figure 1A); we used the csp-3 deletion mutation n4872, which is likely a null allele [19].
Recently, it was reported that mutations in csp-2 and csp-3 cause ectopic cell deaths in the germline and soma, respectively, and hence that csp-2 and csp-3 inhibit apoptosis [22], [23]. We therefore tested whether csp-1 mutants have ectopic cell deaths indicative of a loss of anti-apoptotic function. Using Nomarski optics and a Pmec-3::gfp transgene that expresses GFP in the six touch neurons (AVM, two ALM, PVM and two PLM neurons) in addition to the FLP and PVD neurons, we examined csp-1 mutants for missing cells that normally survive. We observed that csp-1(n4967) mutants contained a full complement of touch neurons and pharyngeal cells (Table S1). We also noted that csp-1(n4967) failed to cause ectopic cell deaths in sensitized animals carrying the loss-of-function mutation n2812 in the anti-apoptotic gene ced-9, a homolog of human BCL2 (Table S1; data not shown). These results indicate that csp-1 does not have an obvious anti-apoptotic function in the soma. Consistent with a previous report that csp-2 does not affect somatic cells [23], a mutation in csp-2 did not cause ectopic cell deaths in the somatic cells we examined (Table S1). However, we failed to observe the ectopic cell deaths in csp-3 mutants previously reported [22]. Ectopic somatic cell deaths have also been noted in animals with loss-of-function mutations in ced-9 [25] or tat-1 [26], [27], which encodes an aminophospholipid translocase required for the asymmetric distribution of phosphatidylserine on the inner leaflet of the plasma membrane. As expected, we found that ced-9 mutant larvae were missing pharyngeal cells and many touch neurons: more than 80% of PLM neurons were not present in ced-9(n2812) larvae (Table S1). However, we failed to detect the previously reported ectopic cell-death defect of tat-1 mutants (ref. [26], [27]; Table S1); we used the same deletion alleles for csp-3 and tat-1 and assayed the same cells that had been characterized in the previous studies.
To determine whether the C. elegans caspase homologs csp-1, csp-2 or csp-3 promote programmed cell death in the soma, we examined animals carrying csp deletion mutations for extra cells that failed to undergo programmed cell death in the anterior pharynx. As many as 16 extra cells can be counted in the anterior pharynges of mutants with strong defects in programmed cell death, e.g., ced-3(n3692) (ref. [28]; Table 1). Single mutations in csp-1, csp-2 or csp-3 failed to cause detectable defects in programmed cell death (Table 1; data not shown). However, we observed that mutations in csp-1 (but not csp-2 or csp-3) caused the survival of pharyngeal cells in sensitized strains carrying a weak mutation in the caspase gene ced-3 (Table 1). The partial loss-of-function ced-3 mutations n2427 and n2436 cause slight and intermediate defects in apoptosis, respectively (ref [17]; Table 1; data not shown). The n4967 and n5133 mutations, both of which delete the putative active site of CSP-1 (Figure 1A), enhanced the cell-death defects of ced-3(n2427) and ced-3(n2436) mutants, increasing the number of extra cells in their anterior pharynges on average by 1.4 and 2.4 cells, respectively (Table 1). These results are consistent with our RNAi experiments in which csp-1B dsRNA (which likely inactivated all csp-1 transcripts) was injected into the gonads of rrf-3(pk1426); ced-3(n2436) animals and caused an enhanced cell-death defect in their progeny (Figure 1C); we used the rrf-3 mutation to increase sensitivity to RNAi [29]. The cell-death defect conferred by the csp-1(n4967) mutation was rescued by extrachromosomal arrays carrying a 9 kb genomic csp-1 fragment that included the entire csp-1 coding region, 1.5 kb of genomic sequence 5′ of the csp-1A translational start codon and 3.5 kb of genomic sequence 3′ of the csp-1A/B translational stop codon (Figure 1B; Table S2). These results demonstrate that csp-1 encodes a detectable cell-killing activity that contributes to programmed cell death in C. elegans. Mutation of csp-2 and/or csp-3 neither enhanced nor suppressed the cell-death defects of strains mutant for csp-1 and/or ced-3 (Table 1; Table S3), suggesting that csp-1 and ced-3 are the only C. elegans caspase genes with functions in somatic programmed cell deaths.
The development of the anterior part of the C. elegans pharynx involves 16 programmed cell deaths, all of which appear to be sensitive to ced-3 [17], [28], [30]. To test whether specific pharyngeal programmed cell deaths required csp-1, we used GFP reporters to visualize the survival of cells fated to die, specifically the sister cells of the M4 and NSM neurons. csp-1 was partially required in ced-3(n2427) or ced-3(n2436) sensitized strains for the death of the M4 sister cell (Table S4); by contrast, mutation of csp-1 did not affect the cell deaths of the sister cells of the NSM neurons (data not shown). Likewise, csp-1 did not appear to function in the postembryonic programmed cell deaths of the ventral cord or postdeirid sensilla (Table S4). We conclude that csp-1 promotes cell death in a subset of cells fated to die during C. elegans development.
The csp-1 locus produces three known mRNA isoforms [15], all of which include the sequence that encodes the presumptive caspase active site (Figure 1A). The csp-1A transcript contains a long prodomain not present in the other transcripts, and it uses an alternative start site that is 3 kb 5′ to the start site of the csp-1B and csp-1C isoforms. To determine which isoforms are required for the cell-killing activity of csp-1, we peformed experiments in which the csp-1 rescuing transgene was mutated to express: (1) the A isoform only, (2) the B and C isoforms only, or (3) a truncated version of csp-1A including only the prodomain (PD). Extrachromosomal arrays engineered to express only csp-1-PD or the csp-1A isoform failed to rescue the cell-death defect of csp-1(n4967) mutants (Figure 1B; Table S2). By contrast, a csp-1 transgene lacking the csp-1A translation start codon and predicted to express only the csp-1B and csp-1C transcripts rescued the csp-1(n4967) defect in programmed cell death (Figure 1B; Table S2). Consistent with these results, transgenes expressing the csp-1B cDNA, but not the csp-1A cDNA, under the control of the mec-7 promoter efficiently killed touch neurons (Figure 2A–2B; Table 2; data not shown); we also expresed the csp-1C cDNA under the control of the mec-7 promoter and failed to observe killing of the touch neurons (data not shown). Ectopic expression of csp-1B from the ser-2d and flp-15 promoters killed the OLL and I2 neurons, respectively (ref. [31]; N. Bhatla and H.R. Horvitz, unpublished results). However, we noted that tm917, a csp-1 allele that deletes coding regions of only the csp-1A transcript, enhanced significantly (albeit weakly) the cell-death defects of ced-3(n2427) and ced-3(n2436) mutants, increasing the number of extra cells in their anterior pharynges by 0.9 and 1.2 cells, respectively (Table 1). dsRNA targetting the csp-1A prodomain (csp-1-PD) caused a similar slight enhancement of the cell-death defect of ced-3(n2436) mutants (Figure 1C), suggesting that, in addition to the more robust cell-killing activity of the csp-1B transcript, csp-1A might have a weak cell-killing function.
The proteolytic activity of caspases requires an active-site cysteine. Previously, it was shown that the CSP-1B protein can proteolytically process CED-3 in vitro and that this enzymatic activity required the active-site (SACRG) cysteine of CSP-1B, C138 [15]. We tested in vivo whether C138 was necessary by assaying the touch neuron-killing activity of mutant Pmec-7::csp-1B transgenes in which C138 was changed to a serine. We observed that the ectopic cell deaths were entirely dependent on the caspase active site (Table 2). Thus, csp-1B promotes cell death via caspase activity. The cell deaths induced by a Pmec-7::csp-1B transgene resulted in cell corpses with apoptotic characteristics (Figure 2C–2D). When observed with Nomarski optics, the csp-1B-induced cell deaths exhibited a refractile button-like appearance (Figure 2C) similar to that of developmental programmed cell deaths. Transmission electron micrographs of the cell corpses showed some contraction of the cytoplasmic volume and considerable condensation of the nuclear chromatin (Figure 2D), which are general characteristics of apoptotic cells, including those generated by ced-3 cell-killing transgenes (ref. [32]; data not shown). We conclude that csp-1B encodes a functional caspase that promotes programmed cell deaths with apoptotic morphology.
CED-3, like most caspases, is expressed as an inactive zymogen with an inhibitory N-terminal prodomain. Trans-auto-proteolysis of the CED-3 pro-protein at two aspartate residues removes the pro-domain and yields two subunits that form the active caspase [33]. CED-3 auto-activation is dependent on its prodomain and is facilitated by the association of two CED-3 pro-proteins within an octameric complex formed with the Apaf-1 homolog CED-4 [34]–[36]. Under normal cellular conditions, CED-4 is sequestered by CED-9 at mitochondria through a direct protein-protein interaction [37]–[39]. In response to upstream pro-apoptotic signals and the consequent expression of the BH3-domain-only protein EGL-1, which binds to and inhibits CED-9 [40], CED-4 is released from CED-9 and translocates to the nuclear periphery [37], [41], where it facilitates CED-3 activation [38]. Thus, the activation of CED-3 is controlled by an apoptosis pathway involving a BH3-domain-only protein, a member of the Bcl-2 family of apoptosis regulators, and a homolog of the apoptosome complex protein Apaf-1. The basic elements of this apoptosis pathway are evolutionarily conserved in mammals and are responsible for the activation of caspases in response to cell-intrinsic apoptotic stimuli [7].
Consistent with the role of ced-9 in negatively regulating ced-3 activation, it was previously shown that null mutations of ced-9 enhance the touch neuron-killing activities of Pmec-7::ced-3 transgenes [32]. (These experiments were performed using a ced-3(null) background to suppress the ced-3-dependent inviability of ced-9(null) animals.) Furthermore, this enhancement is dependent on ced-4 [32], indicating that the absence of CED-9 activates endogenous CED-4 within the touch neurons and that CED-4 activation elevates CED-3 activity. Unlike the CED-3 zymogen, CSP-1B lacks a long prodomain, suggesting that it might be activated via an alternative mechanism (i.e., independently of CED-4 and CED-9). To determine whether these canonical apoptosis regulators control CSP-1B activation, we introduced the ced-9(n2812) mutation into ced-3(n3692) strains carrying Pmec-7::csp-1B transgenes and assessed the effect of this ced-9 null mutation on PLM survival. In contrast to its effects on Pmec-7::ced-3–mediated PLM killing, ced-9(n2812) failed to enhance PLM killing in Pmec-7::csp-1B strains with a ced-3(n3692) mutant background (Figure 3A). Instead, ced-9(n2812) partially suppressed csp-1B-mediated PLM death (Figure 3A). CED-9 has a poorly understood pro-apoptotic activity in addition to its anti-apoptotic role in CED-4 inhibition [42], and it is possible that this ced-9 pro-apoptotic activity contributed to the deaths of cells expressing ectopic CSP-1B. Nevertheless, our results indicate that csp-1B-mediated cell killing, unlike ced-3-mediated cell killing, is not negatively regulated by ced-9 and suggest that CSP-1B is activated independently of CED-9.
We also observed that the expression of a Pmec-7::csp-1A transgene in ced-3(null) mutant strains failed to cause PLM cell death, even in a ced-9(null) background (Figure S1). These results suggest that the CSP-1A isoform (which contains a long prodomain similar to that of CED-3) does not promote programmed cell death, even in the absence of the anti-apoptotic protein CED-9. A role for csp-1A in cell death cannot be excluded entirely, as it is possible that endogenous CSP-1A requires a co-factor not present in the touch neurons to mediate cell killing.
Since CSP-1B can proteolytically cleave CED-3 in vitro [15], we tested whether the csp-1B cell-killing activing requires the endogenous ced-3 and ced-4 genes. The ced-3(n3692) and ced-4(n1162) mutations weakly suppressed csp-1B-mediated PLM death (Figure 3B), and it is possible that the endogenous csp-1 can in part promote programmed cell death through ced-3. Nonetheless, most csp-1B cell-killing activity was independent of ced-4 and ced-3 (Figure 3B). Loss of endogenous csp-1 failed to suppress PLM death in strains carrying Pmec-7::ced-3 or Pmec-7::ced-4 transgenes (Figure 3C–3D). Together, our results are consistent with a model in which csp-1B promotes programmed cell death at least mostly independently of and in parallel to the canonical apoptosis pathway (Figure 3E).
To determine which C. elegans cells express csp-1, we directly visualized endogenous csp-1 transcripts via fluorescence in situ hybridization (FISH) experiments using Cy5- and ALEXA-labelled probes complementary to the csp-1B transcript (i.e., targeted to all csp-1 transcripts) or to the csp-1A prodomain (specific to the csp-1A trancript). To our surprise, csp-1 mRNA was not detectable in the somatic cells of wild-type or egl-1(n1084 n3082) mutant embryos, larvae or adult hermaphrodites (data not shown). By contrast, csp-1 transcripts were present in the germlines of L4-stage larval and adult hermaphrodites (Figure 4A–4B). This expression was restricted to the late pachytene stage of meiosis I in both L4 larval gonads (in pachytene nuclei adjacent to differentiating sperm) and adult gonads (in pachytene nuclei adjacent to the bend of the gonad arm) (Figure 4A–4B). Both csp-1A and csp-1B/C transcripts were expressed in the adult pachytene germ cells, as indicated by the presence of FISH foci recognized by the csp-1A prodomain probes and foci recognized primarily by the csp-1B probes and only weakly by the csp-1A probes (Figure 4C). Stochastic and ionizing radiation (IR)-induced germline cell deaths occur during the late pachytene stage of oocyte development in adult gonads [43], [44]. However, csp-1 (unlike ced-3) was not required for either stochastic or IR-induced germline apoptosis, even in ced-3(n2436) strains sensitized for defects in germ-cell death (Figure 4D). In these experiments, apoptotic germ cells were identified using a transgene that expresses a functional GFP::CED-1 fusion protein that envelopes dying cells engulfed by the gonadal sheath [45], [46]. We also failed to detect differences in either stochastic or IR-induced germline cell death between csp-1 mutants and wild-type animals in experiments in which apoptotic germ cells were quantified by acridine orange staining or by direct observation of their refractile morphology using Nomarski optics (data not shown). We also noted that the level of csp-1 transcript expression in the germline (as determined by FISH) was not affected by either ionizing radiation or by mutation of egl-1 or ced-3 (data not shown).
Since we detected csp-1 expression in the adult germline but not in somatic cells of the embryo, we tested whether maternally supplied csp-1 transcript was necessary for the zygotic function of csp-1 in programmed cell death. Indeed, in sensitized genetic backgrounds (ced-3(n2427) and ced-3(n2436)), csp-1(+) progeny of csp-1(n4967) hermaphrodites (M−Z+ animals) had more undead pharyngeal cells than the csp-1(+) progeny of csp-1(+) hermaphrodites (M+Z+ animals) or the csp-1(n4967) progeny of csp-1(+) hermaphrodites (M+Z− animals) (Table 3). Thus, csp-1 expressed in the maternal germline is necessary for the csp-1 pro-apoptotic activity in embryonic programmed cell deaths. Given that we could not detect csp-1 expression in either embryos or larvae, it is therefore not surprising that the postembryonic programmed cell deaths of the ventral cord and postdeirid sensilla were unaffected by mutation of csp-1 (Table S4).
Most programmed cell deaths in C. elegans require ced-3 [20]. However, some cells die in mutants completely lacking ced-3. We previously reported that a subset of cells fated to die can be eliminated from ced-3 mutant embryos via a cell-shedding mechansm [19]. In that study, we noted that cell shedding from ced-3 mutants occurs independently of csp-1, csp-2 and csp-3: quadruple mutants lacking all four caspases also generate shed cells, indicating that cell elimination by this mechanism is completely caspase-independent [19]. Like most programmed cell deaths, the cells generated by caspase-independent extrusion are apoptotic in appearance. However, unlike caspase-dependent cell corpses, shed cells do not undergo phagocytosis by engulfing cells. The death of the male linker cell, which also occurs independently of ced-3, is non-apoptotic and requires the heterochronic protein LIN-29, its binding partner MAB-10 [47], and the polyglutamine repeat protein PQN-41 (ref. [18], [24]; Table S5). Previously it was shown that this cell death occurs in double-mutant males in which ced-3 and an additional csp gene (csp-1, csp-2 or csp-3) were inactivated [18]. We have now examined males lacking all four caspases and observed that the linker cell died in 100% of csp-3; csp-1; csp-2 ced-3 mutants (Table S5). The csp-3; csp-1; csp-2 ced-3 quadruple mutants were viable and fertile. Thus, both zygotic and maternal caspase contributions were eliminated. Our results therefore confirm that this cell death is indeed completely caspase-independent.
In addition, cell corpses are visible in the heads of larvae carrying null alleles of ced-3 (ref. [17]; Table 4). All programmed cell deaths in the developing heads of wild-type animals occur embryonically and are engulfed and degraded prior to hatching (ref. [30], [48]; Table 4). To detect ced-3-independent programmed cell deaths in larval heads, we used mutations (e.g., ced-1(e1735), ced-6(n2095) or ced-7(n1996)) that cause defects in cell-corpse engulfment and result in the persistence of many embryonic cell corpses into larval stages (ref. [49], [50]; Table 4). Like most wild-type cell corpses, the ced-3-independent cell corpses were refractile in appearance as observed with Nomarski optics and were not extruded from the animal (data not shown). We also observed that larvae mutant for ced-4 or egl-1 contained similar cell corpses, demonstrating that their generation does not require the canonical pro-apoptotic pathway that mediates most programmed cell deaths (Table 4).
We tested whether the small number of cell corpses visible in ced-3 larval heads are generated by the other C. elegans caspase genes and found that all double, triple and quadruple caspase mutants that we examined contained a small number of refractile corpses (Table 4). For example, 39% of csp-3; csp-1; ced-6; csp-2 ced-3 mutant animals contained at least one refractile cell corpse (Table 4), indicating that these programmed cell deaths occur in animals lacking all C. elegans caspases. We observed caspase-independent cell corpses in different regions of the larval head, including positions internal and external to the pharynx, which suggests that multiple cell lineages – at low frequencies – generated caspase-independent cell corpses. Surprisingly, we discovered that engulfment-competent ced-3 and csp-3; csp-1; csp-2 ced-3 mutants also contained refractile cell corpses (Table 4). The number of cell corpses per ced-3 or csp-1; csp-2 ced-3 larva increased until 12 to 24 hours post hatching (see below; data not shown), indicating that at least some of the cell deaths occurred after embryogenesis. Given that all programmed cell deaths in the head normally occur embryonically and that cell corpses are never observed in the heads of wild-type larvae, we concluded that timing of cell deaths in these ced-3 mutants was delayed. Thus, caspase-independent cell corpses can undergo an inefficient programmed cell death with slow kinetics in the absence of CED-3 activity, indicating that these cells likely die via CED-3-mediated apoptosis in wild-type animals.
Despite the strong causal link between caspase activation and apoptosis, recent studies have demonstrated that many morphological and biochemical changes associated with apoptosis can occur in the absence of caspases [4], [19], [21]. For example, in C. elegans the shed cells of csp-3; csp-1; csp-2 ced-3 quadruple mutants exhibit phosphatidylserine exposure, expression of the pro-apoptotic BH3-only gene egl-1, and chromatin condensation [19]. To determine whether these apoptotic attributes are evident in caspase-independent programmed cell deaths that do not involve extrusion of the dying cell from the embryo, we characterized the cell corpses visible in caspase-deleted larvae (Figure 5 and Figure 6). In most of these experiments, we used strains with the wild-type csp-3 allele, because (1) csp-3 lacks a caspase active-site [15]; (2) although previous studies reported that csp-3 has an anti-apoptotic function in somatic cells [22], we were unable to replicate those findings (Table S1); and, (3) the presence or absence of a csp-3 mutation had no effect on the frequency or appearance of caspase-independent corpses (Table 4; Figure 6B; data not shown). Like ced-3-mediated programmed cell deaths in wild-type animals, the caspase-independent corpses expressed egl-1, the upstream activator of the canonical apoptosis pathway (Figure 5A). Also, these cell corpses displayed phosphatidylserine on their cell surfaces, as indicated by the phosphatidylserine-binding reporter MFG-e8::Venus (Figure 5B), and exhibited many of the morphological hallmarks of apoptosis, including contraction of cytoplasmic volume and, in some but not all cases, condensation of nuclear chromatin (Figure 5C).
Additionally, we noted that the caspase-independent cell corpses frequently stained with acridine orange (Figure 6A), suggesting that these corpses are engulfed, internalized and degraded via endosomal pathways, as are canonical programmed cell deaths [30], [48], [49], [51]. Indeed, we found that the caspase-independent corpses were recognized by CED-1 (Figure 6B), a receptor expressed on engulfing cells required for the efficient phagocytosis of cell corpses [46], [49], [50]. The recognition of caspase-independent cell corpses by CED-1 appeared to be functionally important, as ced-1; csp-1; csp-2 ced-3 larvae contained more corpses than csp-1; csp-2 ced-3 larvae (Figure 6C). Given that ced-1 and other genes that function in cell-corpse engulfment promote programmed cell death [52], [53], it is unlikely that the ced-1(e1735) loss-of-function mutation caused additional cell deaths in the caspase-deleted mutants. Instead, the extra cell corpses in ced-1 mutant larvae likely reflected an engulfment defect, consistent with the comparatively rapid degradation and disappearance of most caspase-independent corpses in ced-1(+) larvae within the 36-hour period after hatching (Figure 6C). We conclude that caspases are not required for programmed cell deaths to be recognized by the engulfment machinery, internalized and degraded. In short, many aspects of apoptosis, including phagocytosis – the ultimate fate of apoptotic cells – can occur without caspases. We conclude that a parallel, caspase-independent pathway contributes to programmed cell death in C. elegans and can execute most cellular changes associated with apoptosis.
Our experiments revealed unexpected complexities in the execution of apoptosis in C. elegans. While the CED-3 caspase is clearly the primary effector of programmed cell death, we demonstrated the existence of additional caspase-dependent and caspase-independent contributions to developmental apoptosis. Specifically, we observed that maternally-expressed caspase gene csp-1 (but not csp-2 or csp-3) promotes the deaths of a subset of cells programmed to die during C. elegans embryogenesis (Figure 1 and Figure 4; Table 1 and Table 3). Furthermore, ectopic expression of the csp-1B isoform of csp-1 is sufficient to cell-autonomously kill cells that normally survive. These ectopic apoptotic cell deaths require the active site cysteine (C138) of CSP-1B, indicating that a caspase-like proteolytic function is responsible for its cell-killing activity (Table 2). The C. elegans genome therefore expresses at least two pro-apoptotic caspases, CED-3 and CSP-1B, to mediate programmed cell deaths. Nevertheless, the additional caspase activity conferred by csp-1 cannot account for ced-3-independent programmed cell deaths that have been observed in C. elegans. For example, the non-apoptotic death of the male linker cell and the extrusion of shed cells were already known to be caspase-independent [18], [19]. Here we demonstrate that cells in caspase-deleted animals can undergo an apoptosis-like programmed cell death followed by engulfment, indicating that the complete apoptotic program can occur in the absence of caspases. Thus, in addition to CED-3 and CSP-1B, there are caspase-independent cell-killing activities that contribute to programmed cell deaths.
The caspases CED-3 and CSP-1B appear to be regulated differently. The auto-activation of CED-3 is facilitated by the Apaf-1 homolog CED-4 in a protein-protein interaction that requires the CED-3 prodomain [34]–[36]. In the absence of a pro-apoptotic signal, CED-9 sequesters CED-4 [37], thereby preventing its association with the inactive CED-3 proprotein. The CSP-1B proprotein lacks a long prodomain, suggesting that it is not activated through an association with the CED-4 octamer in cells undergoing apoptosis. Consistent with this expectation, we observed that the cell-killing activity of csp-1B transgenes, unlike that of ced-3 transgenes, was not negatively regulated by ced-9 (Figure 3). Furthermore, based on our genetic experiments (Figure 3) and the in vitro studies of Shaham [15], it does not appear that CSP-1B is activated by CED-3. We therefore propose that CSP-1B is regulated by a mechanism different from the canonical programmed cell death pathway that activates CED-3 and that CSP-1B likely promotes cell killing in parallel to CED-3 (Figure 3E).
There are no known or candidate regulators of csp-1. It is possible that csp-1 is controlled entirely at the transcriptional level and that csp-1 contributes a minor, sub-lethal pro-apoptotic activity to all cells within the C. elegans embryo. Indeed, only using sensitized backgrounds with partial defects in programmed cell death did we detect the pro-apoptotic function of csp-1. Nevertheless, we expect that it will be possible to identify regulators and effectors of csp-1 through genetic screens for mutants that modify the cell-killing activity of csp-1B transgenes.
Given the minor contribution of csp-1 to programmed cell death and the lack of a detectable role of csp-2 or csp-3 in apoptosis (Table 1; Table S1; data not shown), it is tempting to speculate that the csp genes have non-apoptotic functions in C. elegans. In C. elegans, ced-3 functions in axon regeneration following laser axotomy [54]. In mammalian and Drosophila neurons, caspases have functions in dendritic pruning, axon guidance and the synaptic changes underlying long-term depression [14]. Caspase function is also required for the maturation of Drosophila sperm [55]. Interestingly, we observed robust expression of csp-1 in the germlines of L4 and adult hermaphrodites, specifically in the late pachytene nuclei (Figure 4). We also observed temporally and spatially restricted csp-2 and csp-3 mRNA expression in the late pachytene nuclei of the L4 larval germline (data not shown), suggesting that the csp genes might have functions in germ cell development. However, mutant hermaphrodites and males carrying all tested combinations of csp-1, csp-2 and csp-3, including the triple csp mutant were viable, fertile and failed to exhibit obvious brood-size defects that would suggest abnormalities in sperm or oocyte differentiation (data not shown).
Genetically encoded cell-killing activities provide an efficient and convenient method for determining cellular function through cell ablation. Killer genes such as ced-3 have been used under the control of various promoters to ablate specific cells [32], [45], [56], [57]. However, the potent cell-killing activity of ced-3 transgenes can cause organismic inviability, particularly if the promoter expression is not exclusive to a small number of cells (see below). csp-1B overexpression using the mec-7 and flp-15 promoters efficiently killed the touch and I2 neurons, respectively (Figure 2; Table 2; N. Bhatla and H.R. Horvitz, personal communication). The mec-7 and flp-15 promoters are relatively strong, as they also robustly induced gfp expression in these cells, such that the neural processes were visible with a dissecting microscope equipped with fluorescence optics. By contrast, the odr-1 promoter did not produce detectable GFP expression in the neurites of the AWB, AWC and I1 neurons, and csp-1B under the control of the odr-1 promoter failed to kill these cells even when injected at plasmid concentrations as high as 100 ng/µl (N. Bhatla and H.R. Horvitz, unpublished results). Thus, high levels of csp-1B expression might be required to kill most cells, making the use of csp-1B as a cell-ablation tool appropriate in situations in which the promoter sequence strongly drives expression in targeted cells and/or weakly promotes expression in additional cells not intended to be targets. For example, the Pmec-7::csp-1B constructs, which were injected at a concentration of 15 ng/µl, produced csp-1B expression outside of the touch neurons that was detectable by fluorescence in situ hybridization. However, this level of csp-1B expression was sub-lethal and did not induce cell death or other cellular defects outside of the touch neurons (data not shown). By contrast, Pmec-7::ced-3 constructs were toxic to the animals when injected at concentrations above 1 ng/µl, suggesting that cells are very sensitive to ectopic ced-3 and that using ced-3 as a cell ablation tool is potentially problematic when promoter expression is not restricted to a small number of targeted cells.
Although the csp-1 gene contributes a cell-killing activity to normal programmed cell deaths (Table 1), csp-1 and the other csp genes are not responsible for the ced-3-independent programmed cell deaths present in the heads of ced-3 larvae (Table 4). These deaths, like those of the male linker cell (ref. [18]; Table S5) and the embryonic shed cells [19], are caspase-independent – a surprising result in light of our observations that these cell corpses are morphologically apoptotic (Figure 5) and are engulfed (albeit with slower kinetics) like normal programmed cell deaths (Figure 6). Thus, the complete apoptotic program including cell-corpse internalization can occur in the absence of caspases in C. elegans, suggesting that the cellular changes accompanying apoptosis do not require proteolysis by the caspase family of proteases. Moreover, it is clear that apoptotic programmed cell deaths are achieved through the integration of independent cell-killing activities from CED-3, CSP-1B and an unknown caspase-independent source.
Given the minor cell-killing effects of the CSP-1B and the caspase-independent pathways, why might cell-killing activities in addition to that of CED-3 have evolved? It is possible that different cells, even within the set of C. elegans cells fated to die, are differentially sensitive to pro-apoptotic signals and that additional caspase and caspase-independent pathways ensure efficient and complete cell death under diverse environmental and developmental conditions. Interestingly, the postembryonic programmed cell deaths of the ventral cord are more sensitive to weak ced-3 mutations than are the embryonic programmed cell deaths in the presumptive anterior pharynx: ced-3 mutations that have weak effects in the anterior pharnyx typically have stronger effects in the ventral cord (ref. [17]; data not shown). We observed a complementary function for csp-1, which promotes apoptosis in the anterior pharynx (Table 1) but not in the ventral cord (Table S4).
In summary, multiple pro-apoptotic caspases function in programmed cell death in C. elegans, Drosophila and vertebrates. Furthermore, as we and others have shown, there are additional caspase-independent contributions to programmed cell deaths in C. elegans. We identified C. elegans caspase-independent cell deaths that are essentially identical to wild-type programmed cell deaths based on their apoptotic appearance and their recognition and internalization by engulfing cells. We expect that caspase-independent pro-apoptotic activities are present in other metazoans and that their identification will be of major importance to our understanding of cell death in the contexts of development and disease.
All C. elegans strains were cultured as described previously [58] and maintained at 20°C. We used Bristol N2 as the wild-type strain, and the mutations used in our experiments are listed below:
LG I. unc-75(e950), ced-1(e1735), csp-3(n4872, tm2260, tm2286), nIs177[Pceh-28::gfp] [59]
LG II. csp-1(n4967, n5133, tm917), mab-10(n5117), lin-29(n836)
LG III. ced-4(n1162, n3158), ced-6(n2095), ced-7(n1996), ced-9(n1653, n2812), tat-1(tm1034), nIs308[Pmec-7::csp-1B, Pmec-3::gfp], nIs400[Pced-1::ced-1ΔC::gfp] [19]
LG IV. csp-2(n4871), ced-5(n1812), dpy-20(e1282), unc-30(e191), ced-3(n2427, n2436, n2452, n3692), nIs309[Pmec-7::csp-1B, Pmec-3::gfp]
LG V. egl-1(n1084 n3082), bcIs39[Plim-7::ced-1::gfp] [45], nIs342[Pegl-1::4×NLS::gfp] [59], qIs56[Plag-2::gfp]
LG X. ced-8(n1891), bzIs8[Pmec-4::gfp] [22], nIs106[Plin-11::gfp] [52]
Unknown linkage. nIs290[Pmec-3::gfp]; nIs307[Pmec-7::csp-1B, Pmec-3::gfp], nIs368-370[Pmec-7::csp-1B(C138S), Pmec-3::gfp], nIs398[Pdyn-1::mfg-e8::Venus] [19], [60]
Extrachromosomal arrays. nEx1646[Pdyn-1::mfg-e8::Venus] [19], [60], nEx1465-71[csp-1(+) (pDD027)], nEx1604-9[csp-1B/C only (pDD030)], nEx1614-16[csp-1A only (pDD029)], nEx1617-19[csp-1-PD (pDD028)]
The Pmec-7::ced-3, Pmec-7::ced-4 [32], Pdyn-1::mfg-e8::Venus [60], Plim-7::ced-1::gfp [45], Pced-1::ced-1ΔC::gfp [46], Plin-11::gfp [52], Pegl-1::gfp and Pceh-28::gfp [59] plasmids were described previously. The csp-1 rescuing plasmid (pDD027) was constructed using PCR to amplify a 9 kb fragment of the csp-1 genomic locus with the primers 5′-gtaacgccagggttttcccagtcacgacggtgatccttcggagcttcag and 5′- acgaggatatccgcattgag. The resulting amplicon was ligated via the TOPO-TA subcloning protocol into the pCR2.1 vector (Invitrogen). pDD028 (csp-1-PD), pDD029 (csp-1A only), and pDD030 (csp-1B/C only) were constructed using site-directed PCR mutagenesis. Two early stop codons in the csp-1B/C isoforms were generated in pDD028 using the primer 5′-ccgagaacggacgcctagtaatcgaaccataaac and its reverse complement. The csp-1B/C start codon was mutated to an alanine codon in pDD029 using the primer 5′-gactctcagagtcgagcgccgagaacggacgcc and its reverse-complement. Two early stop codons in the csp-1A isoform were generated in pDD030 using the primer 5′cctgaaaacgatagaagataattgataatcacaattcgacgatgatttgg and its reverse complement. The Pmec-7::csp-1A plasmid (pDD003) was constructed using PCR to amplify the csp-1A cDNA from pDD006 using the primers 5′-gcggctagcatggtcctgaaaacgatagaag and 5′-gcgccatggttacatcgaccttgaaaagtgcc, which incorporate the restriction sites NheI and NcoI, respectively, into the resulting amplicon. The csp-1A amplicon was digested with NheI and NcoI and then ligated into the vector pPD52.102. The Pmec-7::csp-1B plasmid (pDD002) was constructed by using PCR to amplify the csp-1B cDNA from pDD001 using the primers 5′-gcggctagcatgccgagaacggacgccaag and 5′-gcgccatggttacatcgaccttgaaaagtgcc, which incorporate the restriction sites NheI and NcoI, respectively. The csp-1B amplicon was digested with NheI and NcoI and then ligated into the vector pPD52.102, which encodes the mec-7 promoter. The Pmec-7::csp-1B(C138S) plasmid (pDD005) was constructed from pDD002 using PCR with the primers 5′-tggatgaactatacaaatagctgcgctccagcgcgttcgt and its reverse complement. The RNAi plasmid pL4440::csp-1-PD (pDD060) was constructed using PCR to amplify the prodomain encoding fragment of the csp-1A cDNA with the primers 5′-gcgagatctatggtcctgaaaacgatagaag and 5′-cgcctcgagatggcgggtttcagctgggtc, which incorporate the restriction sites BglII and XhoI, respectively. The resulting csp-1-PD amplicon was digested with BglII and XhoI and then ligated into pL4440. The RNAi plasmid pL4440::csp-1B (pDD061) was constructed using PCR to amplify the csp-1B cDNA with the primers 5′-gcgagatctatgccgagaacggacgccaag and 5′-cgcctcgagttacatcgaccttgaaaagtgcc, which incorporate the restriction sites BglII and XhoI, respectively. The resulting csp-1B amplicon was digested with BglII and XhoI and then ligated into pL4440.
The in vitro transcription, purification, preparation and microinjection of csp-1-PD (pDD060) and csp-1B (pDD061) dsRNA were performed as described previously [61].
The fixation of embryos and larval and adult animals, the conjugation of Cy5 or ALEXA594 fluorescent probes to in situ oligo probes, and the hybridization of oligos to fixed samples were performed as described previously [62]. All images were acquired using an inverted Nikon TE-2000 compound microscope equipped for fluorescence microscopy (Prior Scientific). Images were acquired with a PIXIS camera (Princeton Instruments) controlled by MetaMorph software (Molecular Devices) and modified for publication with ImageJ software (NIH). The “total csp-1” set of probes included 32 distinct 20-nucleotide sequences complementary to csp-1B (Biosearch Technologies, Inc). This set of oligos was conjugated to the fluorophore Cy5 (GE Healthcare) and hybridized to all three csp-1 mRNA isoforms (csp-1A, csp-1B and csp-1C). The “csp-1A” set of probes included 32 distinct 20-nucleotide sequences complementary to the region of csp-1A that encodes the prodomain. This set of oligos was conjugated to the fluorophore ALEXA594 (Invitrogen) and hybridized specifically to the csp-1A mRNA isoform. Probe sequences are listed in Table S6.
The numbers of undead cells that failed to undergo programmed cell death in the anterior pharynges and postdeirid sensilla of L3 larvae were determined by direct observation using Nomarski optics as described previously [28]. Persistent cell corpses in larval heads also were quantified by direct observation using Nomarski optics; for this assay, larvae were staged by the time of hatching. For other cell-death assays, the ventral cord cells of young adults, the M4 neuron and its undead sister cell of L3 larvae, the touch neurons of L4 larvae, and the germ cell corpses of adult hermaphrodite gonads were identified using previously described GFP reporter transgenes [45], [52], [59]. For experiments involving ionizing radiation, L4 larvae were exposed to gamma irradiation from a Co-60 source. All strains were analyzed using a Zeiss Axioskop II compound microscope equipped for fluorescence microscopy. Images were acquired with an ORCA camera (Hammamatsu) controlled by OpenLab software (Perkin Elmer) and modified for publication using ImageJ (NIH).
L1-stage larvae were fixed, stained and sectioned for transmission electron microscopy as described previously [43]. Stained sections were imaged with a JEM-1200EX II microscope (JEOL) using an AMT XR41 CCD camera.
|
10.1371/journal.pbio.1000545 | The Translation Initiation Factor 3f (eIF3f) Exhibits a Deubiquitinase Activity Regulating Notch Activation | Activation of the mammalian Notch receptor after ligand binding relies on a succession of events including metalloprotease-cleavage, endocytosis, monoubiquitination, and eventually processing by the gamma-secretase, giving rise to a soluble, transcriptionally active molecule. The Notch1 receptor was proposed to be monoubiquitinated before its gamma-secretase cleavage; the targeted lysine has been localized to its submembrane domain. Investigating how this step might be regulated by a deubiquitinase (DUB) activity will provide new insight for understanding Notch receptor activation and downstream signaling. An immunofluorescence-based screening of an shRNA library allowed us to identify eIF3f, previously known as one of the subunits of the translation initiation factor eIF3, as a DUB targeting the activated Notch receptor. We show that eIF3f has an intrinsic DUB activity. Knocking down eIF3f leads to an accumulation of monoubiquitinated forms of activated Notch, an effect counteracted by murine WT eIF3f but not by a catalytically inactive mutant. We also show that eIF3f is recruited to activated Notch on endocytic vesicles by the putative E3 ubiquitin ligase Deltex1, which serves as a bridging factor. Finally, catalytically inactive forms of eIF3f as well as shRNAs targeting eIF3f repress Notch activation in a coculture assay, showing that eIF3f is a new positive regulator of the Notch pathway. Our results support two new and provocative conclusions: (1) The activated form of Notch needs to be deubiquitinated before being processed by the gamma-secretase activity and entering the nucleus, where it fulfills its transcriptional function. (2) The enzyme accounting for this deubiquitinase activity is eIF3f, known so far as a translation initiation factor. These data improve our knowledge of Notch signaling but also open new avenues of research on the Zomes family and the translation initiation factors.
| The highly conserved signaling pathway involving the transmembrane receptor Notch is essential for development, and misregulation of this pathway is linked to many diseases. We previously proposed that the Notch1 receptor is monoubiquitinated during its activation. With the aim of identifying a deubiquinating enzyme that could regulate Notch activation, we demonstrated that eIF3f, known previously as part of the multiprotein translation initiation factor eIF3 complex, harbors an enzymatic activity that acts on Notch. The activated form of Notch is able to interact with eIF3f only in the presence of the E3 ubiquitin ligase Deltex, and Notch needs to be deubiquitinated before it can be cleared and its intracellular domain can enter the nucleus and fulfill its transcriptional function. Our results further decipher the molecular mechanisms of Notch signaling activation, showing that ubiquitination and deubiquitination events are required. Additionally, we show that beyond acting as a translation initiation factor, eIF3f fulfills other functions and has an intrinsic enzymatic activity.
| Notch signaling relies on two consecutive cleavages of the receptor after binding of its ligand expressed by a neighboring cell. These two processing steps successively performed by a protease of the ADAM family and by the γ-secretase complex can occur only if the activated receptors on one side, the ligands on the other side, undergo post-translational modifications and trafficking. Some of these complex events begin to be elucidated [1]–[7]. They essentially depend on ubiquitination events affecting the ligand and/or the receptor, and probably regulating sorting and trafficking of the activated versus non-activated molecules. Eventually, after proteolytic release the intracellular portion of Notch (hereafter named NIC) enters the nucleus, where it functions as a transcriptional co-activator of Notch target genes.
In mammals, the Notch1 receptor was proposed to be monoubiquitinated before its γ-secretase cleavage; the targeted lysine has been localized to its submembrane domain [8]. Investigating how this monoubiquitination is regulated may be crucial for understanding Notch receptor activation and downstream signaling. Ubiquitination is a reversible process, and deubiquitinating enzymes (DUBs) remove the ubiquitin moieties from ubiquitinated substrates, thus allowing a tight control of these modifications [9]. A potential deubiquitination step could either affect NIC production by γ-secretase, NIC release from the endocytic vesicles, NIC entry into the nucleus, NIC interaction with its transcriptional cofactors, NIC transcriptional activity, or NIC stability. With the aim of identifying a DUB involved in Notch signaling, we established a screening strategy using shRNA vectors targeting the putative and known DUBs of the human genome [10]. Here, we report the identification of eIF3f as a DUB targeting the activated Notch receptor and positively regulating Notch signaling.
eIF3f (for eukaryotic translation initiation factor 3 subunit f) is one of the 13 subunits (named eIF3a-m) of the translation initiation factor eIF3. eIF3 stimulates many steps of the translation initiation pathway, including assembly of the eIF2-GTP/met-tRNA complex to the 40S ribosome to form the 43S preinitiation complex (PIC), mRNA recruitment to the 43S PIC complex, impairment of the 40S ribosome to join the 60S prematurely, and scanning the mRNA for AUG recognition [11]. eIF3 has no known enzymatic activity so far, but it has an intriguing degree of homology with two other complexes whose functions appear unrelated: the COP9 signalosome and the 19S proteasome lid. All three complexes, forming the Zomes family, consist of subunits with either PCI (Proteasome-COP9 signalosome-initiation factor 3 domain) or MPN (for MPR1-PAD1-N-terminal domain) signature domains and share a common 6PCI +2 MPN domain stoechiometry. The mammalian eIF3 also has an additional five non-PCI-MPN subunits [12]. The MPN domain of CSN5 (COP9 Signalosome subunit 5) harbors a metalloprotease motif referred to as the Jab/MPN domain-associated metallopeptidase (JAMM) motif and regulates the activity of E3 ubiquitin ligases by deneddylation of the cullin component. On the other hand, a JAMM-containing subunit associated with the 19S proteasome lid (Rpn11, [13]) also harbors DUB activity, accounting for substrates deubiquitination before they enter the proteasome channel. Interestingly, eIF3f contains a JAMM domain, making it a putative DUB [14],[15]. We show here that it harbors a DUB activity acting on Notch signaling.
In order to identify DUBs involved in the Notch signaling pathway, we set up an immunofluorescence screen. We used Notch ΔE [16], a mutant form of Notch deleted of most of its extracellular domain. ΔE mimics the ADAM-cleavage product and therefore represents a constitutively active form, which is monoubiquitinated and endocytosed [8] before being cleaved by γ-secretase to liberate NIC [17]. We made use of the V1744 antibody to monitor the production and the localization of NIC, whereas anti-myc antibody detected all Notch products. U2OS cells were co-transfected with vectors encoding ΔE and with each individual pool of a shRNA library targeting the 91 known or putative DUBs encoded by the human genome (called shDUB library; see Table 1) [10]. With ΔE alone (Figure 1A, Panel A), we observed a membrane and endomembrane-localized myc labeling corresponding to the unprocessed Notch ΔE, but also a nuclear myc labeling (A3) corresponding to NIC, co-stained with the V1744 antibody (A4). The same pattern was observed with almost every shRNA pool of the library (exemplified in Panels B1–4), whereas four pools abolished ΔE expression (those targeting eIF3h, PRP8 and USP54, and Pool 2, see below). However, with an shRNA pool targeting eIF3f (panels C), we observed a partial extra-nuclear V1744 labeling (C4), as well as an increase in the average proportion of extranuclear/nuclear myc staining (see Figure S1, Dose 1). This suggests that NIC production and localization was affected when eIF3f was knocked down.
The shRNA library actually contained two pools (each containing four shRNA-encoding plasmids) targeting eIF3f: Pool 1, whose effects are shown in Figure 1A, and Pool 2, in the presence of which no ΔE-positive signal could be detected by immunofluorescence (unpublished data). We isolated the eight shRNAs from these pools (six different shRNAs in total) and tested them individually for their efficiency in knocking down eIF3f: Pool 1 (P1) contains shRNA #1 to #4, Pool 2 (P2) contains shRNA #1, #2, #5, and #6. HEK293T cells were transfected with each of these shRNAs and the levels of endogenous or overexpressed human HA-tagged eIF3f were analyzed by Western blotting. Anti-α-Tubulin was used as a loading control (Figure 1B, bottom panel), since this protein is very stable and reflects the total protein content in transfected and non-transfected cells. shRNAs #1 and #2 significantly affected the level of endogenous and transfected eIF3f (Figure 1B, Lanes B and C and quantification under the lanes), similar to the Pool 1 (P1, Lane I). The additional shRNAs from Pool 2 (#5 and #6 in Lanes J and K) and Pool 2 itself (P2, Lane L) almost completely abolished eIF3f expression in this assay. The effect of each of these shRNAs on eIF3f protein level was correlated with its ability to inhibit ΔE expression in immunofluorescence experiments and probably reflects eIF3f requirement for translation (Figure S1). Indeed shRNAs #3 and #4, which exhibited a minor effect on the expression of the eIF3f protein (Figure 1B, Lanes D–G), were the most efficient in inducing extranuclear V1744 labeling in a dose-dependent manner (Figure 1C). Nevertheless, further increasing the dose of these two shRNAs separately or co-transfecting them resulted in an effect on eIF3f expression (Figure 1B, Lanes E, G, and H) and an extinction of the ΔE signal in immunofluorescence (Figure S1 and unpublished data). Since the shRNAs that affect Notch signaling target different sequences of eI3Ff, and since this effect is complemented by transfection of wt eIF3f (see below), we can exclude an off-target effect of these shRNAs. Taken together, these results suggest that a specific but mild knockdown of eIF3f is responsible for the altered Notch localization observed in Figure 1.
We next tested the effect of eIF3f on the ubiquitination of activated Notch. HEK293T cells were transfected with vectors encoding ΔE and a 6xHis-tagged Ubiquitin. We added increasing amounts of shRNA #3, without reaching the doses used in Figure 1B, Lane E, in the presence or not of a murine form of eIF3f (meIF3f). The murine eIF3f cDNA exhibits a three base pair change in the sequence targeted by shRNA #3 and is therefore refractory to its effect. Proteins were extracted in denaturing conditions and ubiquitinated proteins were purified on Nickel-charged beads. Finally, whole cell extracts and ubiquitinated products were analyzed by Western blot to quantify the levels of ubiquitinated Notch (Figure 2A). Transfection of shRNA #3 led to a dose-dependent accumulation of monoubiquitinated ΔE (ΔEUb) and of monoubiquitinated NIC (NICUb) (Lanes C, D compared to B), whereas the levels of Notch in the extracts remained stable (Lanes G to J). The same effect was observed using shRNA #4 (unpublished data). Interestingly, overexpression of meIF3f abolished the effect of shRNA #3 and accumulation of ubiquitinated Notch could no longer be seen (compare Lanes E to D). The ubiquitination levels of other forms of Notch were also analyzed in the presence or absence of shRNA #3 (Figure 2B): the membrane-anchored ΔE-LLFF (Lanes D–F), which undergoes the same processes as ΔE (i.e., monoubiquitination and endocytosis) but which is mutated in the γ-secretase cleavage site and consequently does not generate NIC [18]; the nuclear Notch (NIC, Lanes G–I) and the non-activated full-length Notch (FL, Lanes J–L). The only forms whose ubiquitination was affected by shRNA #3 and/or meIF3f were those corresponding to activated and still membrane-anchored forms of Notch, namely ΔE and ΔE-LLFF. We also tested as controls NDFIP2 (a transmembrane protein located in the endosomal compartment [19]) and Deltex1 (an E3 ubiquitin ligase genetically identified as involved in Notch signaling). These two proteins showed no change in ubiquitination in the presence of shRNA #3 or overexpressed meIF3f (Figure S2 and unpublished data). These results suggest that eIF3f is specifically involved in an ubiquitination/deubiquitination process targeting activated Notch.
The eIF3f protein contains an MPN domain, also found in some DUBs of the JAMM family, such as AMSH, Rpn11, and CSN5 [20]–[23]. However, the position of the amino acids constituting the catalytic site of these DUBs is not strictly conserved in eIF3f, although histidines and acidic residues can still be found and can possibly form a signature of metalloprotease (HEX2HX2GX2H). We mutated six amino acids of eIF3f, including two histidines and two acidic residues of the putative catalytic site. We then tested in parallel WT meIF3f and this putative catalytic mutant (meIF3f Mut) (Figure 2C). Whereas WT meIF3f was able to prevent shRNA #3–induced NICUb accumulation in a dose-dependent manner (Lanes D, E compared to C), meIF3f Mut was not, although it was expressed at similar levels (Lanes D to G). This shows that meIF3f indeed complements the effect of shRNA #3 and that an intact MPN domain of eIF3f is necessary for its effect on monoubiquitination of activated Notch. Taken together, these results strongly suggest that eIF3f could act as a DUB targeting activated Notch.
In order to test whether eIF3f exhibits a deubiquitinase activity, we first used a functional in bacteria assay using GFP fused to ubiquitin (Ub-GFP) [24]. The peptide bond between ubiquitin and GFP can be cleaved by ubiquitin- or UbL-deconjugase activities, which are absent from the bacterial genome. Bl21 bacteria were transfected with plasmids encoding GST alone, GST-fused to human WT eIF3f or WT MPN domain, or His-tagged murine eIF3f WT or mutant MPN domain (Mut), together with a vector encoding Ub-GFP fused to an S-Tag at its C-terminus (Ub-GFP-S-Tag) (Figure 3A). As control DUB, we used BPLF1 WT (an EBV DUB of the cysteine-protease family [24]) and its catalytically inactive form BPLF1 CM. After induction of protein expression, the bacteria were lysed and cleavage of Ub-GFP was assessed in Western blot by the appearance of free GFP-S-Tag using anti-S-Tag antibody (Figure 3A, upper panel). Anti-GST or anti-His antibodies were used to verify protein expression (Figure 3A, bottom panel). As expected, we observed released GFP with BPLF1 WT but not its CM mutant (compare Lanes A–C). Interestingly, we also detected protease activity with heIF3f WT, heIF3f MPN WT, and meIF3f MPN WT (Lanes D, E, and F, respectively) but not with the mutant meIF3f MPN Mut (Lane G).
To further investigate whether eIF3f could act as an ubiquitin-specific protease, we performed an in vitro assay using Ubiquitin Vinyl Sulfone (Ub-VS), a functional probe that covalently binds the active site of DUBs and consequently inhibits DUB catalytic activity (Figure 3B). HeLa cells were transfected with vectors encoding HA-tagged meIF3f, either WT or Mut. Flag-tagged BPLF1 WT and CM were used as controls. A portion of the corresponding whole cell extracts was incubated in vitro with Ub-VS. Total extracts incubated or not with Ub-VS were finally analyzed by Western blot using anti-HA and anti-Flag antibodies to monitor the covalent binding of Ub-VS to the DUBs (Figure 3B, Lanes A to H) and to check protein expression (Lanes I to N), respectively. We observed a partial upshift corresponding to the size of the Ub-VS probe with meIF3f WT (Lane F compared to E) and with BPLF1 WT (Lane B compared to A), indicating that both are able to bind Ub-VS. In contrast, we did not observe any shift with meIF3f Mut (Lane H compared to G) nor BPLF1 CM (Lane D compared to C), showing that they are indeed unable to bind Ub-VS. We obtained the same results when the DUBs were expressed in bacteria (unpublished data).
Taken together, these results show that eIF3f exhibits a deubiquitinase activity, carried by its MPN domain. Moreover it confirms that the mutant we generated (Mut) by replacing six amino acids of the metalloprotease-like sequence of eIF3f is a catalytically inactive form of eIF3f.
In order to identify the site of action and the target of eIF3f in the Notch activation cascade, we first overexpressed ΔE and murine eIF3f in U2OS cells. We observed by immunofluorescence a low frequency of vesicular colocalization of the two proteins (Figure 4, Panels A and B). As eIF3f might require a cofactor or target an intermediate protein to regulate Notch ubiquitination, we tested several components of the Notch signaling pathway that could be associated with trafficking. Some DUBs are known to be recruited to their substrate indirectly via an E3 ubiquitin ligase [25],[26], so we particularly focused on the E3 ubiquitin ligases of the Notch pathway known to be associated with the endocytic machinery: Itch/AIP4 and Deltex1 (hereafter designated as DTX) ([27] and therein). In contrast to Itch/AIP4, DTX significantly colocalized with eIF3f when both proteins were coexpressed (Figure 4, Panel C). In addition, the presence of DTX strikingly increased the colocalization of ΔE and eIF3f, the three proteins being associated to the same vesicles (Figure 4, Panel D). These results suggest that DTX could recruit eIF3f to activated Notch.
To verify this hypothesis, we performed co-immunoprecipitation experiments in HEK293T cells transfected with vectors encoding VSV-tagged DTX, Flag-Itch/AIP4, and HA- or Flag-tagged forms of eIF3f (schematized in Figure 5A). We pulled down eIF3f and analyzed the whole cell extracts and the immunoprecipitates by Western blot (Figure 5B, 5C and Figure S3). DTX co-immunoprecipitated with eIF3f WT and (1–192), but also with eIF3f (188–361) and (91–361) (compare Lanes B, C to D, E in Figure 5B and 5C). No co-immunoprecipitation was detected with Itch/AIP4 (Figure S3). As a control, the endogenous eIF3a, another subunit of the eIF3 complex, only co-immunoprecipitated with eIF3f WT and (91–361) (Panels B, C). This observation suggests that DTX is able to physically interact with eIF3f and that the domain(s) necessary for this interaction is different from the domain necessary for eIF3f to incorporate the eIF3 complex. We then performed the reverse experiment by pulling down VSV-tagged DTX and could detect the various forms of eIF3f coimmunoprecipitating with DTX (Figure 5D): WT or Mut eIF3f (Lanes B, C) as well as the deletion mutants (Lanes D–F). To confirm these results under conditions where the proteins are not overexpressed, we established by retroviral transduction murine cell lines expressing low amounts of VSV-DTX and S-tagged forms of WT eIF3f or of a mutant of the active site (HDI to AAA). eIF3f or DTX was immunoprecipitated first and their association was confirmed in both cases (Figure 5E).
We then performed co-immunoprecipitations in HEK293T in the presence of Notch ΔE (Figure 6A). While DTX co-immunoprecipitated with WT eIF3f but also with meIF3f Mut (Figure 6A, Lanes B–E), ΔE did not (Lanes I and J), unless DTX was cotransfected (Lanes D and E). This strongly suggests that a tripartite interaction occurs between activated Notch, DTX, and eIF3f, DTX being required for the Notch-eIF3f interaction. In addition, we verified that neither DTX nor ΔE could be co-immunoprecipitated with eIF3f from cell extracts that were transfected separately and mixed (unpublished data). We then repeated these experiments using other forms of Notch: ΔE-LLFF, NIC, or the non-activated full-length Notch (FL) (Figure 6B). Only two forms could co-immunoprecipitate with eIF3f in the presence of DTX: ΔE and ΔE-LLFF (Lanes C and D). No signal was detected with FL and NIC (Lanes E and F) even with a longer exposure. It is of note that NIC produced from ΔE (Lanes C, G) and detected by the V1744 antibody was not co-immunoprecipitated with eIF3f either. These results show that the tripartite interaction between Notch, DTX, and eIF3f occurs preferentially with activated, membrane-associated, but γ-secretase unprocessed, forms of Notch.
Our results (Figures 2 and 6) suggest that the target of eIF3f is preferentially a γ-secretase unprocessed form of activated Notch. The fact that we could detect ubiquitinated NIC in the presence of shRNAs #3 and #4 (Figure 2A) might thus appear paradoxical, however it might be the consequence of γ-secretase cleavage of non-deubiquitinated Notch ΔE. We also observed that shRNAs #3 and #4 led to a partially extranuclear localization of NIC (Figure 1C). This suggests that NIC carrying residual monoubiquitination could be impaired in nuclear translocation. In order to test this possibility, we tried to mimic NICUb by attaching an ubiquitin to the N-terminus of NIC (UBIC construct). Given that ΔE, and consequently any putative non-deubiquitinated NIC, has been shown to be monoubiquitinated on K1749, we used a K1749R mutant of NIC (NIC(KR)) on which no ubiquitin can be conjugated. In addition, the attached ubiquitin was mutated on the internal Lysine residues 29 and 48 to impair polyubiquitination and also on the 2 C-terminal glycines to prevent proteolysis by a DUB enzyme [28]. U2OS cells were transfected with NIC, NIC(KR), or the new construct UBIC. We observed by immunofluorescence that NIC or NIC(KR) were mostly nuclear (96% in average), whereas UBIC was partially retained in the cytoplasm (18% see Figure S4, Panel A). In addition, we monitored the transcriptional activity of these three forms by cotransfecting U2OS cells with increasing doses of expression vectors together with a Notch-reporter gene (CSL-luciferase, [29],[30]) and an internal control reporter (pRL-TK). As shown in Figure S4, Panel B, UBIC is significantly less active than NIC or NIC(KR), although it is expressed at a comparable level (see Western blot in Figure S4, Panel B bottom). One possibility to explain the localization and the drop in transcriptional activity of UBIC is that the ubiquitin moiety partially prevents access to the NLS of NIC.
In order to test whether eIF3f could act on Notch signaling under more physiological conditions, we performed a coculture assay using a CSL reporter strategy. U2OS cells stably expressing Notch FL were transfected with a CSL-Luciferase Notch reporter and increasing doses of meIF3f WT, meIF3f Mut, or meIF3f (188–361). pRL-TK vector, encoding Renilla luciferase under the control of the Notch-insensitive thymidine kinase promoter, was also cotransfected as an internal control. These cells were then cocultured with OP9 cells stably expressing or not the Notch ligand Delta-like1 (Dll1) [31], and relative luciferase activity was finally determined by normalizing CSL-firefly luciferase with renilla luciferase. In parallel cell extracts from the same transfections were analyzed by Western blot (Figure 7A, bottom). In the presence of Dll1, the relative luciferase activity of the CSL reporter gene increased 20-fold (Figure 7A, Lanes A, B), showing that Notch was indeed activated by Dll1. While the presence of meIF3F WT did not modify Dll1-dependent Notch activation (Lanes C to F), meIF3F Mut and meIF3f (188–361) repressed it in a dose-dependent manner, respectively, reaching 35% and 54% of reduction (Lanes G–J and K–N, respectively). Thus, meIF3f Mut and meIF3f (188–361) inhibit Notch transcriptional activity and behave as dominant-negative forms of eIF3f, in accordance with the fact that both lack an intact catalytic site but are able to bind DTX and Notch. We also tested the effect of shRNAs targeting endogenous eIF3f in this assay. As represented in Figure 7B, Dll1-induced Notch stimulation was inhibited in the presence of increasing doses of either P2 (Lanes C–E) or shRNA #1 (Lanes F–H), although Notch1 overall level remained constant (see bottom of Figure 7B). In contrast, a shRNA pool targeting AMSH, another DUB of the JAMM family, had no effect on Notch activation (Lanes I–K). Furthermore, the same shRNAs have no effect on NIC-mediated transcriptional activation (Figure S5), excluding any effect of the shRNAs on Notch-associated transcription factors. Therefore, inhibition of eIF3f DUB activity impairs the production of a transcriptionally active NIC in a dose-dependent manner.
All together, our results indicate that eIF3f, after being recruited to vesicular Notch via DTX, could act as a DUB on a monoubiquitinated but not γ-secretase processed form of Notch, thus positively modulating Notch signaling activity.
With the aim of identifying a DUB that could regulate Notch activation, we reached two new and provocative conclusions: the activated form of Notch needs to be deubiquitinated to enter the nucleus and fulfill its transcriptional function, and the DUB accounting for this activity is eIF3f, known so far as a translation initiation factor. When reconstituting a Notch activation system by co-culturing Notch1 receptor- and Dll1 ligand-expressing cells, we observed that Notch-dependent transcriptional activation of a reporter gene was specifically affected when expressing mutant forms of eIF3f where the active site was mutated or when eIF3f was partially knocked down. Therefore eIF3f acts as a positive regulator of Notch signaling, independently of its function in translation (see below). We have also demonstrated that a monoubiquitinated form of Notch ΔE was stabilized when eIF3f was slightly knocked down. This effect was accompanied by the appearance of a mono-ubiquitinated form of NIC, which is probably excluded from the nucleus. Therefore we conclude that Notch deubiquitination is necessary for its full activity in the nucleus.
We have not formally identified the form of Notch that is the substrate of the DUB activity. We cannot exclude that it is a monoubiquitinated non-nuclear NIC, resulting from γ-secretase cleavage of activated Notch. Such a form has never been detected before, in contrast to the nuclear polyubiquitinated NIC, which appears subsequently under the action of the E3 ubiquitin ligase Sel10 [32],[33]. An artificial construct mimicking to a certain extent the monoubiquitinated NIC (UBIC) was partially retained in the cytoplasm and as a consequence was transcriptionally less active than NIC. This suggests that the ubiquitin moiety could mask the proximal NLS [34] and impair interaction with importins. Nevertheless three observations argue against the possibility of NIC being the natural substrate of eIF3f: first, NIC was not obviously associated with endosomes when it was excluded from the nucleus in the presence of eIF3f shRNA. As eIF3f colocalizes with Notch to endosomal structures, its target should be retained to these structures (Figure 4). Second, NIC produced from ΔE was not detected as coimmunoprecipitating neither with eIF3f nor with its mutant in the presence of DTX, which would have probably been the case if it were the DUB substrate (Figure 6). And third, the effect of shRNAs targeting eIF3f were only detected on activated but still membrane-anchored forms of Notch, including one unable to produce NIC (ΔE-LLFF, see Figure 2). Therefore we favor the hypothesis that the eIF3f substrate is the activated, monoubiquitinated but still membrane-anchored Notch (mimicked by ΔE in transfection experiments, see [8]). The presence of monoubiquitinated NIC when eIF3f activity is impaired would be due to a partial activity of γ-secretase on non-deubiquitinated ΔE.
Finally, our results show that eIF3f is recruited to Deltex1 and Notch ΔE-containing vesicles in transfected cells. They can thus form a tripartite complex where DTX serves as a bridging factor between Notch and eIF3f. Deltex1 belongs to the RING family of E3 ubiquitin ligase, however its target in Notch signaling remains to be determined [35]. It localizes to the endocytic pathway [27], interacts with Notch, and thus could serve as a scaffolding protein enabling Notch trafficking and modifications, and eventually Notch signaling.
Our results suggest that eIF3f, known so far as a component of a translation initiation factor complex, is itself able to act as a DUB during Notch activation. In contrast to a mutant of the active site, a WT form of eIF3f is able to complement the inhibition of activity mediated by eIF3f shRNAs and to restore deubiquitination of Notch ΔE. However we cannot completely rule out the possibility that another DUB, associated with eIF3f, could account for Notch deubiquitination. This is, for example, the case for CSN5, which, in addition to its own isopeptidase activity, is associated with USP15, both being required for proper processing of polyubiquitinated substrates bound to p97/VCP [36]–[38]. Such a putative eIF3f-associated DUB, the knockdown of which would not affect protein translation, should have been identified during the screen. Moreover, eIF3h, the other MPN-containing subunit of eIF3 associated with eIF3f [12], has neither histidines nor acidic residues that could confer a DUB activity. The fact that eIF3f was the only DUB identified argues against the possibility of an associated DUB. eIF3f was not identified in other screens, in particular in one recently performed in Drosophila [39], probably because a strong extinction of this translation factor had a more severe and broad phenotype than a Notch phenotype, even on external sensory organ development. We actually were able to pick up this factor thanks to the presence of relatively inefficient shRNAs in the library.
The human genome encodes 14 JAMM proteins, seven of which have a complete set of the conserved residues for Zn2+ coordination [40]. Among them, six (AMSH, AMSH-LP, BRCC36, Rpn11, MYSM1, and CSN5) have been reported to have isopeptidase activity on ubiquitin or ubiquitin-like proteins. It is of note that the JAMM sequence of eIF3f cannot be aligned with those of the proteins constituting the MPN+ group (Rpn11, Csn5, AMSH; see [15]). Nevertheless the four polar and the additional glutamate residue in a more N-terminal region of the domain are still present in eIF3f, although not arranged in the MPN+-defined pattern, and the mutation of some of these amino acids affects eIF3f enzymatic activity and function in Notch signaling. It must be noted that the MPN+ consensus domain has been generated using a small number of proteins and that divergent members may well exist, including eIF3f.
Using an in bacteria assay, we have demonstrated that full-length eIF3f or its isolated MPN domain exhibit DUB activity and that the active site indeed involves the amino acids that we had targeted in our inactive mutant. The relatively low activity that we detected in this assay, as compared to the BPLF1 control, might be due to the specificity of this DUB, which is limited by the nature of the P'1 and P'2 residues in the substrate sequence, thus preventing cleavage of non-specific substrates. However we have confirmed by Ub-VS fixation that WT but not mutant eIF3f exhibits DUB activity in mammalian cells. Among the members of the JAMM family, AMSH and AMSH-like seem to be specific for Lys63-linked polyubiquitin chains [41], while MYSM1 is specific for monoubiquitinated H2A [42]. Sato et al. [43] suggested that the specificity of AMSH family members for Lys63-linked polyubiquitin chains is primarily due to their interaction with the proximal ubiquitin, involving a single domain containing two characteristic insertions that are not conserved in eIF3f. However, DUBs may have a multidomain structure, and some of them associate with other proteins, including E3 ubiquitin ligases [23],[44]. We show that Notch is only able to interact with eIF3f in the presence of the E3 ubiquitin ligase DTX and that a monoubiquitinated form of Notch is probably the substrate targeted by the DUB. Therefore any reconstituted in vitro system will be difficult to set up.
As JAMM DUBs are commonly found in association with large protein complexes, on one hand, and eIF3f belongs to the eIF3 complex, on the other hand, it will be of great interest to determine whether eIF3f can work as a DUB outside of the translation complex or whether the active form of eIF3f is associated with the whole translation initiation complex. If it were the case, beside an analogy and homology in the molecular architecture of the other Zomes members (Cop9 signalosome and proteasome, [45]), eIF3 would also harbor an enzymatically-conserved organization.
Beyond acting as a translation initiation factor, eIF3f fulfills other functions, probably independently of the eIF3 complex. For instance, it was shown to inhibit HIV-1 replication [46] by modulating the sequence-specific recognition of the HIV-1 pre-mRNA by the splice factor 9G8. It was also recently proposed to serve as a scaffold in coordinating mTor and SGK1 actions on skeletal muscle growth [47] and to interact genetically and physically with TRC8, an E3 ubiquitin ligase of the RING family with several TM domains [48]. On the other hand, there are at least two different types of eIF3 complexes in the cell, localized in different subcellular fractions. One complex lacks eIF3a and eIF3f, while the other consists of eIF3a–c and eIF3f. Phosphorylated eIF3f may predominantly localize to the nucleus and join a complex containing at least b and c during apoptosis. Therefore, the nuclear eIF3 complex is likely to have functions other than translation initiation [49]. In yeast, affinity purification and LC-MS/MS was employed to characterize the eIF3 interactome, which was found to contain 230 proteins [50]. This led to the proposal that eIF3 assembles into a large supercomplex, the translasome, which contains elongation factors, tRNA synthases, 40S and 60S ribosomal proteins, chaperones, and the proteasome. On the other hand, eIF3 also associates with importins-β, a critical event for normal cell growth. These data suggest that translasomes are dynamically localized within the cell, and eIF3 could shuttle between the cytoplasm and the nucleus in a cell cycle-dependent manner. All these data are in agreement with the hypothesis that eIF3f, and maybe part of eIF3, could have multiple functions in the cells. Our results now add an enzymatic activity to these various properties; it remains to be elucidated whether this activity is necessary to fulfill all eIF3f functions.
Recent data suggest that other translation factors, many of which exist in several copies in the genome (as does eIF3f), have “part-time jobs” outside of their usual function in translation [51]. It is the case, for example, for eIF4A, which exhibits an RNA helicase activity in translation initiation and which was shown to act in a translation-independent manner as a negative regulator of Dpp/BMP signaling in drosophila [52]. eIF4A affects Mad protein level and was suggested to act as an E3 ubiquitin ligase.
In parallel also to ribosomal proteins, which have been recently shown to function outside of the ribosome and to be recruited in some cases for a function unrelated to the ribosome or its synthesis [53], we propose that eIF3f would be another example of a housekeeping protein that also plays a specific role in the Notch signaling pathway.
shDUB library was described in [10] (see also [40],[54],[55]). Additive shRNAs are listed in Table 1. The pool targeting AMSH used as a control in Figure 7B is the number 139 of the library.
All myc-tagged Notch constructs (ΔE, ΔE-LLFF, FL, and NIC) were already described [18],[56] and were gifts from R. Kopan (Washington University, St. Louis, MO). These Notch constructs are all deleted from aa 2183 of murine Notch1 and fused to a hexameric myc tag at the carboxy terminus. Notch1 retroviral vector encodes a full-length human Notch1 with an HA-epitope tag inserted between EGF repeats 22 and 23 (gift of J. Aster, Harvard medical school, Boston, USA).
All WT and deletion mutants of eIF3f (containing N-terminal tags), as well as GST-fused eIF3f constructs, were gifts from S. Leibovitch (Montpellier, France), except the catalytic mutant. meIF3f Mut was generated from HA-tagged WT meIF3f by two successive site-directed mutagenesis using, respectively, the oligonucleotides 5′-GCCACAGGCGCTGCAGCCACAGAACACTCAGTGCTG-3′ and its complementary DNA, and 5′-GACATCACAGCTGCAGCCCTGCTGATCCATGAG-3′ and its complementary DNA (HDI(aa181–183) mutated in AAA and IHE(aa190–192) mutated in GIL). 6xHis-tagged Ubiquitin construct was a gift from M. Treier (European Molecular Biology Laboratory, Heidelberg, Germany).
HIS-tagged eIF3f MPN constructs were built by inserting the amino acids 95 to 226 of meIF3f WT or meIF3f Mut in pet28a vector using NdeI and BamHI restriction sites.
Retroviral meIF3f vectors were constructed first by adding a S-Tag in the N-terminus end of meIF3f WT or meIF3f mutated in the catalytic domain (HDI (aa181–183) mutated in AAA). Then S-tagged meIF3f constructions were inserted in pMSCVpuro vector using HindIII and ClaI restriction sites. The puromycin resistance gene was replaced by S-Tagged meIF3f, whose expression is by the way under the control of PGK promoter.
CSL-Luc was a gift from T. Honjo (Kyoto University, Japan) and is referred to as pGa981-6 in [30]. NDFIP2 was a gift from S. Kumar (Adelaide University, Australia) and is referred to in [19]. Both BPLF1 constructs were described in [24].
We used anti-myc 9E10 and anti-VSV P5D4 antibodies. Anti-Notch IC and anti-Dll1 antibodies were described, respectively, in [57] and [31].
Other antibodies for Western blot were supplied by Abcam (polyclonal anti-S-Tag), Bethyl (polyclonal anti-eIF3a), BioLegend (polyclonal anti-eIF3f), Cell Signaling (anti-Notch V1744), Covance (monoclonal anti-HA; polyclonal anti-HA), Invitrogen (polyclonal anti-GFP), Novagen (monoclonal anti-S-Tag), and Sigma (monoclonal anti-Flag M2; polyclonal anti-Flag; monoclonal anti-β-tubulin; monoclonal anti-α-tubulin, monoclonal anti-β-Actin). Secondary antibodies for immunofluorescence were supplied by Molecular Probes (Alexa Fluor conjugates).
U20S-FL cell line was established by retroviral transduction. High titers of recombinant HA-tagged Notch FL viruses were obtained 48 h after transfection of the Plat-E ecotropic packaging cell line with retroviral expression plasmids. After retroviral transduction of the U2OS cell line, clonal populations were obtained by limiting dilution. OP9-Dll1 cell line was described in [31].
MEFs stably expressing VSV-DTX and S-Tag-meIF3f were established from the DTX-expressing MEFs [27] by retrotransduction of S-tagged meIF3f vectors.
U2OS cells were grown on glass coverslips and were transiently transfected using FuGeneHD transfection reagent (Roche, Mannheim, Germany) for 24 h. Cells were fixed with 4% paraformaldehyde and permeabilized with PBS containing 0.2% Triton X-100 for 5 min before the incubation with appropriate antibodies. Cell preparations were mounted in Mowiol (Calbiochem, Merck Biosciences, Darmstadt, Germany) and images acquired using an AxioImager microscope with ApoTome system with a 63× magnification and AxioVision software (Carl Zeiss MicroImaging Inc., Le Pecq, France).
293T cells were harvested 24 h after transfection and lysed in 8 M urea, 0.1 M NaH2PO4, 10 mM Tris-Hcl (pH 8), 1% Triton X-100, and 20 mM Imidazole at room temperature. His-Ub conjugated proteins were purified on chelating Sepharose beads (Pharmacia), previously charged with Nickel. Ni-bound proteins were washed extensively with the same buffer, then with a pH 6.3 buffer, and eluted in Laemmli before Western blot analysis.
293T cells were collected 24 h after transfection, washed in PBS buffer, and lysed in 50 mM Tris-HCl (pH 7,9), 400 mM NaCl, 5 mM MgCl2, 1% Triton X-100, supplemented with protease inhibitor cocktail (Roche). Cell extracts were cleared by centrifugation at 14,000 rpm for 20 min at 4°C. Immunoprecipitations were performed with the appropriate antibodies in the same buffer. When indicated, the immunoprecipitates were eluted by peptide competition (2 mg/mL) for 1 h at 4°C. Samples were denatured in Laemmli buffer for SDS-PAGE resolution, and immunoblots were performed as described previously [18].
In bacteria functional assay: Bl21 bacteria were transfected with GST-fusions or His-tagged encoding vectors together with Ub-GFP plasmid. The bacteria were selected on agar plates and cultured in LB medium containing Chloramphenicol and Ampicillin or Kanamycin. Exponential bacteria cultures (OD = 0.4) were treated with IPTG 0.5 mM for 16 h at 23°C. Bacteria were harvested, washed, resuspended in PBS containing 20 mM N-Ethylmaleimide, and sonicated for 30 s on ice. After centrifugation (12,000 rpm, 10 min), clear supernatants were measured for GFP fluorescence and protein concentration. Samples were analyzed by two SDS-PAGE resolutions using for each sample 100 units of fluorescence for Ub-GFP assay and 10 µg extracts for protein expression analysis.
Ub-VS assay: HeLa cells were collected 36 h after transfection, washed with PBS, and lysed in 50 mM Tris-Cl pH 7.4, 150 mM NaCl, 1 mM DTT, 1 mM EDTA, 1 mM PMSF, and 0.5% NP40. Protein concentrations were measured. 10 µg of WCE were incubated for 1 h at 37°C with 1 µg of Ub-VS functional probe (Boston Biochem) in a 30 µL final volume of labeling buffer (50 mM Tris-Cl pH 7.4, 250 mM sucrose, 5 mM MgCl2, and 1 mM DTT). Finally, WCE and Ub-VS-treated samples were analyzed by Western blot.
20,000 U2OS-FL cells/cm2 were grown and transiently transfected using FuGeneHD transfection reagent (Roche, Mannheim, Germany). 24 h after transfection, U2OS-FL cells were cocultured with 35,000 OP9-Dll1 cells/cm2. 18 h later, cocultures (done in triplicates) were lysed using Passive Lysis buffer (Promega). A fraction of cell lysates were transferred to a white 96-well plate (Berthold). Firefly and Renilla luciferase activities were measured using the luminometer Centro XS (Berthold). For Western blot analysis, part of cocultures were lysed with 8 M urea, 0.1 M NaH2PO4, 10 mM Tris-Hcl (pH 8), 1% Triton X-100, and 20 mM Imidazole.
|
10.1371/journal.pntd.0001892 | Impact of Wolbachia on Infection with Chikungunya and Yellow Fever Viruses in the Mosquito Vector Aedes aegypti | Incidence of disease due to dengue (DENV), chikungunya (CHIKV) and yellow fever (YFV) viruses is increasing in many parts of the world. The viruses are primarily transmitted by Aedes aegypti, a highly domesticated mosquito species that is notoriously difficult to control. When transinfected into Ae. aegypti, the intracellular bacterium Wolbachia has recently been shown to inhibit replication of DENVs, CHIKV, malaria parasites and filarial nematodes, providing a potentially powerful biocontrol strategy for human pathogens. Because the extent of pathogen reduction can be influenced by the strain of bacterium, we examined whether the wMel strain of Wolbachia influenced CHIKV and YFV infection in Ae. aegypti. Following exposure to viremic blood meals, CHIKV infection and dissemination rates were significantly reduced in mosquitoes with the wMel strain of Wolbachia compared to Wolbachia-uninfected controls. However, similar rates of infection and dissemination were observed in wMel infected and non-infected Ae. aegypti when intrathoracic inoculation was used to deliver virus. YFV infection, dissemination and replication were similar in wMel-infected and control mosquitoes following intrathoracic inoculations. In contrast, mosquitoes with the wMelPop strain of Wolbachia showed at least a 104 times reduction in YFV RNA copies compared to controls. The extent of reduction in virus infection depended on Wolbachia strain, titer and strain of the virus, and mode of exposure. Although originally proposed for dengue biocontrol, our results indicate a Wolbachia-based strategy also holds considerable promise for YFV and CHIKV suppression.
| Mosquito-transmitted viruses such as dengue, yellow fever and chikungunya, are responsible for significant morbidity and mortality throughout tropical and sub-tropical regions of the world. These viruses are primarily transmitted by Aedes aegypti, a mosquito that due to its close association with humans has historically been difficult to control. An innovative control strategy involving the release of mosquitoes infected with the intracellular bacterium Wolbachia is currently being developed. This approach is based on the recent discovery that Wolbachia reduces infection of mosquitoes with dengue virus, malaria parasites and filarial nematodes. In the current study, we demonstrated that Wolbachia also blocks infection of chikungunya and yellow fever viruses in Ae. aegypti. The degree of virus inhibition depended on the strain of Wolbachia, the route of virus exposure, the virus strain and the titer of virus that the mosquitoes were exposed to. The implementation of Wolbachia-based control strategies has the capacity to transform the way that mosquitotransmitted diseases are controlled in the future.
| Mosquito-transmitted viruses cause significant human morbidity and mortality throughout the world and impose particularly heavy health and economic burdens on developing countries. Dengue, caused by infection with any of the four dengue virus (DENV) serotypes, is currently the leading arboviral disease, with millions of cases of classic dengue fever and tens of thousands of deaths annually due to hemorrhagic disease [1]. Yellow fever virus (YFV) has been implicated in an estimated 200,000 clinical cases and 30,000 human deaths annually in the equatorial regions of Africa and South America [2], [3]. Recently, chikungunya virus (CHIKV) emerged as a major threat, with unprecedented outbreaks on islands in the western Indian Ocean, as well as in India, Thailand, Malaysia and Italy [4]. Effective vaccines against all four DENV serotypes and CHIKV are still at various stages of development and clinical trial [5], [6]. Although a highly effective vaccine against YFV has been administered for over 50 years, rapid vaccination of susceptible populations either prior to or during an epidemic is financially and logistically challenging, particularly in developing countries [2], [3].
Current disease control measures focus on the suppression of mosquito vector populations to reduce virus transmission. The primary vector of DENVs, CHIKV and YFV is the mosquito Aedes aegypti, a highly domesticated species that feeds almost exclusively on humans. Its geographic range has expanded with increased urbanization, resulting in increased arbovirus transmission [7]. The primary mosquito control activities are source reduction to eliminate larval habitats, application of larvicides (such as Temephos or s-methoprene), or adulticiding with indoor residual spraying or ultra-low-volume application. While these approaches can be successful, they are often labor-intensive, can be prohibitively expensive to implement and require a sustained commitment from all levels of government [8]. Furthermore, increasing insecticide resistance and concerns with non-target effects on the environment have necessitated the development of alternative approaches to arbovirus control.
An emerging biocontrol approach to reducing transmission of arboviruses is provided by the transinfection of Ae. aegypti mosquitoes with Wolbachia pipientis from other insect hosts [9], [10]. Wolbachia is a maternally transmitted, endosymbiotic bacterium that manipulates host reproduction to enhance its own transmission [11]. Estimated to infect over 60% of insect species [12], Wolbachia can provide its host with nutritional benefits [13] and enhanced resistance to pathogens [14], [15]. Ae. aegypti transinfected with the wMelPop-CLA strain of Wolbachia from Drosophila melanogaster [16] displayed a shortened life span [16], a reduction in blood feeding success [17], [18] and dramatically lowered DENV serotype 2 (DENV-2) infection levels compared to Wolbachia-free control mosquitoes [19]. Although these phenotypes are likely to reduce virus transmission in the field, wMelPop-CLA may also impose fitness costs on Ae. aegypti, such as reduced fecundity due to poor blood feeding success [17], [18] and decreased embryonic viability [20].
A second strain of Wolbachia, wMel that is closely related to wMelPop-CLA and also occurs naturally in D. melanogaster was recently introduced into Ae. aegypti as an additional strain for the biocontrol of dengue [21]. Both strains induce cytoplasmic incompatibility and have high rates of maternal transmission [16], [21], phenotypes necessary for the invasion of Wolbachia in natural populations of mosquitoes. Unlike wMelPop-CLA infected mosquitoes, Ae. aegypti infected with wMel did not suffer any significant deleterious fitness costs when compared to uninfected controls [21]. Similar to wMelPop-infected Ae. aegypti, wMel-infected mosquitoes displayed dramatically reduced replication of DENV-2 [21]. Ae. aegypti infected with wMel were deployed in the field in north Queensland, Australia over the 2010–2011 summer [22]. Wolbachia was able to reach almost 100% fixation in wild mosquito populations only a few months following release, indicating that Wolbachia-infected mosquitoes present a very promising strategy for the control of dengue that is cost-effective and poses minimal environmental and social harm [23], [24].
Although originally developed as a biocontrol tool for DENVs, Ae. aegypti infected with wMelPop-CLA showed reduced infection with CHIKV [19], filarial nematodes [25] and avian malaria [19]. Therefore, wMelPop-CLA infected mosquitoes could potentially be used for biocontrol in areas where human pathogens other than DENVs occur. However, not all strains of Wolbachia protect equally well, with those phylogenetically most closely related to wMel and wMelPop conferring the greatest degree of protection [26]. In Drosophila, wMel confers protection against a range of viruses, for example Drosophila C virus, Flock House virus and Cricket paralysis virus [14], [15]. However, it remains unclear whether wMel infection is able to limit the replication of human pathogens other than DENVs in mosquitoes, information that is critical to evaluating different Wolbachia strains for biocontrol.
Here, we tested the ability of the wMel strain of Wolbachia to limit infection in Ae. aegypti with CHIKV and YFV. We also tested the ability of YFV to replicate in mosquitoes infected with wMelPop-CLA, in order to compare this strain to wMel. We found reduced levels of CHIKV but not YFV infection in wMel-infected Ae. aegypti, although the degree of virus inhibition depended on the mode of infection. By contrast, mosquitoes harboring the wMelPop-CLA strain of Wolbachia showed reduced infection rates with YFV with the extent of reduction virus strain and titer dependent.
Six different lines of Ae. aegypti were used in the experiments. The transinfection of Ae. aegypti with the wMelPop-CLA and wMel strains of Wolbachia and maintenance of infected lines has been previously described 16,21. The wMel-infected line, MGYP2, and its tetracycline-treated counterpart, MGYP2.tet, were exposed to both YFV and CHIKV. Both lines were in the F16 generation for the YFV experiments and the F24 generation for the CHIKV experiments. The original wMelPop-CLA infected line, PGYP1 [16], and the corresponding tetracycline-treated line, PGYP1.tet, were assessed in the YFV experiments only. The PGYP1 line was used in the combined F52 and F53 generations, while the PGYP1.tet was used in the combined F50 and F52 generations. A Wolbachia-uninfected wild-type line of Ae. aegypti, designated Cairns3, which originated from Cairns, Australia, was used as a positive control for all experiments. The YFV-susceptible Rex-D white-eye Higgs strain of Ae. aegypti that originated from Rexville, Puerto Rico, was used as an additional positive control for the YFV experiments [27], [28]. Both the Cairns3 and Rex-D lines had been in colony for >40 generations.
The CHIKV strain was isolated from a patient visiting Melbourne, Australia in 2006 and contained the alanine to valine mutation in the membrane fusion glycoprotein E1 gene (E1-A226V) that has been linked to increased infectivity in mosquitoes, especially Ae. albopictus [29], [30]. The CHIKV stock had previously been passaged three times in African green monkey kidney (Vero) cells prior to use in this study. Two YFV strains that have been characterized in mosquitoes were used for the experiments: BA-55 which was isolated from a fatal yellow fever case in Nigeria in 1987 [31] and Cinetrop 28 (OBS 7549), which was isolated from a yellow fever patient in Bolivia in 1999 (R.B. Tesh, University of Texas Medical Branch, personal communication). The BA-55 strain had been passaged three times in suckling mouse brains, whilst the Cinetrop 28 strain had been passaged twice in C6/36 (Ae. albopictus) cells.
The CHIKV experiments were undertaken in the Biosafety Level 3 (BSL-3) insectary at Queensland Health Forensic and Scientific Services, Brisbane, Australia, and the experiments with YFV were undertaken in the BSL-3 insectary located at the University of Texas Medical Branch, Galveston, Texas, USA. The MGYP2, MGYP2.tet and Cairns3 lines were exposed to CHIKV using both intrathoracic inoculation and oral feeding. Intrathoracic inoculation was employed because it circumvents the midgut infection and escape barriers, and allows a standard amount of virus to be delivered. The six Ae. aegypti lines were exposed to YFV via intrathoracic inoculation and in an infectious blood meal. However, low feeding rates in all lines coupled with unexpectedly low infection rates in control lines that did actually feed, compromised our ability to draw any meaningful conclusions from the YFV oral feeding data. Thus it was excluded from the current paper. For intrathoracic inoculation, immobilized mosquitoes were inoculated with 0.5 µl and 0.22 µl of YFV or CHIKV suspension, respectively. The suspension consisted of stock virus diluted in growth media (GM; Gibco-BRL, Gaithersburg, MD) which contained antibiotics and antimycotics and was supplemented with 10% fetal bovine serum (FBS) for YFV or 3% FBS for CHIKV. For oral feeding, mosquitoes were allowed to feed on a blood meal consisting of CHIKV diluted in commercially available defibrinated sheep blood (Institute of Medical and Veterinary Science, Adelaide, Australia) and 1% sucrose. The blood meal was housed within a membrane feeding apparatus that was fitted with pig intestine as the membrane [32]. Feeding was undertaken at 23°C and a 2 h period was used to ensure that a sufficient number of mosquitoes had imbibed a blood meal. Immediately following virus exposure, mosquitoes were anaesthetized with CO2 and engorged mosquitoes placed in 900 ml gauze-covered containers. All mosquitoes were maintained on 10% sucrose at 28°C, high relative humidity and 12L∶12D light cycle within an environmental growth cabinet.
For the YFV experiments, mosquitoes were processed at either day 10 or 14 post-exposure. Mosquitoes were chilled and the heads removed from the bodies and placed separately in 0.5 ml of GM supplemented with 10% FBS, and antibiotics and antimycotics. Recovery of virus from the head demonstrated that the virus had infected head tissue and was potentially able to be transmitted.
For the CHIKV experiments, infection, dissemination and transmission were assessed using a modified in vitro capillary tube system [33]. Briefly, mosquitoes were anaesthetized with CO2 and the legs and wings removed. The saliva was collected by inserting the proboscis of the mosquito into a capillary tube containing GM with 20% FBS. After 30 min, the contents of the capillary tube were expelled into 600 µl of GM with 3% FBS. The salivary glands were then dissected from the body in a drop of GM with 3% FBS. The dissected salivary glands were washed in a drop of 1% bleach [34], before being rinsed in 2 drops of GM with 3% FBS. The body remnants and salivary glands were placed in separate 2 ml tubes containing 1 ml of GM with 3% FBS and 3 sterile glass beads.
Salivary glands were also excised for visualization of CHIKV infection via an immunofluorescence assay (IFA). The glands were dissected in a drop of PBS and washed in another two drops of PBS, before each gland was placed in separate wells of an 18 well microscope slide. Slides were air-dried, fixed in cold acetone for 30 min, before being stored at −80°C.
Infection, dissemination and transmission rates were analyzed using Fisher's Exact Tests with two-tailed P-values. YFV copy numbers were tested for differences among mosquito lines using Wilcoxon rank sum tests for non-parametric data and P-values <0.05 were considered statistically significant. P-values were adjusted for multiple tests using a Bonferroni correction. All analyses were performed in R [41].
We challenged wMel-infected (MGYP2 line) and uninfected mosquitoes (the tetracycline-treated MGYP2.tet and wild-type Cairns3 lines) with CHIKV in two replicate oral feeding experiments. In both experiments, fewer MGYP2 mosquitoes displayed CHIKV infection compared to MGYP2.tet and Cairns3 lines (Table 1). In experiment 1, virus was detected in significantly fewer MGYP2 (23%) mosquito bodies compared to MGYP2.tet (83%) and Cairns3 (58%) lines (P<0.01). Similarly, in experiment 2, CHIKV was detected in significantly fewer MGYP2 (30%) bodies compared to MGYP2.tet (70%) and Cairns3 (67%) lines (P<0.01).
Next, we used both CC-EIA and a CHIKV-specific TaqMan RT-PCR assay to track dissemination of the virus to the salivary glands and saliva following oral feeding experiments. CHIKV was detected in significantly fewer MGYP2 (20%) salivary glands compared to those from MGYP2.tet (80%) and Cairns3 (66%) lines (P<0.01) in experiment 1, but only when using the TaqMan RT-PCR assay (Table 1, in bold). The TaqMan RT-PCR assay detected more CHIKV-positive saliva expectorates than the CC-EIA (Table 1), suggesting that, if infectious particles were present in the saliva, they were below the threshold of CC-EIA detection. Nonetheless, in both oral feeding experiments CHIKV was detected in significantly fewer saliva expectorates in MGYP2 (10% and 6%) than MGYP2.tet (58% and 47%) (in both experiments P<0.01), but not Cairns3 mosquitoes.
We tested whether wMel infection reduced CHIKV infection rates in Ae. aegypti using intrathoracic inoculations as the mode of virus delivery, in two separate experiments. Inoculations with CHIKV at titers of 104.5 and 106.3 TCID50/ml resulted in similar rates of virus infection among the three mosquito lines in bodies, salivary glands and saliva expectorates, using both CC-EIA and TaqMan RT-PCR for detection (Table 1, experiments 3 and 4). Rates of infection in bodies and salivary glands were consistently high across all three lines. A smaller percentage of saliva expectorates from MGYP2 mosquitoes were virus infected compared to MGYP2.tet and Cairns3 mosquitoes, but the difference was not statistically significant.
Finally, we visualized CHIKV infection in the salivary glands of MGYP2 and Cairns3 Ae. aegypti lines using immunofluorescence assays (IFA). Twelve days following oral exposure, 27% (3/11) and 67% (6/9) of the MGYP2 and Cairns3 lines, respectively, were positive in the IFA. There did not appear to be any consistency in the localization of virus within infected salivary glands from either Ae. aegypti line (Figure 1).
We tested the effect of wMel and wMelPop-CLA (PGYP1 mosquito line) Wolbachia strains on body infection rates, using intrathoracic inoculation with two strains of YFV isolated from Nigeria (BA-55) and Bolivia (Cinetrop 28). Four Wolbachia-uninfected mosquito lines were used as controls: Cairns3, Rex-D, MGYP2.tet and PGYP1.tet. Due to high mortality rates in some lines, mosquitoes exposed to 103.5 TCID50/ml of YFV BA-55 were tested at 10 days post virus exposure. Mosquitoes exposed to104.5 TCID50/ml of YFV BA-55 and 104.0 TCID50/ml of YFV Cinetrop 28 were tested at 14 days post exposure.
At both low (103.5 TCID50/ml) and high (104.5 TCID50/ml) titers of YFV strain BA-55, MGYP2 mosquitoes had the same high (100%) rates of body infection as MGYP2.tet, Cairns3 and Rex-D mosquitoes (Table 2), detected using a YFV-specific Taqman assay. By contrast, at low titer, significantly fewer PGYP1 (12%) mosquitoes displayed body infections compared to the PGYP1.tet (100%), Cairns3 (100%) and Rex-D (100%) lines (P<0.001). PGYP1 mosquitoes also displayed significantly lower rates of body infection compared to MGYP2 mosquitoes (P<0.001). At the higher virus titer a similar number (100%) of PGYP1 mosquitoes was infected with YFV as for the PGYP1.tet, Cairns3 and Rex-D lines. A similar pattern was observed for the Cinetrop 28 strain of YFV, with mosquito body infection rates ranging from 96–100% and no significant differences among the mosquito lines, although this strain was tested at only one titer (104.0 TCID50/ml).
Next, we assessed the impact of the two strains of Wolbachia on YFV dissemination to the head in Ae. aegypti. Using the YFV strain BA-55, MGYP2 mosquitoes displayed similar high rates of dissemination to the head as the MGYP2.tet, Cairns3 and Rex-D mosquitoes (92–100%), at both low and high virus titers (Table 2). At low virus titer, however, no dissemination to the head was observed in PGYP1 mosquitoes compared to control lines PGYP1.tet (100%), Cairns3 (100%), Rex-D (100%) and the MGYP2 (100%) (P<0.001). At high virus titer, a higher (78%) number of PGYP1 mosquitoes had virus in the heads and no significant differences in infection rates were found across the mosquito lines. High head dissemination rates (84–100%) were found across all mosquito lines challenged with the Cinetrop 28 strain of YFV, with no significant differences among them.
Next, we explored the effect of Wolbachia on the number of YFV RNA copies in Ae. aegypti bodies following intrathoracic inoculation with the strains BA-55 and Cinetrop 28, at three different titers. Overall, infection with both Wolbachia strains resulted in a lower median number of virus copies per body in Ae. aegypti compared to control lines. However, PGYP1 mosquito bodies consistently displayed several log fewer copies than MGYP2 lines (Figure 2). Significantly fewer copies were found in MGYP2 compared to Wolbachia-uninfected MGYP2.tet, Cairns3 and Rex-D mosquitoes (P<0.05 in all Wilcoxon rank sum tests) in challenges with YFV strain BA-55 at a titer of 103.5 TCID50/ml. However, virus levels were high in both MGYP2 and the Wolbachia-uninfected lines (medians >108 copies/body; Figure 2A). By contrast, median virus copy in PGYP1 mosquitoes was zero, an 8-log difference compared to MGYP2 (P<0.0001, Wilcoxon rank-sum test) and significantly less than PGYP1.tet, Cairns3 and Rex-D mosquitoes (P<0.0001 in all Wilcoxon rank-sum tests) (Figure 2A).
In challenges with YFV strain BA-55, at the higher titer of 104.5 TCID50/ml, MGYP2 bodies still had significantly fewer virus copies than Rex-D (P<0.001, Wilcoxon rank-sum test) but not the MGYP2.tet or Cairns3 lines (Figure 2B). However, PGYP1 bodies had far less virus than either the MGYP2 or Wolbachia-uninfected lines (P<0.0001 in all Wilcoxon rank-sum tests) (Figure 2B). YFV replicated to a significantly lesser extent in MGYP2 and PGYP1 bodies following challenge with the Cinetrop 28 strain compared to control lines (P<0.0001 in all Wilcoxon rank-sum tests) (Figure 2C). YFV copy numbers were almost 5-logs lower in PGYP1 than MGYP2 bodies (P<0.0001, Wilcoxon rank-sum test) (Figure 2C).
We also explored whether Wolbachia strain and YFV inoculation titer influenced the extent to which the virus replicated in Ae. aegypti heads, using TaqMan RT-PCR quantification of viral RNA (Figure 3). In challenges with YFV strain BA-55 at the lower titer of 103.5 TCID50/ml, significantly fewer RNA copies were detected in MGYP2 heads compared to Wolbachia-uninfected lines (P<0.001 in all Wilcoxon rank-sum tests), although the number of virus copies was still high (Figure 3A). By contrast, no virus was detected in the heads of PGYP1 mosquitoes. In challenges with YFV strain BA-55 at the higher titer of 104.5 TCID50/ml, similar levels of virus disseminated to the head in MGYP2 as in Cairns3 and Rex-D, but not MGYP2.tet mosquitoes (P<0.01, Wilcoxon rank-sum tests) (Figure 3B). Median virus copy number for PGYP1 heads was six logs lower than the median for MGYP2 lines (P<0.0001, Wilcoxon rank-sum test), although virus was present in the majority of PGYP1 heads (Figure 3B). Significantly less virus had disseminated to MGYP2 and PGYP1 heads compared to control, Wolbachia-uninfected lines (P<0.0001 in all Wilcoxon rank-sum tests) when mosquitoes were challenged with the Cinetrop 28 strain (Figure 3C). Interestingly, a higher amount of YFV was found in PGYP1 heads in challenges with the Cinetrop 28 strain (median = 104.8 copies, Figure 3C) compared to BA-55 at either low (median = 0 copies; Figure 3A) or high (median = 103.1 copies; Figure 3B) inoculation titers, suggesting that the extent of YFV replication in the presence of Wolbachia may depend on virus strain.
Mosquito-transmitted viruses such as DENVs, CHIKV and YFV pose a significant health risk for almost half of the world's population. The release of Wolbachia-infected mosquitoes into natural populations has been proposed as a way of reducing DENV transmission while minimizing social and environmental harm [21], [22]. Although developed for dengue biocontrol, Wolbachia-infected mosquitoes may prevent the transmission of other significant arboviruses. Here, we evaluated whether mosquitoes infected with the wMel strain of Wolbachia show limited infection with CHIKV and YFV. To compare the two Wolbachia strains, we tested whether YFV was able to infect and replicate in wMelPop-CLA infected mosquitoes.
Mosquitoes infected with wMel showed significantly reduced rates of CHIKV infection and dissemination to the salivary glands compared to controls, but only in the oral exposure experiments. CHIKV also showed limited dissemination in wMelPop-CLA-infected mosquitoes following oral exposure [19], suggesting that both strains of Wolbachia may be useful candidates for release in CHIKV control programs. By contrast, YFV was much less likely to infect and disseminate in Ae. aegypti infected with wMelPop-CLA compared to wMel strains. The virus was also less likely to replicate in wMelPop-CLA infected mosquitoes, with very high virus loads detected in wMel-infected Ae. aegypti. Our experiments suggest that wMelPop-CLA infected mosquitoes may be the best candidates for YFV biocontrol programs, but were unable to determine the extent of virus replication following oral exposure rather than intrathoracic inoculation. Because virus inhibition with some Wolbachia-virus combinations does not appear to be complete, it is essential that epidemiological models be utilized to establish the threshold of virus inhibition necessary to minimize and prevent transmission in the field.
The wMel strain occurs at high levels in Ae. aegypti ovaries and salivary glands but it is not present in as many body tissues as the wMelPop-CLA strain [19], [21]. In particular, wMel does not accumulate to high levels in the fat bodies, brain and thoracic ganglia, which are involved in secondary replication of arboviruses prior to infection of the salivary glands [42]. YFV may have replicated to a sufficiently high level in these tissues in wMel-infected mosquitoes to allow dissemination to the salivary glands and potentially be transmitted to humans. The presence of wMel in the mosquito did not completely prevent the replication of CHIKV in Ae. aegypti salivary glands, as determined by cell-culture, real-time RT-PCR and IFA analysis.
Host immune pre-activation has been proposed as an explanation for interference with virus replication because Wolbachia-infected Ae. aegypti show significant immune upregulation [19], [25], [43]. Recently, Pan et al. showed that Wolbachia infection activates the Toll immune pathway in Ae. aegypti and the production of antimicrobial peptides which inhibit DENV-2 [44]. However, Wolbachia infection in Drosophila, the endosymbiont's native host, also blocks DENV-2 replication despite the absence of immune upregulation [45]. Therefore, although immune upregulation may explain some of the observed Wolbachia-mediated interference with virus replication, it may not be the whole explanation [45]. Alternatively, Wolbachia may compete with arboviruses for key cellular components [19], [46]. Fluorescence in situ hybridization has revealed that wMelPop-CLA infection is widespread throughout the body of infected Ae. aegypti, and is prevalent in tissues and organs associated with viral replication, including the fat bodies, brain and ommatidia [19]. Importantly, DENV-2 antigen did not co-localize in cells and tissues that were infected with wMelPop-CLA. As DENV-2 and YFV are closely related flaviviruses, it is possible that a similar pattern of virus exclusion is occurring with YFV in wMelPop-CLA infected cells and tissues. In vitro studies have shown that reduced virus replication correlates with the density of Wolbachia in the cell [46]. Further studies are required to determine the precise molecular mechanism underpinning Wolbachia-mediated interference with virus replication.
The discovery that Wolbachia limits arbovirus infection and replication in Ae. aegypti has the potential to fundamentally alter vector control strategies [23]. Importantly, mass release of wMel-infected Ae. aegypti in two communities near Cairns, northern Australia, resulted in >90% frequency within the population six weeks after the first release [22]. Although the Wolbachia-based control strategy was initially proposed to control DENVs, previous work and the current study have demonstrated that releases of Wolbachia-infected Ae. aegypti to control DENVs may have the added benefit of reducing CHIKV and YFV transmission. Alternatively, Wolbachia-infected Ae. aegypti could be released to specifically reduce transmission of these viruses, irrespective of whether DENVs were circulating or not in an area.
We have demonstrated that the level of virus blockage induced by Wolbachia is dependent on the strain of Wolbachia, the strain of virus and the mode of exposure to the virus. Higher YFV intrathoracic inoculation titers allowed more breakthrough of virus in wMelPop-CLA infected mosquitoes. Infection rates, accumulation and dissemination to the head also differed between BA-55 and Cinetrop 28 in the presence of wMelPop-CLA, suggesting that different virus strains may have different replication dynamics in Wolbachia-infected mosquitoes. The CHIKV experiments showed that the TaqMan RT-PCR was generally more sensitive than the CC-EIA in detecting the presence of virus. This was not unexpected, given that the TaqMan RT-PCR does not differentiate between RNA derived from live, packaged infectious virus particles, from that of dead or defective, non-infectious virus. Thus, our experiments illustrate the importance of carefully selecting virus titer, virus strain, mode of delivery and method of virus detection when assessing the potential utility of different Wolbachia strains for biocontrol.
As part of the implementation of Wolbachia-based control strategies, it will be essential to assess the virus blocking ability of different Wolbachia strains against different virus strains, and if necessary, identify additional Wolbachia strains that could be deployed in the field. Commensurate with this will be the need for ongoing monitoring of Wolbachia-infected Ae. aegypti populations to confirm that the virus-limiting phenotype is maintained in successive generations post release.
|
10.1371/journal.pcbi.1005671 | Reconstructing the regulatory circuit of cell fate determination in yeast mating response | Massive technological advances enabled high-throughput measurements of proteomic changes in biological processes. However, retrieving biological insights from large-scale protein dynamics data remains a challenging task. Here we used the mating differentiation in yeast Saccharomyces cerevisiae as a model and developed integrated experimental and computational approaches to analyze the proteomic dynamics during the process of cell fate determination. When exposed to a high dose of mating pheromone, the yeast cell undergoes growth arrest and forms a shmoo-like morphology; however, at intermediate doses, chemotropic elongated growth is initialized. To understand the gene regulatory networks that control this differentiation switch, we employed a high-throughput microfluidic imaging system that allows real-time and simultaneous measurements of cell growth and protein expression. Using kinetic modeling of protein dynamics, we classified the stimulus-dependent changes in protein abundance into two sources: global changes due to physiological alterations and gene-specific changes. A quantitative framework was proposed to decouple gene-specific regulatory modes from the growth-dependent global modulation of protein abundance. Based on the temporal patterns of gene-specific regulation, we established the network architectures underlying distinct cell fates using a reverse engineering method and uncovered the dose-dependent rewiring of gene regulatory network during mating differentiation. Furthermore, our results suggested a potential crosstalk between the pheromone response pathway and the target of rapamycin (TOR)-regulated ribosomal biogenesis pathway, which might underlie a cell differentiation switch in yeast mating response. In summary, our modeling approach addresses the distinct impacts of the global and gene-specific regulation on the control of protein dynamics and provides new insights into the mechanisms of cell fate determination. We anticipate that our integrated experimental and modeling strategies could be widely applicable to other biological systems.
| A systematic characterization of the proteomic changes during the process of cell differentiation is critical for understanding the underlying molecular mechanisms. However, protein expression can be largely affected by changes in cell physiological state, which hampers the detection of regulatory interactions. Here we proposed an integrated experimental and computational framework to reconstruct regulatory circuits in mating differentiation of budding yeast Saccharomyces cerevisiae, in which distinct cell fates are triggered by alteration in pheromone concentration. A modeling approach was developed to decouple gene-specific regulation from growth-dependent global regulation of protein expression, allowing us to reverse engineering the gene regulatory circuits underlying distinct cell fates. Our work highlights the importance of model-based analysis of proteomic data and delivers new insight into dose-dependent differentiation behavior of budding yeast.
| Retrieving gene regulatory networks from experimental measurements lies at the foundation for deciphering the mechanistic basis of cellular responses. To date, several strategies have been proposed to reconstruct biological networks. It is possible to infer network connectivity directly from genome-wide localization analysis, which takes advantages of high-throughput techniques to identify genomic sites bound by transcription factors (TFs) [1–4]. However, uncovering the physical interactions is insufficient to bring insight into the underlying regulatory mechanisms and recapitulate the dynamics of the system. Another strategy makes use of the simultaneous measurements of network elements and requires reverse engineering methods to deduce the network structure. A plethora of algorithms have been proposed to reconstruct gene regulatory networks in different organisms, and their performance has been quantitatively assessed [5–8]. Well-established methods include statistical methods based on correlation and mutual information [9, 10], ordinary differential equation (ODE) model [11], Bayesian networks [12] and Boolean network models [13, 14]. Prior knowledge about the organization of the biological network can be further incorporated into the workflow to facilitate the reverse engineering process [15, 16].
Despite these research achievements, several challenges exist in the reconstruction of biological networks. Gene expression profiles are widely used to retrieve transcriptional regulatory networks [8] with the implicit assumption that the activity of a TF is proportional to its mRNA level. However, the expression level of TFs is also subject to post-transcriptional regulations. Earlier studies employing simultaneous measurements of the transcriptome and proteome showed substantial differences between the mRNA and protein abundance either at the population level or the single-cell level [17–19]. On the other hand, although proteomic data provides a more reliable estimation of gene activity, it is not a good indicator of gene regulatory events. Physiological changes that involve global variations in ribosome number, metabolite concentration and growth rate can also affect protein synthesis and dilution, contributing to a layer of regulation that is independent of gene-gene interactions [18, 20–22]. This is especially important for investigating the gene regulatory networks underlying cell fate switches, in which distinct cell fates are often associated with very different physiological parameters (such as growth rate and biogenesis). Recently, several analyses applied dynamic modeling of protein life cycles to characterize the effect of different factors on the variations in protein abundance [23, 24], and their results indicated the critical role of kinetic modeling in decoupling the influence from different levels of regulation.
In this work, we incorporated high-resolution gene expression and cell growth profiling into kinetic modeling to study the cell fate determination in yeast mating response. The yeast mating response pathway is among the best-characterized models in the study of signal transduction, in which the external signal is transmitted through a mitogen-activated protein kinase (MAPK) cascade. This signal eventually activates transcription factor Ste12, which initiates downstream gene regulatory programs (Fig 1A). Despite a wealth of detailed information revealed by past studies [25], a less-studied perspective of the mating response is the cell fate decision governed by changes in the pheromone concentration [26–28]. While growth arrest and shmoo-like morphology is triggered when cells sense a high concentration of pheromone, yeast cells initiate a chemotrophic elongated growth along the pheromone gradient in response to a lower dose of pheromone. Due to the complexity of gene expression programs induced by pheromone stimulation, the mechanism underlying the mating differentiation switch remains elusive. Therefore, it represents a unique opportunity for quantitative investigations into whether and how divergent gene expression networks leading to distinct cell fates can be stimulated in a dose-dependent way.
Through kinetic modeling of protein abundance, we found that the observed protein dynamics in the mating response were partially determined by changes in the physiological state of the cell. Therefore, a model-driven framework named protein synthesis decoupling analysis (PSDA) was developed to decouple gene-specific regulation information from the influence of global physiological regulation. Based on the temporal order of gene-specific changes from PSDA, the putative regulatory networks were then reconstructed using a Boolean network model [14, 29]. These model analysis revealed network rewiring during cell differentiation and suggested a pheromone-dependent regulation of the TOR signaling pathway [30]. In summary, our results highlight the importance of considering the global physiological effects on gene expression control and provide mechanistic insights into the cell fate decisions triggered by different doses of pheromone.
To quantify the effect of physiological constraints on protein dynamic changes and reconstruct the gene regulatory networks in the yeast mating response, we developed a high-throughput microfluidic device that features a throughput of 96 experiments in one single run, continuous control of the medium and an automatic image processing pipeline. The system allows for simultaneous measurements of cell mass accumulation and protein expression level (Fig 1B). We used our platform to track the expression of ~200 fluorescently tagged genes as well as the growth dynamics of yeast cells in response to high and intermediate levels of pheromone (0.59 μM and 5.9 μM). These data offered a comprehensive view of the downstream gene regulatory response with unprecedented temporal resolution. In our experiment, the yeast strains from a green fluorescent protein (GFP)-tagged library [31] with BAR1+ background were cultured and loaded into a 96-well microfluidic device (S1A Fig), in which each strain was confined within the observation region of an insulated chamber for time-resolved analysis. The chemical condition of the chamber was controlled in an accurate and continuous way, and the concentration of pheromone was set to a high or intermediate level about 1 h after cell loading. In each experiment, phase contrast and fluorescence images of cells under pheromone stimulation were acquired at an interval of 5-min for 6 hours, producing >20000 images with single-cell resolution from a single chip. An image processing pipeline was adopted to automatically track the fluorescence intensity of each cell and estimate the growth rate via quantification of the accumulation of cell mass. To study the underlying mechanism of mating differentiation switching, we focused on a set of 195 selected genes including 79 Ste12 regulated genes [3], 52 transcriptional factors that are critical in the regulation of protein expression, and 64 manually selected genes that are representative of different functional groups relating to mating response.
We monitored the protein expression and growth dynamics of the selected yeast strains from the GFP library in response to two different concentrations of pheromone (first two columns of Fig 2A and 2B, S1 Dataset). Consistent with previous studies [26, 27], different cell fate responses, characterized by distinct morphological changes were induced depending on pheromone levels (S1B Fig). When exposed to a high concentration of pheromone, cells underwent a sudden cell cycle arrest within 60 min and formed multiple short projections. Intriguingly, the protein abundances of most examined genes were first up-regulated and then relaxed to their original levels towards the end of the experiment (Fig 2A, left column). These results exhibit a substantial deviation from the transcriptional regulation reported before, because previous microarray data showed that a reduction in the transcript levels of at least 48% of examined genes under the same condition [32]. In addition, we observed the growth rate of shmooing cells underwent a sudden arrest, falling to about half of its normal value within 60 min (Fig 2A, middle column). Under an intermediate level of pheromone stimulation, yeast cells arrested in cell cycle synchronously and initiated a chemotrophic elongated growth. Accordingly we observed a gradual slowdown of growth rate, in accord with the linear mode of cell mass accumulation in elongated cells (Fig 2B, middle column). The protein abundances of most examined genes were also up-regulated, but to a lesser extent and were more fluctuating over time (Fig 2B, left column). These results suggest that cell cycle progression and cell growth may influence the protein dynamics, in accordance with previous studies [33].
To calculate the actual protein synthesis rate from our data, we next employed a mass-action kinetic model of protein abundance and investigated the interdependence between the measured fluorescence and the growth dynamics. In this model, the changes in protein concentration were considered to be due to protein synthesis and decay; the latter was attributed to protein degradation and cell growth dilution (see Methods). To calculate the synthesis rate for each protein, we took into account the expression P(t) and growth profiles γ(t) generated from the time-resolved measurements (first two columns of Fig 2A and 2B) and the protein degradation rate d obtained from a genome-wide analysis of protein half-lives [34]. The kinetic model is solved in a discrete manner so that the protein concentration change between two time points can be expressed as ΔP(t) = Δtα(t) − Δt(d + γ(t))P(t). The calculated results of protein synthesis rate α(t) are presented in the third columns of Fig 2A and 2B. Under the high pheromone level, the synthesis rates of most examined proteins were slightly up-regulated followed by a significant decrease to a much lower level. Because the average mRNA level of examined genes is not significantly reduced throughout the same time period [32], our results suggest that the genome-wide regulation of translation rate is responsible for the global decrease in protein synthesis rate. This global reduction in translation rate may compensate the decrease in protein dilution rate caused by cell cycle arrest and lead to an overall decline of protein abundance after the initial induction (Fig 2A, right column). In contrast, in cells responding to an intermediate level of pheromone, the protein synthesis rate was more fluctuating with a general trend of a more prolonged increase (Fig 2B, right column). These results showed that distinct cell fates are associated with different dynamic changes in protein synthesis rate, as protein biogenesis might be strongly inhibited in the shmooing cells but not in the elongated growing cells.
Since gene-gene interactions and variations in physiological parameters can affect protein dynamics at the same time, we investigated their impacts on protein expression control individually through kinetic modeling. We first examined to what extent the global physiological regulation inherent to each cell strain accounted for the dynamic changes in protein expression. We used a similar kinetic model as that used in the estimation of protein synthesis rate. The only difference is that the synthesis rate of each protein was replaced by a rescaled control rate S(t) generated as follows: (1) we estimated the dynamic changes of global protein synthesis rate for two phenotypes, which revealed the physiological constraints on protein expression control in the pheromone response (S3 Fig); (2) the global synthesis rate of each protein was generated by rescaling the normalized control rate to model the differences in their basal expressions. The rescaling factor is proportional to the steady state protein abundance prior to pheromone stimulation. The protein dilution and degradation rates were set to the per-strain values. Therefore, by estimating changes in protein synthesis and dilution that are independent of gene-specific interactions, we were able to obtain the protein dynamic changes caused by global regulation. We found the simulated protein abundance profiles showed a trend of transient up-regulation even though there is no gene-gene interactions (Fig 3A, S3 Fig). The Pearson correlation between the predicted and observed fold change was 0.37 (p-value = 8.5 x 10−6), indicating that about 14% of the variance in the protein level can be explained by the influence of physiological factors. Thus our results suggest that although physiological regulation is responsible for the global trend of protein dynamics, the protein dynamics of different genes are mainly determined by gene-specific regulations.
We assume that the discrepancy between the measured expression and the above-mentioned calculation stems from gene-specific regulation. To quantitatively dissect the gene-specific transcriptional regulation, we developed a computational framework named Protein Synthesis Decoupling Analysis (PSDA) that employs a robust and cross-grain estimation of the time window and fold change of gene regulation events (Fig 3A). We used a modified kinetic model of protein life cycle, in which protein synthesis is the product of the gene-specific regulatory term R(t) and the global synthesis rate S(t) related to the physiological status. Therefore, rate parameters except S(t) contributed to the global regulation of protein abundance and R(t) was used to quantitatively assess the protein dynamic changes that could not be explained by the global regulation for each gene, thereby identifying gene-specific regulation dynamics. To filter out the noise inherent to experimental data and capture the main features of gene regulation, the gene-specific regulation term R(t) was parameterized as a pulse-like function [18, 24]. The parameters were estimated utilizing a global optimization algorithm, namely differential simulated annealing (DSA) [35]. In each round of parameter evaluation, we simulated the dynamics of protein concentrations and set the value of the objective function to the sum of squared errors between the simulated and observed trajectory. The Markov chain length was set to 100 and the maximum round of function evaluation was set to 2000 for each gene. Distinct modes of regulation were deconvoluted from similar dynamic patterns of protein abundance in different cell fate responses (S2 Dataset). Notably, our algorithm is capable of identifying the transcriptional down-regulation of genes despite of the induction of overall protein expression that is purely due to the alterations of physiological parameters. For instance, although the expression of FKH1 showed a 1.5-fold induction in cells undergoing shmoo formation, PSDA revealed a reduction in its protein synthesis in the last 4 h. This reduction in the protein abundance of Fkh1, which is a cell cycle regulator, could account for the cell cycle arrest observed under this condition (Fig 3B).
Our PSDA approach yielded the gene-specific dynamics that are consistent with previous microarray results [32] (S4 Fig), for most genes examined in this study. For example, synthesis of BAR1 is up-regulated after 1.5 h (Fig 3B, gray shaded window in the panel labeled “Bar1”), contributing to a delayed negative feedback that modulates the response duration. Our analysis also revealed up-regulation of genes in the MAPK cascade, such as FUS3 and STE5, consistent with previous microarray data [32] (Fig 3B, gray shaded window in the panels labeled “Fus3” and “Ste5”). More importantly, the PSDA approach can identify time intervals of regulatory events that are otherwise invisible from the protein expression profiles. For instance, we found synthesis of FUS3 was up-regulated during the intermediate phase, while several typical mating genes, such as FAR1, were induced in the late phase of experiment (Fig 3B, gray shaded window in the panels labeled “Fus3” and “Far1”). Therefore, PSDA can identify gene-specific regulatory events from protein expression profiles and reveal the temporal patterns of regulatory events in a quantitative manner.
To further characterize the regulatory modes of cells committed to the shmooing fate, we clustered genes into 6 groups according to their temporal patterns of gene-specific regulation, with 2 groups inhibited and 4 groups activated (Fig 3C; S5A Fig). Using a smaller cluster number makes it hard to detect the sequential gene regulatory steps. For example, reducing the cluster number could merge C3 and C4 into a new cluster and we cannot distinguish genes that are activated in the early phase form that activated in the intermediate phase. On the other hand, subdividing the genes into more clusters would increase computation complexity and the number of possible networks, but does not alter the general network topology (see Discussion below). The first cluster includes RP genes such as RPL1B and TFs, such as ABF1 and RAP1, which are important regulators of RP synthesis [36]. The synthesis rate of these genes is rapidly down-regulated before recovering to the pre-treatment value, which might lead to decreases in ribosome synthesis (Fig 3C, C1). The second cluster contains genes that are inhibited with a time lag of ~2 h. This cluster is enriched with cell cycle genes (S1 Table), including FKH1/2, which are TFs that mediate the expression of genes in M phase (Fig 3C, C2). Genes in cluster 3 show a rapid and transient induction of expression and are primarily involved in stress-activated signaling responses, e.g., OPY2 in the high-osmolarity glycerol (HOG) pathway and MDS3 that associates with the TOR pathway. DIG1, an important regulator of the pheromone response pathway, is also included in this cluster (Fig 3C, C3). Genes encoding the components of MAPK pathway, such as KSS1 and FUS3 are enriched in cluster 4 and are transiently induced after cluster 3 genes (Fig 3C, C4). Cluster 5 includes the genes participate in execution of yeast mating and fusion, such as FUS1 and FAR1. These genes are up-regulated slightly later than cluster 4 genes and exhibit a more prolonged induction patterns (Fig 3C, C5). Finally, cluster 6 genes, such as ISW1 and SNF2, are up-regulated very late in the response and are primarily involved in chromatin remodeling (Fig 3C, C6).
Because genes in the same cluster show similar temporal patterns, we hypothesized that they might share same upstream regulators. So we treated each cluster as a ‘meta-gene’ and generated its activation/inhibition time sequence via a threshold model (S5B Fig), which allows us to reconstruct the putative interactions among the clusters by analyzing the time trajectory using a Boolean network model [14, 16]. A wide range of parameter values was used in the threshold model to investigate the robustness of our results (S5B Fig). The discrete trajectory resulted in 4.1 x 108 possible networks, including 96 minimal networks without redundant edges. We further incorporated prior knowledge to further confine the network structure (Fig 4A). The regulatory network and representative genes in each cluster are illustrated in Fig 4B. The “input” of the network is the TFs activated by pheromone that can directly regulate gene expression, such as Ste12 and Tec1. The reconstructed network successfully captures the core structure of pheromone response pathway and reveals genetic interactions in accordance with previous knowledge of the system (Fig 4B, solid arrows). The network structure is robust to the choice of the cluster number. Subdivision of the largest cluster (C6) into smaller clusters would increase the number of possible networks, while leaving the general network structure unchanged (S6 Fig). More importantly, our results also suggested novel putative interactions in the gene expression programs induced by pheromone (Fig 4B, dashed arrows). For example, the stress response genes in cluster 3 might repress translation by inhibiting cluster 1 genes. Additionally, in the early phase of the response, cluster 2 cell cycle genes may repress the expression of cluster 6 genes, implying a coordinated regulation of cell cycle and chromatin remodeling during the mating response. In summary, these results support the use of our experimental and computational approaches in dynamic network analysis and imply the potential interactions of the canonical mating pathway with various signaling and cellular processes, offering a more comprehensive picture of the yeast mating response.
To investigate the mechanisms underlying dose-dependent mating differentiation, we next reconstructed the gene regulatory network for cells committed to chemotropic elongated growth in response to the intermediate pheromone level. To this end, we examined the deconvoluted gene-specific regulations of the six meta-genes (classified in Fig 3C) for elongated growth and compared their time trajectories with those of the cells committed to shmoo formation upon a high pheromone dose. Interestingly, although we observed striking differences in protein dynamics of all six meta-genes between the two cell fates, only the genes in cluster 1 and 3 exhibited a qualitative difference. Cluster 1 genes are down-regulated during shmoo formation in response to the high level of pheromone stimulation, but are up-regulated during elongated growth in response to the intermediate pheromone dose. In contrast, cluster 3 genes are transiently induced during shmoo formation, but are repressed during elongated growth. All the other clusters showed quantitative differences between the responses to different doses of pheromone, in which the gene regulation are towards the same direction and only differ in the extent of changes (Fig 5A).
Based on the dynamic changes of the meta-genes in response to the intermediate level of pheromone, we generated a putative regulatory network that can recapitulate the protein dynamics during elongated growth (S7A and S7B Fig). We then compared this network structure with that of shmoo formation response (Fig 5C–5D). Consistent with the analysis of protein dynamics in Fig 5A, we found that the core network structure consisting of the interactions between clusters 2, 4 and 5 remained unchanged in the two cell fates. The major difference lies in cluster 1 and 3. During elongated growth in response to an intermediate pheromone level, the RP genes in cluster 1 are induced but the stress response genes in cluster 3 are repressed. In contrast, during shmoo formation triggered by a high pheromone dose, cluster 1 genes are inhibited whereas cluster 3 genes are induced. Furthermore, since cluster 3 undergoes a slightly faster response than cluster 1 during shmoo formation, it is possible that the stress response genes in cluster 3 may contribute to the repression of RP genes in cluster 1. To examine the robustness of the uncovered networks, we further investigated the attractor landscape for cell fate determination in the mating response (S7C and S7D Fig). Although re-wiring of the regulatory network alters the attractor landscape, the two biological phenotypes emerge as the largest attractors, revealing the dynamic robustness of the cell fate commitment. Taken together, our modeling analysis suggested that the dose-dependent regulation of stress response genes might lead to the phenotypic differences associated with distinct cell fates, in which the cells undergoing elongated growth have an enhanced cell growth and biogenesis capacity, whereas the cells committed to shmoo formation feature a reduced biogenesis capacity.
To further validate the model results, we experimentally explored the molecular connections between the pheromone response pathway and the RP or stress response genes. In S. cerevisiae, induction of stress response genes and repression of RP genes are simultaneously associated with the responses to nutrient limitation or stresses and are mediated by the general stress responsive pathways, such as TOR or protein kinase A (PKA) pathways [37, 38]. Thus, we examined major stress responsive regulators in yeast, including Sfp1—a stress responsive transcription factor [30], Hog1 –a stress-activated protein kinase [39], Msn2 –a general stress responsive TF in the PKA pathway [40], Yap1 –a TF involved in transcriptional response to various stresses [41], and Tpk1 –the catalytic subunit of PKA [42]. These regulators can govern the expression of RP and/or stress response genes and their activities are primarily controlled via nucleocytoplasmic translocation. We found that pheromone stimulation had no effect on the localization of Hog1, Msn2, Yap1 or Tpk1 (S7 Fig). In contrast, whereas localized in the nucleus under vegetative or elongated growth conditions, Sfp1 translocated from the nucleus to the cytoplasm in response to a high level of pheromone stimulation leading to shmoo formation (Fig 6; S8 Fig; S1 Video). Intriguingly, Sfp1 has been best known as a stress responsive TF primarily involved in the TOR signaling pathway [30, 43]. Under optimal growth conditions, the Tor kinases are active and Sfp1 localizes in the nucleus where it activates the expression of RP and ribosomal biogenesis genes; in response to nutrient limitation or chemical stress, Tor kinases are inhibited and Sfp1 is translocated from the nucleus to the cytoplasm. Therefore, our results suggest a potential cross-regulation of the TOR-dependent ribosomal biogenesis pathway that only occurs in shmooing cells. This pathway crosstalk could serve as a molecular switch that mediates the dose-dependent network rewiring and cell fate determination during mating differentiation.
It is a challenging task to resolve regulatory networks from experimental observations in eukaryotic organisms, despite continuous advances in reverse engineering methods in the past decades [8]. One reason for this is that regulation of protein abundance is not only determined by the gene-gene interactions, but also subject to the availability of cellular resources and changes in growth dynamics. In this paper, we employed an integrated approach that combines high-throughput experiments, dynamic modeling of the protein life cycle and a reverse engineering method to investigate the regulatory mechanisms underlying distinct cell fates in the yeast mating response. Based on the high-resolution profiling of the protein abundance and growth dynamics of yeast cells, we found that gene expression can be strongly affected by the global changes in cell physiological state, hampering the identification of gene-specific regulation events that are critical for network discovery. Here, we solved this problem by taking a model-driven approach (PSDA) to assess the influence of global regulation and deconvolute the gene-specific regulation, which enables us to further reconstruct regulatory networks underlying different mating responses.
Our PSDA method confers advantages of robust detection of regulation events and little data requirement, but is not without limitations. For example, we assumed the regulation of protein level mainly takes place in the synthesis part. This assumption, supported by previous studies [23, 44], helps to reduce the number of free parameters while preserving the capacity to identify regulatory events. However, it may overlook dramatic changes in protein degradation that is crucial for some specific responses. Also PSDA quantitates the summarized effect of regulations in transcriptional and translational levels and therefore it cannot distinguish between the two effects. Further incorporation of time-resolved transcriptome data may enable more accurate quantification of transcriptional and translational regulation and comparison with relevant methods [24].
The putative regulatory networks were generated without empirical parameterization of the regulation functions, yet successfully recapitulated the dynamics of the system underlying different phenotypic responses. Our reconstructed network includes genes encoding the components of canonical pheromone response pathway that regulate signal transduction, cell cycle arrest and cell fusion. In addition, our network analysis suggests a conditional regulation of RP and stress response genes, which guided our identification of the dose-dependent translocation of a stress responsive transcriptional factor, Sfp1. Since the localization of Sfp1 is directly regulated by the activity of Tor kinases [30, 45], it is possible that Tor activity is only diminished in shmooing cells, but not in elongated cells. Intriguingly, a previous study [33] showed using microarray time course analysis that a high dose of pheromone down-regulates 54 ribosomal proteins. This down-regulation is independent of pheromone-induced cell cycle arrest and the kinetics is similar to that observed in response to rapamycin treatment. Their findings are consistent with our modeling prediction and experimental results, pointing to a possible cross talk between the mating pathway and the TOR pathway. A systematic epistasis analysis of the pheromone response pathway would provide us further insights into the mechanisms of this pathway crosstalk.
Taken together, our experimental and computational approaches represent an integrated framework for analyzing the proteomic dynamics during cell differentiation. Given the growing interests in large-scale protein dynamics and network analysis, we anticipate that our strategies would be widely applied to enable systems-level understanding of other biological systems.
The yeast cell strains used in this study were selected from a chromosomally GFP-tagged library. Strains were grown to saturation at 30°C and further diluted and cultured for another 8 hours to reach the exponential growth phase before the microfluidic experiments. The alpha factor (Sigma-Aldrich, St. Louis, MO) level was set to 5.9μM/L for the high concentration and 0.59μM/L for the intermediate concentration.
A high-throughput microfluidic chip was used in our fluorescence experiment, which allows a maximum of 96 parallel experiments across 2 different conditions (S1A Fig). Our chip was fabricated with PDMS (polydimethylsiloxane, RTV615, Momentive Performance Materials Inc.) using standard soft lithography technology. Each strain was loaded into an individual channel and a constant flow rate of 400 μl/h for fresh media was achieved. Phase contrast and fluorescence images of yeast cells were generated via a Nikon Ti-E microscope in combination with NIS-Elements software every 5 min. A cell culture incubator around the microscope was used to maintain a temperature of 30°C.
We developed a kinetic model of protein dynamic change in which:
dP(t)dt=α(t)−(d+γ(t))P(t)
where P(t) is the protein concentration at time t, α(t) represents the protein synthesis rate and γ(t) denotes protein dilution rate, which equals the exponential cell mass accumulation rate. d is the protein degradation rate estimated from a genome-wide measurement of protein half-lives [34].
We suspected that the per-gene regulations in the response mainly took place in the protein synthesis term. Our kinetic model for protein concentration was thus modified to the following form:
dP(t)dt=R(t)S(t)−(d+γ(t))P(t)
in which S(t) is the global synthesis rate that reflects the changes of cellular resources related to protein synthesis, and R(t) is the temporal regulation term of the protein. We adopted an efficient parameter optimization algorithm named differential stimulated annealing (DSA) to estimate the regulation parameters for each protein [35]. The calculation was based on a customized MATLAB version of the algorithm. For detailed information, see S1 Text.
In the Boolean network model, each node represents a biological species. Si(t)∊{0, 1} is used to denote the state of node i at time t. Regulation from node j to node i is represented by the coefficient aji, which is positive for activation and negative for inhibition. Update of the node states is described by the following Boolean functions:
{Si(t+1)=θ(∑jSj(t)aji),∑jSj(t)aji≠0Si(t+1)=Si(t),∑jSj(t)aji=0
where θ is the Heaviside step function as follows: θ(x) = 1 when x > 0 and θ(x) = 0 when x < 0. From a given initial state, the state of the system is updated until it reaches a steady state known as an attractor.
|
10.1371/journal.pntd.0005732 | High prevalence of epilepsy in onchocerciasis endemic regions in the Democratic Republic of the Congo | An increased prevalence of epilepsy has been reported in many onchocerciasis endemic areas. The objective of this study was to determine the prevalence of epilepsy in onchocerciasis endemic areas in the Democratic Republic of the Congo (DRC) and investigate whether a higher annual intake of Ivermectin was associated with a lower prevalence of epilepsy.
Between July 2014 and February 2016, house-to-house epilepsy prevalence surveys were carried out in areas with a high level of onchocerciasis endemicity: 3 localities in the Bas-Uele, 24 in the Tshopo and 21 in the Ituri province. Ivermectin uptake was recorded for every household member. This database allowed a matched case-control pair subset to be created that enabled putative risk factors for epilepsy to be tested using univariate logistic regression models. Risk factors relating to onchocerciasis were tested using a multivariate random effects model. To identify presence of clusters of epilepsy cases, the Kulldorff's scan statistic was used. Of 12, 408 people examined in the different health areas 407 (3.3%) were found to have a history of epilepsy. A high prevalence of epilepsy was observed in health areas in the 3 provinces: 6.8–8.5% in Bas-Uele, 0.8–7.4% in Tshopo and 3.6–6.2% in Ituri. Median age of epilepsy onset was 9 years, and the modal age 12 years. The case control analysis demonstrated that before the appearance of epilepsy, compared to the same life period in controls, persons with epilepsy were around two times less likely (OR: 0.52; 95%CI: (0.28, 0.98)) to have taken Ivermectin than controls. After the appearance of epilepsy, there was no difference of Ivermectin intake between cases and controls. Only in Ituri, a significant cluster (p-value = 0.0001) was identified located around the Draju sample site area.
The prevalence of epilepsy in health areas in onchocerciasis endemic regions in the DRC was 2–10 times higher than in non-onchocerciasis endemic regions in Africa. Our data suggests that Ivermectin protects against epilepsy in an onchocerciasis endemic region. However, a prospective population based intervention study is needed to confirm this.
| An increased prevalence of epilepsy has been reported in many onchocerciasis endemic areas. Between July 2014 and February 2016, house-to-house epilepsy prevalence surveys were carried out in the Bas-Uele, Tshopo and Ituri province of the Democratic Republic of the Congo, in areas with a high level of onchocerciasis endemicity. Of 12, 408 people examined in the different health areas 407 (3.3%) were found to have a history of epilepsy. A nested case control analysis demonstrated that before the appearance of epilepsy, compared to the same life period in controls, persons with epilepsy were around two times less likely to have taken Ivermectin than controls. Our data suggests that Ivermectin protects against epilepsy in an onchocerciasis endemic region. However, a prospective population based intervention study is needed to confirm this.
| A high prevalence of epilepsy (>1%) has been reported in many onchocerciasis endemic regions. In Africa, a special form of epilepsy referred to as “nodding syndrome” has been reported solely in onchocerciasis endemic areas [1]. The Democratic Republic of the Congo (DRC) is the country with the largest population at risk for onchocerciasis, with approximately 40.3 million people exposed and 29.7 million currently treated [2,3]. In the DRC, despite 16 years of Community Directed Treatment with Ivermectin (CDTI), the therapeutic coverage of Ivermectin is not consistent spatially or from year to year, and some areas remain untreated (have received no CDTI at all). According to the WHO, the geographic coverage for the country in 2007 was 72.2% and the therapeutic coverage 43.5% and in 2015, 93.3% and 73.3% respectively [3,4]. The inconsistent therapeutic coverage combined with the high level of onchocerciasis endemicity, may thus have direct consequences for the epilepsy prevalence in endemic areas in the DRC. In 2014, we documented the prevalence of epilepsy to be 2.9% in the village of Dingila and 2.3% in Titule, both in the Bas-Uele province [5]. In Titule, epilepsy showed a marked spatial pattern with clustering of cases occurring within and between adjacent households. Individual risk of epilepsy was found to be associated with living close to the nearest fast flowing river where blackflies (Diptera: Simuliidae)–the vector of Onchocerca volvulus (O.v)–oviposit and breed [5]. A small case control study in Titule suggested that Ivermectin may protect against onchocerciasis associated epilepsy or OAE [6].
In this study, we investigated in door-to-door village surveys whether there was increased epilepsy prevalence in other onchocerciasis endemic foci of the DRC, in the provinces of Bas-Uele, Tshopo and Ituri (previously part of the “Oriental Province”). Moreover, in a nested case control study we examined whether there was evidence for annual intake of Ivermectin being able to provide significant protection against epilepsy.
Definition: A case of epilepsy was defined as a patient who reported at least 2 unprovoked seizures without fever or any acute illness [7].
Between July 2014 and February 2016, house-to-house epilepsy prevalence surveys were carried out in 3 localities of Bas-Uele covering 164 households, 24 localities of Tshopo covering 1322 households and 19 localities of Ituri province covering 570 households (Fig 1).
The localities were selected on the basis of a historically high level of onchocerciasis endemicity based on Rapid Epidemiological Monitoring of Onchocerciasis (REMO) (1998 or 2003) data provided by the National Onchocerciasis control program (Programme National de Lutte contre l’Onchocercose, PNLO) [8]. In Ituri localities within one health area were chosen because they had never been included in the CDTI program and as such never received Ivermectin.
In the three provinces, houses were selected starting from the localities centres’, selecting every third household. If household members of the selected house were not at home, the next house was visited; all households were geo-located (handheld Garmin 62Cs GPS; ±4m accuracy). This method was implemented in all study sites except in the Tshopo province, in the Wanierukula health zone, where all houses on a 40km stretch of the national road linking Kisangani (Tshopo province) to the Ituri province were visited. The latter region is referred to as the “PK30-PK70” region (as it does not exist, administratively) and contains 1182 unique households across 2 health areas consisting of 14 localities. In every area, all household heads and parents of children present in the household were interviewed in their local language. For every consenting household, a one page questionnaire was completed (available as supporting information). Age and sex of every household member and their individual history of Ivermectin uptake each year between 2000–2015 were recorded. Screening for epilepsy was performed using the five questions questionnaire validated by Diagana et al [9]. The research team who visited the households consisted of one or two local health care workers or Ivermectin community distributers and a medical doctor. If a person with epilepsy was identified, family members were asked by a doctor to describe the type of seizures (or to show what happens during a seizure), to report on the precipitant circumstances, the duration of seizures, whether they were associated with uncontrollable tongue biting or passing of urine or stool, whether there were episodes of absence (sudden episodes of decreased consciousness of sudden onset and short duration) with or without nodding of the head. Questions were asked in the local language Boa, Lingala, Kiswahili or Lendu according to the site. The final diagnosis of epilepsy was made by a medical doctor.
For household members with epilepsy confirmed by the doctor, the age of the first epilepsy episode was noted. For those who developed epilepsy recently, the exact month of the first epilepsy attack was recorded. Onset of epilepsy was considered recent if a first seizure appeared in the 12 months before the survey date.
The prevalence of epilepsy in the populations surveyed was calculated across selected health areas of the three provinces (Ituri, Tshopo, and Bas-Uele). The data on the prevalence of skin onchocerciasis lesions (leopard skin) and therapeutic coverage were obtained through household interviews. The prevalence of people with onchocerciasis nodules (determining the level of onchocerciasis endemicity) were either obtained from the PNLO database or measured in the field using the standard REMO procedure [8]. The number of years of Ivermectin distribution in each health area was also obtained from the PLNO database. Ivermectin coverage per health area was calculated as the number of individuals in the health are who reported to have taken Ivermectin in 2014 over the number of individuals eligible to receive Ivermectin in that year. As eligibility criterium for Ivermectin treatment we only used age > 5 years old; we did not took into account whether a women was pregnant or breast feeding.
Case control pairs were generated from individuals in the prevalence survey database by matching individuals from the same health area, of the same gender and birth year. In this case control sub-study, only cases were included who were old enough to be eligible for Ivermectin distribution the year before they became epileptic (at least 6 years old at epilepsy appearance), and who lived in health areas where Ivermectin was distributed under the CDTI program in the year before they became epileptic.
Univariate binomial logistic mixed regression models were constructed to identify whether there was an association between the epilepsy status and the individual Ivermectin treatment history. A random effect term was included for pair identity. For cases (persons with epilepsy), this was considered as Ivermectin treatment in the year immediately before epilepsy was identified, and the proportion of years where Ivermectin was received in the years before and after the reported onset of epilepsy. For controls (persons without epilepsy) the Ivermectin history was considered similarly, around the age at which their paired case became epileptic.
The association between epilepsy and potential risk factors, including the proportion of Ivermectin doses received on occasions where the individual was eligible and the presence of onchocerciasis-associated skin lesions, was assessed through the construction of a binomial logistic mixed regression model using the whole survey data base, without case control pairing. Age and gender were controlled for as categorical fixed effects, and the health area included as a random effect to control for clustering at this level.
To identify presence of clusters of epilepsy cases, the Kulldorff's scan statistic [10] as implemented in SaTScan software (https://www.satscan.org/) was used. The spatial scan statistic tests for spatial randomness of cases over the identified region. It defines a set of potential cluster areas (spatial circles of varying size), each consisting of a collection of cases. The most “unusual” cluster is then identified using the likelihood ratio test statistic which is based on the alternative hypothesis that the risk of the disease is greater inside than outside the circle. The most likely cluster is the circle with the maximum likelihood. For the epilepsy dataset, cases were defined as individuals with epilepsy while controls are those without epilepsy. The Kuldorff’s scan statistic was performed for each of the three administrative regions separately: Ituri, Tshopo and Bas-Uele province.
All statistical analyses (aside from the scan statistic above) were implemented using the R statistical computing environment [11]. The generalized linear mixed model was implemented using the package “lme4” [12].
The study was approved by the Institutional Review Board of Ngaliema hospital in Kinshasa and the ethical review board of the University of Antwerp. Written informed consent was obtained from the head of the family and parents/guardians of children. For people who could not write, consent was obtained by finger printing.
Of the 12,408 people examined in the different health areas, 407 (3.3%) were found to have a history of epilepsy. Across the localities, the mean number of household members was 5.65 (s.d. = 3.35). Whilst the median ages of the persons with epilepsy (18 years, IQR = 13, 23.75) and without epilepsy (16 years, IQR = 6, 35) were relatively similar, the age distributions of the persons with and those without epilepsy were significantly different (Two Sample Kolmogorov-Smirnov test, D = 0.26, P<0.001) (Fig 2).
The median age of onset of epilepsy was 9 years; the modal age of onset was 12 years. The highest prevalence of epilepsy was observed in the 10–19 years age group (Table 1).
Epilepsy appeared most frequently in the 10–15 year age group (Fig 3).
In 41 families there were at least 2 persons with epilepsy (1.5% of all households), and in 9 families at least three (0.3% of all households). Amongst the persons with epilepsy, 30% (123) lived in households with another person with epilepsy. All persons with epilepsy had experienced seizures in the last 5 years.
A high prevalence of epilepsy was observed in all health areas in the 3 provinces: 6.8–8.5% in Bas-Uele, 0.8–7.4% in Tshopo and 3.6–6.2% in Ituri (Table 2).
Ninety six cases and controls perfectly matched for village, age and gender were identified (Table 3).
Within the case control pairs, before the appearance of epilepsy, compared to the same life period in controls, persons with epilepsy had taken less frequently Ivermectin than controls (Table 4).
After the appearance of epilepsy, there was no difference of Ivermectin intake between cases and controls.
Using the whole survey database, male gender, onchocerciasis skin lesions and being treated, at least once with Ivermectin, were associated with a significantly higher risk of being a person with epilepsy (Table 5).
Among the patients without leopard skin, those who used Ivermectin were twice as likely (OR = 2.04) to have epilepsy as compared to those who did not use Ivermectin. However, among the patients with leopard skin, thus showing signs of onchocerciasis, those who used Ivermectin were 84% less likely (OR = 0.16) to have epilepsy than those who did not use Ivermectin.
With regard to clustering of epilepsy cases, based on the Kulldorff scan statistics, 5 most likely clusters of epilepsy cases were detected in Ituri, 9 in Tshopo and 2 in Bas-Uele. However, most identified clusters had non-significant p-value (>0.1). In fact, of the 2 most likely clusters, one in Ituri and the other one in Tshopo (Fig 4), only one had p-value less than 0.05.
The cluster in Ituri had p-value of 0.0001 with a radius of 1.41 km and is located around the Draju sample site area with the Muda and Kuda rivers as nearest potential Simulid breeding sites; however the local vector remains to be identified. On the other hand, the cluster in Tshopo had p-value of 0.085 with a radius of 2.43 km and is located in between the Makana and Salambongo sample sites close to the Mobi and Onane rivers where members of the Simulium neavei complex were found on crabs. The most northern cluster in the Tshopo Province is located around the Tshopo rivers where Simulium damnosum s.s. were aggressively biting the villagers at the time of the study (data on blackflies exposure per province to be published elsewhere).
The observed prevalence of epilepsy in communities of onchocerciasis endemic regions in the DRC was comparable with those reported in other onchocerciasis endemic areas in Africa [14] but higher than the epilepsy prevalence (0.4–1%) reported in most studies in the rest of the world [15–19]. In a study of 586,607 residents in five Health and Demographic Surveillance System centres in sub-Saharan Africa, only 1,711 (0.29%) individuals were diagnosed as having active convulsive epilepsy [20]. However, we did observe large differences in the prevalence of epilepsy in different health areas, although the causes of these differences were not easy to explain. The localities involved in this study were spread across a large geographical (and ecological) range, which is likely to encompass different levels of risk of exposure to the blackfly vector. In the Ituri province, the prevalence of epilepsy was higher in the village where Ivermectin had never been distributed. In Aketi on the other hand, despite 13 years of CDTI and a relatively high coverage of Ivermectin, the prevalence and incidence of epilepsy and the prevalence of onchocerciasis skin lesions was high. Wela is a village situated only 100m from the Agu rapids on the Itimbiri river, a historically known Simulium damnosum breeding site. In 1999, in Wela, before the introduction of CDTI, 98% of the men examined presented with onchocerciasis nodules (REMO). A high S. damnosum density was observed in Aketi in 2016, but we did not investigate the infection rate of the flies. It would be important to investigate whether in the past in the Aketi health zone, the Iverrmectin therapeutic coverage has been much lower than the one reported in 2014. In this health zone, we plan qualitative research to obtain information about the adherence to the annual CDTI program and a case control study to investigate whether other parasitic infections such as cysticercosis could explain the high prevalence of epilepsy. Alternative hypotheses may be i) that in this health zone, the population of O.v. may have developed Ivermectin resistance, ii) a higher vectorial capacity of S. damnosum sensu stricto compared to other vectors (unidentified so far in the Ituri province and S. damnosum-complex in Tshopo province), iii) higher density of infective vector and hence an increased level of exposure to Simuliidae infective bites. The infestation rate of the Simulium spp. in the three foci would help answering some of these questions.
The peak incidence of epilepsy in patients in this study was around the age of 12. This aligns with other onchocerciasis endemic African regions, and can most likely be explained by the equally high incidence of O.v. infection in these age groups and the cumulative nature of the O.v. infestation [21]. This peak incidence is in contrast with the epilepsy situation in industrialized countries and in non-onchocerciasis endemic regions in Africa where most onset of epilepsy is observed in the very young (< 5 years old) and in the older population [21]. It is important to note that not only was there a high prevalence of epilepsy in the 10–19 age group, but also in the 20–29 age group. The latter is in contrast with the epilepsy prevalence reported in onchocerciasis hyperendemic regions where Ivermectin has not yet been introduced. For example, in a study from 1991 in Kyarusozi sub-county in western Uganda, 91% of the epilepsy cases were below the age of 19 [22]. This epilepsy age shift towards older age groups after several years of annual Ivermectin distribution is an argument that Ivermectin may reduce the incidence of epilepsy in children [23].
Around a third of the persons with epilepsy lived in a household with at least one other person with epilepsy. In a study in Europe, only 9.5% of persons with epilepsy had first- or second- degree relatives with seizures [24]. In Ituri, evidence of a significant clustered distribution of households with family members with epilepsy compared to those without epilepsy was observed. This aligns with observations from other onchocerciasis endemic areas [25] and most likely reflects shared exposure to infective bites of the local vector(s). The fact that only a relevant epilepsy cluster was observed in the Logo Health zone in Ituri may be because this was also the only zone, included in the survey, where in certain villages Ivermectin was never distributed. A (spatial) point process model in order to predict the risk of epilepsy due to the infective vector density could be of interest. However, this was not possible given the available data (sample locations were too irregular and clustered only in specific areas).
Similar to the results of a previous small case control study performed in Titule (Bas-Uele), the current case-control matched pairs analysis demonstrated that before the appearance of epilepsy, compared to the same life period in controls, persons with epilepsy had taken Ivermectin less frequently than controls. Considering the whole population, significant associations were identified between Ivermectin use, the presence of skin lesions and epilepsy status, even with fixed factor controlled for age and gender. Among patients with leopard skin, those who used Ivermectin were 84% less likely to have epilepsy than those who did not use Ivermectin. Leopard skin is a clinical sign that indicates that the person has been heavily exposed to O.v. and/or that the person did not take Ivermectin while infected with onchocerciasis. This finding may indicate that also when infected with onchocerciasis Ivermectin may reduce the risk for epilepsy. Among patients without leopard skin, those who used Ivermectin were twice likely to have epilepsy as compared to those who did not use Ivermectin. This could be explained by the fact that people use Ivermectin because of itching caused by onchocerciasis. People know that the itching disappears with the intake of Ivermectin. In a previous study, individuals reporting itching were more likely to be persons with epilepsy [6]. After the intake of Ivermectin, microfilariae disappear from the skin together with the itching and people stop scratching. Therefore, because of decreased scratch lesions and skin inflammation people taking Ivermectin are less likely to develop Leopard skin lesions but they are more likely to have been infected with the O.v. parasite in the past and therefore are more likely to have epilepsy. There are no arguments for suspecting that Ivermectin is able to cause seizures. Indeed, Ivermectin is not known to pass the blood brain barrier in humans. Only in certain dogs, collies and Australian shepherds because of MDR1 polymorphism, Ivermectin is able to cause seizures [26]. Ivermectin is rapidly absorbed. Therefore if Ivermectin could induce seizures, we would expect to see the onset of these seizures shortly after the administration of the Ivermectin. This was never reported despite the many millions of doses of Ivermectin that have been distributed for more than 20 years. Moreover, there is anecdotal evidence that in persons with onchocerciasis Ivermectin reduces the frequency of epilepsy [27]. In Aketi town, people were very eager to receive Ivermectin even more than once a year because the drug was known to reduce the itching and the community distributors treated themselves 3 to 4 times a year (Tepage F, personal communication).
Our study has several limitations. In most individuals, epilepsy was only reported and not observed. Only in the Makana health area (Tshopo) and Ituri (Logo and Rethy health areas), were individuals with epilepsy examined by a neurologist (JMK). A second neurologist, Dr D Mukendi recently visited the villages of Wela and Makoko and confirmed all the epilepsy cases that were identified during the 2015 survey. Laboratory investigations, to identify the cause of the epilepsy were not performed. In our definition of epilepsy, we did not specify a time limit for laps between seizures, and we mainly included convulsive epilepsy. Therefore, comparison of our results with published epilepsy prevalence data is difficult.
Additionally, the prevalence assessment was performed during short visits to the health zone with no accompanying qualitative studies. As discussed above, the exact nature of the relationship between Ivermectin use and epilepsy may be complex and as such it is likely to require knowing why people were taking or not taking Ivermectin. The different epilepsy surveys were performed with a similar methodology, however not always by the same research team. Certain data obtained during the surveys may suffer from recall bias. For example even age and age/year of onset of the epilepsy may be very imprecise. Questions could be raised about the reliability of the history of Ivermectin use, in certain individuals over a period of more than 10 years. Available REMO data are from 2003 and 2008 and are of little relevance after many years of CDTI. Indeed in the epilepsy case control studies we performed in Tshopo and Ituri a reduction in nodule carriers was observed.
In conclusion, the prevalence of epilepsy in villages in onchocerciasis endemic areas in the DRC was 2–10 times higher than in non-onchocerciasis endemic regions in Africa. Our study confirms previous findings that epilepsy is associated with O.v. infestation. Our data suggests that Ivermectin protects against epilepsy in an onchocerciasis endemic region. This finding suggests that in onchocerciasis endemic regions, the physiopathology of the epilepsy characterized by seizures starting between the ages of 3 to 18 years, is triggered by the O.v. parasite. Because microfilariae are not considered to invade the brain, the mechanism of this epilepsy could be through an auto-immune mechanism as was recently proposed for nodding syndrome [28,29]. However, a prospective population-based intervention study, ideally with Ivermectin twice a year, is needed to confirm that indeed Ivermectin protects against epilepsy in onchocerciasis endemic regions.
|
10.1371/journal.pntd.0005797 | Temperature modulates dengue virus epidemic growth rates through its effects on reproduction numbers and generation intervals | Epidemic growth rate, r, provides a more complete description of the potential for epidemics than the more commonly studied basic reproduction number, R0, yet the former has never been described as a function of temperature for dengue virus or other pathogens with temperature-sensitive transmission. The need to understand the drivers of epidemics of these pathogens is acute, with arthropod-borne virus epidemics becoming increasingly problematic. We addressed this need by developing temperature-dependent descriptions of the two components of r—R0 and the generation interval—to obtain a temperature-dependent description of r. Our results show that the generation interval is highly sensitive to temperature, decreasing twofold between 25 and 35°C and suggesting that dengue virus epidemics may accelerate as temperatures increase, not only because of more infections per generation but also because of faster generations. Under the empirical temperature relationships that we considered, we found that r peaked at a temperature threshold that was robust to uncertainty in model parameters that do not depend on temperature. Although the precise value of this temperature threshold could be refined following future studies of empirical temperature relationships, the framework we present for identifying such temperature thresholds offers a new way to classify regions in which dengue virus epidemic intensity could either increase or decrease under future climate change.
| Recurrent, rapidly intensifying epidemics of dengue–the world’s most prevalent mosquito-borne viral disease–pose a challenge to healthcare systems throughout the tropical and subtropical world. An acute disease that tends to respond well to proper treatment, the sometimes intense nature of dengue epidemics has been known to overwhelm healthcare systems and elevate the morbidity and mortality of patients left without adequate medical treatment under peak epidemic conditions. Here, we quantify the temperature dependence of dengue epidemic intensity by quantifying two distinct determinants of epidemic growth rate: the average number of secondary infections arising from each primary infection and the average time between successive infections in humans. Our results show that the time between successive infections in humans decreases steadily with increasing temperatures, whereas the average number of secondary infections peaks at intermediate temperatures. Altogether, this suggests a peak temperature for dengue epidemic intensity. Applying this result to global temperature projections under future climate change scenarios suggests that dengue epidemics in many regions of the world could become more intense under future temperature increases.
| Dengue virus (DENV) is a mosquito-borne pathogen that infects hundreds of millions of people each year across as many as 128 countries [1]. Along with numerous other arthropod-borne viruses (arboviruses), including chikungunya and Zika viruses [2,3], DENV causes epidemics with considerable public health impact. Rapidly growing, intense epidemics can overwhelm healthcare systems [4], leaving those infected without adequate medical treatment and with a significantly elevated risk of mortality to a disease that is seldom fatal when proper treatment is available [5].
A number of factors can lead to variability in the frequency and severity of arbovirus epidemics, including importation probability [6], host susceptibility [7], and climatic conditions [8]. In particular, temperature is known to be a major driver of spatial and temporal variability in arbovirus transmission, as indicated by empirical studies of relationships between temperature and several epidemiologically important vector and pathogen traits, including mosquito lifespan [9–11], incubation time of the pathogen in the mosquito [9,10,12], the rate at which mosquitoes engage in blood feeding [9,13], and mosquito density [14].
Analyses of the effects of temperature on vector-borne pathogen transmission have focused primarily on the basic reproduction number R0 through the effects of temperature on the aforementioned vector and pathogen traits [11,15,16]. Defined as the average number of secondary infections arising from a primary infection in a fully susceptible population, R0 is a fundamentally important epidemiological quantity, because it is informative about the conditions under which a pathogen can invade, or be eliminated from, a host population. The generation interval, which is the period of time separating sequential infections, is the temporal analogue of R0. Through a fundamental mathematical relationship [17], R0 and the generation interval are related to the epidemic growth rate r, which is defined as the per capita change in incidence per unit time and characterizes the dynamics of early-stage epidemic growth in a susceptible population. Because the relationship between r and temperature has never been characterized for arboviruses, there is little scientific basis for understanding how epidemic growth rates may be related to temperature.
Our goal was to quantify the effects of temperature on DENV epidemic growth rates by first establishing a probabilistic description of DENV generation intervals as a function of temperature. We then combined our generation interval calculations with a temperature-dependent formulation of the basic reproduction number, R0, and solved for the epidemic growth rate r as a function of temperature. This new capability to calculate r as a function of temperature allowed us to identify temperature ranges that maximize r and to classify regions by their potential for increasing or decreasing epidemic growth rates based on their current and future temperatures. Our results and the accompanying code are made freely available online at https://github.com/asiraj-nd/arbotemp to facilitate the incorporation of temperature-dependent descriptions of these quantities into future studies.
We first describe our formulation of each of three major metrics of mosquito-borne pathogen transmission: the generation interval, the basic reproduction number R0, and the epidemic growth rate r. The first two are calculated a priori as a function of many of the same temperature-dependent parameters, whereas the third is derived from the first two using a fundamental mathematical relationship among all three. We then describe analyses that we performed to evaluate the epidemiological significance of these three different measures of how temperature impacts dengue virus transmission.
We define the generation interval as the elapsed time between a primary human infection and a secondary human infection deriving from that primary human infection via two bites from the same individual mosquito [18]. To derive a quantitative, probabilistic description of the generation interval for dengue, we adapted an existing framework that defines the generation interval as a sum of random variables for each of four sequential, constituent phases of the transmission cycle [19]. Similar to a recent analysis for Plasmodium falciparum malaria [20], we furthermore quantified each of these phases of the transmission cycle as dependent on temperature (Fig 1). Following Huber et al. [20], we defined these phases as: (1) the intrinsic incubation period (IIP); (2) the period between onset of symptoms in humans and subsequent transmission to mosquitoes (human-to-mosquito transmission period, HMTP); (3) the extrinsic incubation period (EIP); and (4) the period between a mosquito becoming infectious and subsequent transmission to humans (mosquito-to-human transmission period, MHTP) (Fig 1). Below, we describe the derivation and parameterization of each of these phases of the transmission cycle as four independent random variables based on available data [13,21,22]. To obtain a single random variable describing the generation interval as a whole, we took the sum of the four constituent random variables in Fig 1 by applying the convolution theorem, which involves taking the inverse Fourier transform of the product of the Fourier transforms of each random variable [23].
The basic reproduction number (R0) is defined as the average number of secondary infections in humans originating from a single primary human infection introduced into a fully susceptible population. We used the formal definition of R0 for mosquito-borne pathogens based on a set of classic “Ross-Macdonald” assumptions [29], which takes the temperature-dependent form
R0(T)=m(T)bca(T)2e−μ(T)n(T)μ(T)γ,
(1)
where m(T) is the mosquito-to-human ratio as a function of temperature T, μ(T) is the mean daily mortality rate of adult mosquitoes at temperature T, b and c are human-to-mosquito and mosquito-to-human infection probabilities, a(T) is the mosquito biting rate as a function of temperature, 1/γ is the average duration of infectiousness in humans, and n(T) is the mean extrinsic incubation period at temperature T. We note that the mean daily mortality rate of adult mosquitoes, μ(T), is the inverse of the mean for the MHTP distribution used in obtaining the generation interval distribution, while the mean extrinsic incubation period, n(T), is the mean for the EIP distribution, also used in obtaining the generation interval distribution.
Our parameterization of the ratio c/γ equaled the integral of the non-normalized HMTP curve describing the infectiousness of humans to mosquitoes over time [22], as noted in the section describing the generation interval. The parameter b did not appear in our description of the generation interval, because it affects only the magnitude of transmission (i.e., R0) rather than its timing (i.e., generation interval). This parameter is poorly understood empirically, so we chose a value of b = 0.4 consistent with a previous model [30]. We described biting rate a as a function of temperature T (i.e., a(T)) using two temperature-dependent estimates based on the average duration of the Ae. aegypti gonotrophic cycle [9,31], similar to how gonotrophic period was incorporated into the generation interval. This process involved weighting the temperature-dependent length of the first cycle and the temperature-dependent length of each subsequent cycle based on the probability of the mosquito surviving to a given number of cycles (see S1 Appendix for mathematical derivation). To capture one potential effect of temperature on the ratio of mosquitoes to humans m, we assumed that m(T) = λ / μ(T) consistent with equilibrium assumptions of a mosquito population with adult mortality rate μ(T) and constant parameter λ, which is the ratio of the daily rate of adult female mosquito emergence and the number of humans subject to feeding by the mosquitoes represented by m(T) [32]. Because values of λ are highly variable in space and time for reasons other than temperature variation, we examined the sensitivity of the value of λ across a range of values 0.0–0.5. We arrived at 0.5 as an upper limit for λ by dividing an upper limit for R0 based on independent estimates (maximum of 7.8 [33]) by all other terms on the right-hand side of Eq 1 (19.73 at 32.5°C). This is equivalent to assuming that one new adult female mosquito emerges from larval habitats every other day for each human at risk of biting within a given population.
To account for uncertainty associated with values of R0 that we calculated, we generated 1,000 Monte Carlo samples from the uncertainty distributions of each model parameter as described in each of the references [9,12,13] in which those parameters were originally described. For μ(T) and n(T), we took random draws of their parameters consistent with published descriptions of uncertainty in the parameters of these functions from their original sources [13,14]. For a(T), we used nonlinear least-squares estimates of the first gonotrophic period’s ρ parameter in the model by Focks et al. [9] by refitting it to their data, resulting in mean 8.83x10-3 and standard deviation 3.8x10-4. We assumed similar uncertainties (standard deviation) around the ρ parameter for the second gonotrophic period proposed by Otero et al. [31]. We then took random draws from normal distributions describing uncertainties in these two parameters and weighted the resulting two temperature-dependent biting rates (inverses of the gonotrophic periods) according to the probability of the mosquito surviving to a given number of gonotrophic cycles, as described in S1 Appendix. A summary of parameters and their default values is available in S4 Table.
Given temperature-dependent formulations of R0(T) and the DENV generation interval g(t) described above, we solved for the corresponding epidemic growth rate r(T) as a function of temperature by applying the result
1R0(T)=∫0∞e−r(T)tg(t)dt
(2)
from Wallinga and Lipsitch [17]. Although this does not yield an explicit relationship between r and T that can be probed analytically, it does provide a way of numerically characterizing the impacts of temperature on r. We further note that this approximation of r(T) assumes a fully susceptible, well-mixed population of mosquitoes and hosts.
We first derived a formulation of the generation interval for dengue, stochastic variability therein, and its dependence on temperature based on the assumptions described above. We then performed analyses of the relationship between temperature and r, including identification of the temperature that maximizes r and how incremental changes in r driven by changes in temperature can be attributed to distinct contributions from changes in R0 versus changes in the generation interval. For comparison with our detailed formulation of the epidemic growth rate r, we examined two approximations of the generation interval commonly used in transmission models: a fixed-length generation interval and an exponentially distributed generation interval. For each, we considered two formulations: one with a mean generation interval of 16 days [34] and one with temperature-dependent mean generation interval as calculated using our method.
We next considered how average monthly temperature data at 5 km x 5 km resolution for each month of the year based on historical records (average for 1950–2000) [35] may change epidemic growth rates under climate change scenarios. For this analysis, we used three different scenarios for mean temperature in 2050 (average for 2040–2060) corresponding to Representative Concentration Pathways (RCPs) that describe a set of alternative trajectories for the atmospheric concentration of key greenhouse gases: RCP 8.5, high greenhouse gas concentration scenario; RCP 6.0, medium baseline (or high mitigation) scenario; and RCP 4.5, intermediate mitigation scenario [36]. We obtained gridded population estimates for the year 2000 from the Global Rural/Urban Mapping Project [37] and for 2050 by projecting values from 2000 onward according to medium-fertility population projections for each country [38]. We excluded regions from this analysis where Ae. aegypti occurrence probabilities fall below 0.8, a threshold value that separates two distinct modes of local occurrence probabilities globally [39,40].
Potential for diurnal temperature fluctuations to influence DENV transmission has been suggested by temperature effects on extrinsic incubation period (EIP) and mosquito survival [10]. We examined potential effects of diurnal temperature fluctuations on the generation interval, basic reproduction number R0, and epidemic growth rate r by introducing an 8°C diurnal temperature range (DTR) around all mean temperatures. We assumed a sinusoidal progression within the day with a decreasing exponential curve at night [9,41]. We also assumed an absolute maximum temperature for Ae. aegypti survival of 37.73°C over three consecutive hours and 40.73°C in any single hour, as well as a maximum temperature of 45.9°C in any hour of the day for DENV incubation to take place, similar to assumptions of another recent model of temperature-dependent viral transmission by Ae. aegypti mosquitoes [42].
We developed a probabilistic description of the DENV generation interval by sequentially summing random variables associated with each phase of the transmission cycle (Fig 1). Allowing each of these component random variables to depend on temperature (Fig 2A) resulted in a description of the generation interval that was itself strongly dependent on temperature and captured variability and uncertainty in the underlying components (Fig 2B). For example, mean generation interval halved from 30 to 15 days with a change in temperature from 25 to 35°C. Sensitivity of the mean generation interval to changes in temperature was nonlinear, with steeper changes at more extreme temperature values (Fig 2B) due to increasing steepness of the relationships between temperature and the component random variables (Fig 2A).
The basic reproduction number, R0, was also sensitive to temperature, as it includes the same temperature-dependent random variables as the generation interval. At low temperatures, increases in temperature caused a steady increase in R0 due to a shortening extrinsic incubation period and increasing biting frequency (Fig 2A and 2C). Beyond a peak temperature of 32.5°C, R0 decreased rapidly with increasing temperatures due to rapidly increasing mosquito mortality (Fig 2A and 2C). This result contrasted with a lower peak temperature (~29°C) that was obtained in our analysis (not shown in figures) under an assumption that biting rate did not depend on temperature.
Effects of temperature on the DENV generation interval and R0 contributed to similar effects on epidemic growth rate, r. Under mean estimates of model parameters, r increased with temperature until it peaked at 33°C (Fig 2D). Under 1,000 Monte Carlo samples of model parameters, peak temperature for r varied within a relatively narrow band with 95% of values falling between 32.6 and 33.2°C (Fig 3). As both R0 and the generation interval are temperature-dependent, changes in r due to temperature occur through both components. At a constant mosquito emergence rate λ, changes in R0 accounted for the majority of changes in r, although changes in the generation interval accounted for a greater degree of change near extreme and peak temperature regions (Fig 4; S1 Fig).
Allowing for diurnal temperature fluctuations (8°C daily temperature range for all mean temperature values) shortened the mean generation interval and increased its variance relative to a scenario with no diurnal temperature fluctuation (S2 Fig). Similarly, R0 decreased when DTR was considered, as the temperature at which R0 peaks decreased from 32.5 to 30.9°C due to the effect of daily temperature extrema (under DTR) on mosquito survival (S3 Fig). The combined effect of these changes on epidemic growth rate was a slight decrease, while the temperature at which the epidemic growth rate peaks remained close to its value under a scenario with no diurnal temperature fluctuation (Fig 3; S2–S4 Figs).
Because a fully detailed generation interval distribution is beyond the capabilities of many commonly used modeling frameworks [43], we examined the correspondence between epidemic growth rates r calculated under our detailed approach and under four less detailed approximations of the generation interval that are commonly used in transmission models. A fixed-length generation interval yielded a consistently better approximation of our detailed calculations of r as a function of temperature than did an exponentially distributed generation interval (Fig 5A vs. 5B). Calculations of r under the fixed-length approximation tended to match calculations of temperature-dependent r under the detailed generation interval distribution particularly well in temperature ranges of significance to epidemics (i.e., where r > 0) (Fig 5B). For both fixed-length and exponential generation interval distributions, allowing their mean values to follow the temperature-dependent model improved their correspondence with our detailed formulation of temperature-dependent r (Fig 5A & 5B vs. 5C & 5D). These differences in r resulting from different assumptions about the distribution and temperature dependence of r could be of significance to epidemic projections, given that differences in r as small as 0.01 can lead to differences in incidence projections of an order of magnitude only a few months into an epidemic (Fig 6).
Our result that the temperature threshold for maximum r was relatively constant around 33°C (95% CI: 32.6–33.2°C) offers a useful reference point. In a given area and with other factors held constant, an increase in temperatures beyond this threshold would imply first a rise and then a fall in r between present and future. An increase in temperatures that never exceeds this threshold would imply an increase in r between present and future. At most times of year in most regions of the world that are suitable for DENV transmission, temperature increases by 2050 are expected to fall into the latter category (i.e., remaining below 33°C), suggesting that temperature changes could increase epidemic growth rates in those areas (S1–S3 Tables, S5–S16 Figs). On the other hand, temperature increases by 2050 in regions such as India and the African Sahel are expected to exceed 33°C during April-June, potentially resulting in lower epidemic growth rates in those areas during a portion of the year (S1–S3 Tables, S5–S16 Figs).
The central advance that we have made is the development of a probabilistic description of the generation interval for dengue virus (DENV) that is based on first principles of transmission, synthesizes pertinent data for DENV and Ae. aegypti, and characterizes the generation interval as a function of temperature. Although there is little data with which to independently validate our calculations, the mode of our generation time distribution at optimal temperatures for transmission (approximately 16 days at 28–32°C) accords with independent estimates of this quantity based on statistical analyses of spatiotemporal dengue case data from Thailand (15–17 days) [34]. Combining this result with a temperature-dependent description of the basic reproduction number, R0, we obtained a temperature-dependent description of the epidemic growth rate, r. All of these quantities were estimated explicitly for DENV but are also relevant for other arboviruses such as chikungunya and Zika, given their similar ecology and given that many of the parameters we used are not specific to any one virus but instead to their common vector.
The generation interval has a wide range of applications in epidemiology, including the identification of sources of infection [44], the establishment of causal linkages between cases [45], and the characterization of temporal variation in transmission [8,46]. These and other studies have typically assumed a static generation interval of either fixed length [47] or with some standard statistical distribution [48]. Our result that the generation interval for DENV is not static but is instead highly dynamic with respect to temperature highlights that transmission models for DENV and other arboviruses could be systematically inaccurate by excluding temperature-dependent effects. Future work will be needed to address the existence and significance of any such inaccuracies, but our results about the sensitivity of r to the form of the generation interval and temperature dependency therein suggest that these effects could be substantial.
Our calculations of R0 are consistent with the notion that temperature plays an important role in determining optimal conditions for transmission (i.e., peak R0 at 32.5°C) and for delimiting conditions where transmission is sustainable (i.e., R0 > 1, Fig 2B and 2C). However, these results are only valid for a given value of the ratio of new adult mosquitoes to humans λ, which we allowed to vary within a plausible range due to the fact that it depends on a wide range of factors other than temperature. In particular, λ depends on the availability and quality of aquatic habitats for mosquitoes [49] and sociocultural factors that affect contact between people and mosquitoes [50]. Some studies have used temperature-based R0 calculations to delimit geographic ranges of other vector-borne diseases such as malaria [16], but we used R0 solely as part of an intermediate step to link the generation interval with epidemic growth rates.
Although R0 is important for quantifying threshold conditions for pathogen persistence, it is not well suited for characterizing temporal dynamics of transmission [51]. By combining temperature-dependent descriptions of R0 and the generation interval, our results offer a new way to characterize the intensity of dengue epidemics as a function of temperature. One common concern about analyses based on R0, and estimates of r based on R0, is whether they are relevant beyond the context of a novel pathogen in a fully susceptible population. Estimates of r based on the effective reproduction number, R [17], offer a more generalizable alternative to estimates of r based on R0, which is what we have considered in this study. To consider how the distinction between R0 and R might impact our results, we note the relationship R = R0S, where S is the proportion of a population that is susceptible. This linear relationship between R0 and R implies that extrapolating our results below S = 1 should result in behavior similar to how our temperature-dependent estimates of r vary with changes in λ, given that λ also affects R0 linearly. Perhaps most importantly, this reasoning implies that the temperature at which epidemic growth rates peak should be applicable across contexts in which either the susceptible fraction S or the mosquito-human ratio λ vary. Still other factors affecting R0 and r could vary across contexts—e.g., species or strain differences [52]—that could be important for some future applications.
One limitation of our approach is that the precise value of the temperature threshold for maximum r could be subject to revision as understanding of the relationships between temperature and transmission parameters improves. In previous work [15], revised assumptions about the effects of temperature on transmission parameters were shown to affect prior understanding of the relationship between temperature and R0 for malaria. Independently validating our calculations with epidemic data could be one way to address these uncertainties, but epidemic growth rates based on case reports can be difficult to compare across sites. Even if factors such as temperature are consistent across sites, still others may vary and have major impacts on epidemic growth rates, including mosquito abundance [39], population immunity [53], and reporting rates [54]. Due to these and other variations across locations, Johansson et al. [55] found no detectable association between temperature and large-scale epidemic dynamics. Our results make important progress towards being able to resolve the roles of and complex interactions among these factors in future studies.
Based on current understanding of relationships between temperature and transmission parameters, our result that r consistently peaks around 33°C (95% CI: 32.6–33.2°C) led us to examine which populations globally could remain below, newly exceed, or further surpass this temperature under future climate change scenarios. We found that most people currently living in areas at risk for DENV transmission could be subject to increased epidemic growth rates by 2050 under a range of scenarios about future temperature increases. For most DENV-endemic areas, this would have little effect on the overall burden of disease, which is already high, but it could affect transmission dynamics, making epidemics more intense. At the same time, there are a number of important caveats to bear in mind about these projections. First, transmission depends not only on temperature but also other abiotic variables, such as rainfall, in complex ways [56]. Second, the effects of abiotic variables may be outweighed by changes in human factors, such as economic development, urbanization, demography, and population immunity [57,58]. Third, long-term projections of dengue are highly variable and conflicting [59], making the long-term effects of any single change such as temperature nearly impossible to anticipate.
Although r will vary across different regions for different reasons, our finding that temperature changes under future climate change could elevate epidemic intensity of dengue in some areas suggests a categorically new way in which climate change might impact infectious disease transmission [60]. Our quantification of these effects focused on DENV, but these results also offer tentative, but plausible, estimates of how epidemics of other viruses transmitted by Ae. aegypti mosquitoes, such as chikungunya and Zika, might be impacted under future climate change. Our qualitative results apply even more broadly, implying that temperature has the potential to shape multiple aspects of vector-borne parasite life history and to influence multiple aspects of the temporal dynamics of associated diseases.
|
10.1371/journal.pntd.0004636 | Zika Virus Outbreak in Rio de Janeiro, Brazil: Clinical Characterization, Epidemiological and Virological Aspects | In 2015, Brazil was faced with the cocirculation of three arboviruses of major public health importance. The emergence of Zika virus (ZIKV) presents new challenges to both clinicians and public health authorities. Overlapping clinical features between diseases caused by ZIKV, Dengue (DENV) and Chikungunya (CHIKV) and the lack of validated serological assays for ZIKV make accurate diagnosis difficult.
The outpatient service for acute febrile illnesses in Fiocruz initiated a syndromic clinical observational study in 2007 to capture unusual presentations of DENV infections. In January 2015, an increase of cases with exanthematic disease was observed. Trained physicians evaluated the patients using a detailed case report form that included clinical assessment and laboratory investigations. The laboratory diagnostic algorithm included assays for detection of ZIKV, CHIKV and DENV. 364 suspected cases of Zika virus disease were identified based on clinical criteria between January and July 2015. Of these, 262 (71.9%) were tested and 119 (45.4%) were confirmed by the detection of ZIKV RNA. All of the samples with sequence information available clustered within the Asian genotype.
This is the first report of a ZIKV outbreak in the state of Rio de Janeiro, based on a large number of suspected (n = 364) and laboratory confirmed cases (n = 119). We were able to demonstrate that ZIKV was circulating in Rio de Janeiro as early as January 2015. The peak of the outbreak was documented in May/June 2015. More than half of the patients reported headache, arthralgia, myalgia, non-purulent conjunctivitis, and lower back pain, consistent with the case definition of suspected ZIKV disease issued by the Pan American Health Organization (PAHO). However, fever, when present, was low-intensity and short-termed. In our opinion, pruritus, the second most common clinical sign presented by the confirmed cases, should be added to the PAHO case definition, while fever could be given less emphasis. The emergence of ZIKV as a new pathogen for Brazil in 2015 underscores the need for clinical vigilance and strong epidemiological and laboratory surveillance.
| Zika virus (ZIKV) has been identified in 2015 in Brazil for the first time, causing outbreaks of an illness characterized by skin rash and absent or low grade and short-termed fever. It is difficult to distinguish ZIKV from Dengue (DENV) or (CHIKV) based on the acute clinical presentation. The virus is closely related to DENV, and therefore antibody tests also have problems distinguishing between the two viruses due to cross-reactivity. Recent findings suggest that in a minority of ZIKV cases neurological disease can develop, and that babies born from mothers reporting a ZIKV-like illness during pregnancy may suffer from congenital abnormalities, in many cases a small head or brain. Here we report about an outbreak of ZIKV disease in Rio de Janeiro in the first half of the year 2015, which reached its peak in May/June 2015. This is the first published description of a ZIKV outbreak in Latin America. It is interesting to note that confirmed cases appeared as early as January 2015. Cases were confirmed based on the detection of the viral genome in the blood of the patients. The clinical characterization of the confirmed cases and unconfirmed cases proved to be very similar. Itching or itching rash has been suggested to be added to the case definition issued by the Pan American Health Organization (PAHO).
| Since 2015, Brazil is faced with the challenge of three co-circulating arboviruses of major public health importance. For the last 30 years, dengue virus (DENV) infection has been the main mosquito-transmitted threat in the country, which has suffered several epidemics caused by all four serotypes, [1, 2] fostered by the widespread presence of its main vector, Aedes aegypti, in densely populated urban areas. [3–5] In spite of all efforts, a sustained reduction of the Ae aegypti population has not been achieved.
The emergence of Chikungunya virus (CHIKV) and Zika virus (ZIKV) in Brazil [6, 7] poses new challenges to clinicians and public health authorities due to overlapping clinical features and the fact that validated serological assays for ZIKV that can reliably distinguish between acute disease and past exposure are currently not available. Most importantly, ZIKV infections are suspected to be associated with congenital abnormalities [8] and with neurological complications such as Guillan-Barré syndrome (GBS) [9], while CHIKV infections have been associated with persisting arthralgia [10].
Both ZIKV as well as DENV are members of the family Flaviviridae, whereas CHIKV is an alphavirus of the Togaviridae family. ZIKV was first described in humans in Africa in 1952 [11] and has recently caused epidemics in the Pacific region. [12–14] In May 2015, ZIKV was identified in Northeast Brazil associated with an outbreak of acute exanthematous disease in the state of Bahia, [6, 15] followed by several other locations, including the state of Rio de Janeiro. [6, 15–17] ZIKV was considered to cause a benign infection, leading to a self-limited disease consisting of rash, fever (often of low intensity and short-termed or even absent), and mild arthralgia and therefore did not receive much scientific attention until recently [14].
However, new findings suggest an association of recent ZIKV disease with GBS amongst adults[18] and with congenital abnormalities (e.g. microcephaly and ocular lesions in neonates) born from women reporting a ZIKV-like disease during pregnancy. [19–21]
Here we describe the clinical, epidemiological, and virological features of patients with acute exanthematous disease at a tertiary reference centre in Rio de Janeiro during an outbreak of ZIVK in the first half of 2015. We analyze the temporal and spatial occurrence of suspected ZIKV patients over the time of the outbreak and compare the profile of clinical signs and symptoms in PCR-confirmed ZIKV patients with the current PAHO case definition for ZIKV disease.
The study was conducted at the Laboratorio de Pesquisa Clínica de Doenças Febris Agudas of the Instituto Nacional de Infectologia Evandro Chagas (INI), which is part of the regional Dengue-Research Center of Fundação Oswaldo Cruz (FIOCRUZ), a reference center of the dengue network in Rio de Janeiro, Brazil. Patients with acute febrile disease seen at this outpatient clinic are either referred from other health units in Rio de Janeiro, or are spontaneously seeking care as they live in the nearby district of ‘Manguinhos’—an impoverished neighborhood close to the FIOCRUZ research center with a human development index of 0.726, a population of 36,160 persons in an area of 261 square kilometres, and an estimated number of 10,816 households.
Since January 2015 physicians at INI/FIOCRUZ observed an increase of cases characterized by rash, with either absent or low-grade and short-termed fever, and sometimes associated with arthralgia and/or conjunctivitis. The disease was clinically distinct from DENV which led to a systematic syndromic investigation utilizing a specific laboratory algorithm, which included diagnostics assays for detection of DENV, CHIKV, and ZIKV (after the reports of ZIKV transmission in the Northeast of Brazil[22]). The investigations were performed as part of the ongoing study on “Detection of unusual clinical presentations of dengue”, which was reviewed and approved by the local Ethics Committee (CAAE 0026.0.009.000–07). Written informed consent was obtained from all patients or their legal representatives.
For the purpose of the analysis presented here we concentrate on the first half of the year 2015. After the identification of ZIKV infection in a patient with no travel history outside Rio de Janeiro in May 2015[16], additional prospective surveillance at the outpatients' clinic was initiated in order to specifically identify patients with suspected ZIKV disease. Patients with acute onset of generalized macular or papular rash were considered suspected cases of ZIKV. We interviewed patients using a standard case report form (CRF) to collect information about demographic features, clinical signs and symptoms, and the severity of the disease (S1 File). Blood samples were collected during the acute phase (i.e., within 7 days after the onset of symptoms) and during the convalescent phase. Samples were stored at -70°C and analyzed in the Flavivirus Diagnostic and Reference Laboratory of Fiocruz.
Acute phase serum samples were tested by qRT-PCR for ZIKV [23] and RT-PCR for DENV RNA [24]. CHIKV qRT-PCR testing was performed for a random sample of 25% of patients as no patient tested positive and there was no on-going transmission of CHIKV in Rio at the time of this stud1y. Cases were classified as ZIKV infection if ZIKV RNA was detected in the serum.
Phylogenetic analysis of nucleic acid sequences derived from 10 random ZIKV positive samples (out of 119) was performed using 327 base pairs of the envelope protein (GeneBank accession number KT381874). The tree was inferred using the maximum likelihood algorithm based on the Tamura 3-parameter model as implemented in MEGA 6. The numbers shown to the left of the nodes represent bootstrap support values > 70 (1,000 replicates). The tree was rooted with Spondoweni virus.
The home addresses of the cases were georeferenced using Google maps. Statistical analyses and maps were performed using software R version 3.2.2.libraries RCurl, RJSONIO, ggmap and ggplot2. [25]
Between 1st of January and 31st of July 2015, 364 suspected cases of ZIKV disease were identified based on the presence of acute onset rash with or without fever. Of these, 262 (71.9%) were tested and 119 (45.4%) were confirmed by the detection of ZIKV RNA through qRT-PCR assay. All of the samples with sequence information available clustered within the Asian genotype (Fig 1). RT-PCR assays using consensus primers for nucleic acid of other arboviruses, including DENV and CHIKV, were negative in these patients. No DENV RNA was detected in any of the 143 acute phase serum samples that tested negative for ZIKV.
Suspected ZIKV cases started to appear in January 2015, peaking in May (Fig 2). Their geographical distribution is shown in Fig 3. In spite of the clustering in the surrounding neighborhood ‘Manguinhos’, cases residing in other neighborhoods in the Metropolitan area of Rio de Janeiro were also included (Fig 3).
Baseline characteristics of confirmed (tested ZIKV positive) and unconfirmed (tested ZIKV negative) patients are presented in Table 1. The median age of these patients was 37 years (range 9 to 60). A majority of the cases were females (158/262–60.3%), and among these 73% were aged 15–49 years (116/158–73.4%). The same pattern was seen with regard to the confirmed cases only (60.5% female [72/119] and 76.4% among those in the age group of 15–49 years [55/72]). No travel histories were recorded for the confirmed cases, thus infections were acquired locally. Only 38% of the patients recalled exposure to mosquito bites. Arterial hypertension was the most frequent comorbidity recorded, followed by diabetes. Geographic clustering of cases (defined as more than one case in the household, neighborhood or work) occurred in half of all confirmed cases (62/119–51.1%).
The most commonly reported symptoms in the first four days of the disease were rash, itching, prostration, headache and arthralgia (with or without associated oedema) (Table 2). Although fever was not observed at presentation in the majority of patients (64%), 43 patients (36%) reported a history of fever not lasting longer than one day, or the occurrence of a single fever peak on the first day of disease. The median duration of rash was 5.5 days (range 3 to 7), and of arthralgia 9 days (range 2 to 21). Mild hemorrhagic symptoms (petechiae, minor mucosal bleeding) were reported in 21% of the confirmed and 11% of the unconfirmed cases, while no severe bleeding was observed. One patient was hospitalized with neurological symptoms consisting of peripheral nerve impairment, and is described in more detail elsewhere (Brasil et al, in press). No deaths or other severe complications were associated with the ZIKV disease in this series. Four women with confirmed ZIKV were pregnant. One had a miscarriage at 10 weeks of pregnancy, three weeks after the acute disease episode. The remaining three (two with ZIKV infection during the 18th week and one during the 35th week) delivered normal babies without any clinical evidence of infection or congenital abnormalities.
All patients had full blood counts performed. The median leucocyte count of confirmed ZIKV cases was 4,590 cells/mm3 (2,240 to 11,570 cells/mm3), the median platelet count was 201,000 cells/mm3 (102,000 to 463,000 cells/mm3) and the median haematocrit was 41.2% (33.2 to 50.3%).
This is the first report of a ZIKV outbreak in the state of Rio de Janeiro, based on a large number of suspected (n = 364) and laboratory confirmed cases (n = 119). We were able to demonstrate that ZIKV was circulating in Rio de Janeiro as early as January 2015. The outbreak began in the first half of 2015 with a transmission peak in May 2015. The clinical picture of the disease was characterized by the sudden increase of exanthematous disease with absent or short-termed and low-grade fever.
The early documentation as well as the detailed clinical characterization was possible through a prospective syndromic surveillance study, which included a systematic collection of blood samples for the detection of unusual forms of dengue. In fact, 11% of the confirmed cases date back to the months before ZIKV transmission was reported in Northeast Brazil, in May (Fig 2).
The detection of ZIKV RNA in the serum of acutely ill patients and the absence of nucleic acid of other arboviruses provide convincing evidence that the outbreak was caused by ZIKV. Although DENV is common in Rio de Janeiro, none of the 252 patients for whom acute-phase specimens were available, tested positive for DENV, which is surprising and possibly highlights explosive transmission dynamics of ZIKV. Patients were also tested for other pathogens at the attending physician’s discretion, but no other causes were found (i.e. CHIKV, rubella, cytomegalovirus, EBV, toxoplasmosis). Based on this study, there is no evidence that ZIKV was circulating in Rio de Janeiro before January 2015. However, a larger and more representative sample would be necessary to make a more definitive statement about ZIKV circulation in Rio de Janeiro before 2015.
The peak of cases was observed in May (Fig 2), after the rainy season—which is in line with temporal patterns observed in previous outbreaks in Yap Island and French Polynesia. [13, 14] In May 2015, atypically high precipitation was recorded in Rio de Janeiro [26], which may have affected the vector distribution and abundance.
It is noteworthy, that the first DENV epidemic in Rio de Janeiro in 1986 started in the same geographical area from where the majority of ZIKV cases reported here originated [27]. This is likely to be associated with the high human population density, abundant Ae aegypti populations, the precarious socioeconomic status, and lack of infrastructure.
The observation that ZIKV cases clustered within households is interesting and needs to be verified in other settings. If this is a consistent feature, it could be consistent with a high vectorial capacity or hint towards other routes of transmission. The possibility of transmission via mucosal contact is supported by evidence of sexual transmission [28, 29] and by the detection of ZIKV RNA in saliva and urine. [30, 31] Further studies need to be carried out to address this question in more detail, with important implications on epidemiology and disease control of ZIKV.
The age distribution with a bias towards adults is consistent with the overall age profile of the clinic. Female patients were overrepresented with regard to males, which might be due to differences in health care–seeking behavior.
An epidemic of microcephaly in newborns in Northeast Brazil has been the focus of attention by public health authorities in Brazil and worldwide. There is emerging evidence that these congenital abnormalities may be associated with exposure to ZIKV during pregnancy. [32] Among the ZIKV-positive pregnant women followed in our study, one out of four had a miscarriage in the 10th week of pregnancy. However, no further investigations were performed in this case. In a recently published cohort of pregnant women with ZIKV infection, we have described two miscarriages amongst 72 confirmed cases [21], although the real frequency of this and other complications is yet to be determined. More recently, the state of Rio de Janeiro has reported a tenfold increase in microcephaly cases in 2015 compared to previous years [32] and, due to the long-lasting medical and economic consequences, it is of high importance that multidisciplinary studies are conducted to determine the incidence and the phenotypic variability of congenital abnormalities related to ZIKV infection during pregnancy.
The clinical signs and symptoms of ZIKV observed in Rio de Janeiro have many similarities with those described in previous reports and case series. [14] The similarity and frequency of clinical manifestations between confirmed and unconfirmed cases supports the notion that the majority of suspect cases were suffering from ZIKV disease, which could be due to current limits of detecting low quantities of RNA within a potentially short viremic period and the lack of a reliable serological test to ascertain recent exposure and distinguish it from DENV infections.
As in DENV and CHIKV infections, rash was a common feature in ZIKV. However, pruritus/itching was a prominent characteristic of the maculo-papular rash in the majority of confirmed cases (79%) from the onset of symptoms, which can potentially help in distinguishing ZIKV from other arboviral infections. Enlarged lymph nodes, especially in the cervical and retro-auricular regions, were also found frequently, which led clinicians to consider rubella as an important differential diagnosis. More than half of the patients reported headache, arthralgia, myalgia, non-purulent conjunctivitis, and lower back pain, consistent with the case definition of suspected ZIKV issued by PAHO. [8] However, fever was absent in most cases. In our opinion, pruritus, the second most common clinical sign presented by the confirmed cases, should be added to the PAHO case definition.
DENV, which has been circulating in Brazil for many years, was considered an unlikely diagnosis in many patients because fever was only observed in a minority (36%). When present, fever was usually short-termed and the temperature relatively low. From a clinical management perspective, it is important to differentiate ZIKV both from CHIKV and especially from DENV due to the need to monitor cases for possible evolution of severe and life-threatening clinical outcomes. Standard laboratory values (e.g. hematology or biochemistry) did not show a distinct pattern in the confirmed ZIKV patients and are unlikely to be helpful. The leucocyte count in ZIKV patients was in many cases moderately decreased as in other viral infections (< 5,000 cells mm3) and both platelet counts and haematocrit were within the normal range.
In this study, only one patient was hospitalized with febrile illness and a neurological manifestation developed afterwards (manuscript submitted). An increase in cases with Gullain-Barré syndrome (GBS) was reported during the epidemic In French Polynesia in 2013/14 [12], as well as in the Northeast of Brazil during the current outbreak. [6] It remains to be seen if ZIKV is causally related to higher rates of neurological complications.
Our data suggests that in a previously naïve adult population, ZIKV causes a self-limited and mostly benign disease characterized by a pruritic maculopapular rash and absent or low-grade, short-termed fever. ZIKV-like disease was not a notifiable condition in Rio de Janeiro until October 22th, 2015. This partially explains why our data differs from the official figures, [22] as the syndromic surveillance study that was implemented in 2007 (long before the ZIKV outbreak started) is more likely to detect changes in the clinical patterns of dengue compared to the routine health system. Similar to the dynamic in the Northeast of Brazil, a considerable increase in the number of microcephaly in neonates was registered in Rio de Janeiro during 2015. As of January 2016, 122 microcephaly cases were documented—compared to the average of 10 cases per year registered in the previous years. [32] This increase could potentially be attributed to the first wave of ZIKV transmission observed between January and July 2015 in Rio de Janeiro. In Pernambuco, in the Northeast region of Brazil, 1306 cases of microcephaly were reported as of the second epidemiological week of 2016 [32] which are attributed to the ZIKV outbreak in early spring 2015.
At this point, health services must be alerted to the potential for an even larger epidemic during the summer of 2015–2016 spreading to additional locations and affecting the susceptible proportion of the population that was not exposed during the last transmission season.
The emergence of ZIKV as a new pathogen for Brazil in 2015 underscores the ease with which pathogens travel between continents and the need for clinical vigilance and strong epidemiological and laboratory surveillance systems. The phylogenetic analysis somehow is in line with the speculative hypothesis that ZIKV was possibly introduced to Rio de Janeiro during the VI World Sprint Championship canoe race in August 2014, which included teams from four Pacific countries (French Polynesia, New Caledonia, Cook Islands, and Easter Island) where the virus circulated during 2014. [33]
In 2016, Rio de Janeiro will be hosting the Summer Olympics and Paralympics games, which will attract a high number of national and international visitors. The threat caused by ZIKV has far reaching implications on tourism and industry.[34] Reliable and sensitive surveillance of arboviral disease that includes a system for the detection of emerging pathogens is of paramount importance to manage the complex challenges ahead. Our findings have demonstrated that ZIKV was circulating in Rio de Janeiro at least five months before its detection was announced by the health authorities, which must be taken into consideration for future design and implementation of effective syndromic surveillance systems.
|
10.1371/journal.pbio.2002930 | Cell signaling heterogeneity is modulated by both cell-intrinsic and -extrinsic mechanisms: An integrated approach to understanding targeted therapy | During the last decade, our understanding of cancer cell signaling networks has significantly improved, leading to the development of various targeted therapies that have elicited profound but, unfortunately, short-lived responses. This is, in part, due to the fact that these targeted therapies ignore context and average out heterogeneity. Here, we present a mathematical framework that addresses the impact of signaling heterogeneity on targeted therapy outcomes. We employ a simplified oncogenic rat sarcoma (RAS)-driven mitogen-activated protein kinase (MAPK) and phosphoinositide 3-kinase-protein kinase B (PI3K-AKT) signaling pathway in lung cancer as an experimental model system and develop a network model of the pathway. We measure how inhibition of the pathway modulates protein phosphorylation as well as cell viability under different microenvironmental conditions. Training the model on this data using Monte Carlo simulation results in a suite of in silico cells whose relative protein activities and cell viability match experimental observation. The calibrated model predicts distributional responses to kinase inhibitors and suggests drug resistance mechanisms that can be exploited in drug combination strategies. The suggested combination strategies are validated using in vitro experimental data. The validated in silico cells are further interrogated through an unsupervised clustering analysis and then integrated into a mathematical model of tumor growth in a homogeneous and resource-limited microenvironment. We assess posttreatment heterogeneity and predict vast differences across treatments with similar efficacy, further emphasizing that heterogeneity should modulate treatment strategies. The signaling model is also integrated into a hybrid cellular automata (HCA) model of tumor growth in a spatially heterogeneous microenvironment. As a proof of concept, we simulate tumor responses to targeted therapies in a spatially segregated tissue structure containing tumor and stroma (derived from patient tissue) and predict complex cell signaling responses that suggest a novel combination treatment strategy.
| A signaling pathway is a network of molecules in a cell that is typically initiated by stimuli (e.g., microenvironmental cues) acting on receptors and internal signaling molecules to determine cell fate. Signaling pathways in cancer cells are different from those in normal cells, and this difference helps cancer cells to grow and thrive indefinitely. Drugs that target the aberrant signaling pathways in cancer cells (often referred to as targeted therapy) are promising for improving treatment outcomes of many different cancers in patients. However, most patients eventually develop resistance to these drugs. Resistance may already be present in the tumor or may emerge via mutation or via microenvironmental mediation. Tumor heterogeneity, which is characterized by subtle or dramatic differences among tumor cells, plays a key role in the development of drug resistance. Some tumor cells respond well to therapy, while others may adapt to the stress induced by the drug within the microenvironment. Moreover, removal of drug-sensitive cells may result in the competitive release of drug-resistant cells. Here, we present mathematical models to assess the impact of heterogeneity in signaling pathways within tumor cells on the outcomes of targeted therapy. We consider a simplified version of two well-known signaling pathways that modulate the growth of lung cancer cells. By using different targeted therapies, we quantify the effect of pathway inhibition on protein activity and cell viability and developed a mathematical model of the network, which is trained to reproduce these data and to develop a panel of heterogeneous in silico cells. The model predicts potential mechanisms of drug resistance and proposes combination therapies that are effective across the panel. We validate these combination therapies experimentally using the lung cancer cells and integrated the in silico cells into a computational lung tissue model that explicitly captures the microenvironment of lung cancer. Our results suggest that heterogeneity in both the tumor and microenvironment impacts treatment response in different ways and suggest a novel combination therapy for a better response.
| Normal cell signaling is significantly altered in cancer as a result of genetic and epigenetic changes, facilitating uncontrolled proliferation and cell survival [1, 2]. Targeted therapies directly exploit these alterations by blocking the activity of specific proteins typically mutated or abnormally up-regulated [3]. These therapies have elicited dramatic success in controlling the growth of multiple cancers [4–10] but showed little to moderate impact on others [11, 12]. Drug resistance, however, remains a major problem due to both cancer cell–intrinsic (innate and acquired) resistance mechanisms [13] and microenvironment-mediated resistance [14–16].
Tumor heterogeneity is known to contribute to drug resistance [17, 18]. Cancer cells within a tumor exhibit differential genetic and phenotypic characteristics [19]. Genomic heterogeneity leads to cell-to-cell variability in protein expression and activity as genes drive the production of proteins. Protein activity is variable even in genetically identical cancer cell populations in the same microenvironment [20–22]. This cell-to-cell variability arises from intrinsic stochastic fluctuations [23–30] and variation in microenvironmental conditions that affect the protein-signaling network. This variation can affect sensitivity to stimuli, contribute to cell phenotype decisions, and cause clonal cells to differently respond to stimulus and targeted therapies (e.g., erlotinib) [31, 32].
Understanding how cancer cell signaling variation affects targeted therapy outcomes is challenging. Cancer-driven signaling proteins do not function in isolation but rather function in protein complexes that belong to large and complex signaling networks that govern key phenotypic processes such as proliferation, apoptosis, and response to targeted therapy [33, 34]. Furthermore, the cancer signaling network and drug response are modulated by microenvironmental factors [35, 36]. Therefore, experimental data obtained from simplified cell-based experiments in single uniform environments will have limited ability to tease apart the impact of signaling network and microenvironmental variation on targeted therapy outcomes. A number of previous studies have used mathematical models to understand complex signaling networks and improve treatment strategies. Various modeling approaches were developed (reviewed in [37]). Boolean or probabilistic Boolean models were developed to analyze cancer signal pathways and predict treatment outcomes [38–41]. A logical modeling approach was employed to understand various cell signaling pathways [42–47]. An artificial neural network approach was used to map between microenvironments, pathways, and phenotypes [48]. Detailed kinetic models of cell signaling pathways have been studied using systems of ordinary differential equations (ODEs) [49–57]. Most studies, however, ignore signaling heterogeneity or extrinsic variation in microenvironmental cues that will differentially stimulate the signaling network.
To investigate the effects of signaling heterogeneity on targeted therapy outcomes, we develop an integrated approach combining in vitro experiments with three different mathematical models, an intracellular signaling model, a cancer cell population growth model, and a hybrid cellular automata (HCA) model of tumor and stroma (see Fig 1 for an overview).
We consider a simplified signaling network composed of interactions between key proteins in an oncogenic rat sarcoma (RAS)-driven mitogen-activated protein kinase (MAPK) and phosphoinositide 3-kinase (PI3K)–protein kinase B (AKT, PKB) pathway (Fig 2A). This pathway has been studied extensively, and the positive or negative feedback regulations between proteins in the pathway are known [58–62]. Of note, mammalian cells express three RAS gene family members (HRAS, NRAS, and KRAS), and our model is based on empirical data obtained from a lung cancer cell line (A549); KRAS is always activated by a point mutation (Gly12Ser), whereas the other two RAS proteins (HRAS and NRAS) are wild type. Recent studies reported the importance of crosstalk between wild-type and mutant RAS proteins in cancers driven by oncogenic mutant RAS [59, 63, 64]. Therefore, we consider two different types of RAS, wild type (RAS_w) and mutant type (RAS_m). The network node connectivity is based on prior pathway information between signaling proteins [58–62]. Most interactions are feed-forward and positive (Fig 2A, green lines) except the one negative feedback regulation of epidermal growth factor receptor (EGFR) by extracellular receptor kinase (ERK) (Fig 2A, red line). While network connectivity is assumed fixed in the model, the strength of interactions is variable and is modeled using a weight matrix (W). Each element in the network (node xi) is updated by solving the following equation,
dxidt=T(∑j∈N(i)Wijxj)−αxi,xi(0)=0,i=3,4,…,n,T(z)=εtanh(βz),
(1)
where x1,2 correspond to two inputs (growth factor and hepatocyte growth factor [HGF]), and χ3,4…,n correspond to the relative change of protein activities or cell viability due to an inhibitor with respect to untreated conditions (log2 (treated/untreated)) to be consistent with experimental data. Of note, all of the experimental measures in our study are relative values, normalized to the unperturbed condition. The absolute concentration or activity of signaling proteins as well as cell viability are difficult to acquire from experiments performed in our study, namely western blot experiments and cell viability measurement assays. Therefore, the weights in the model represent relative abundance or protein activities in treated conditions compared to treatment-naïve conditions. The model assumes that the rate of change of a variable is determined by the linear combination of neighboring nodes with corresponding weights. This additive linear function has successfully described protein reaction networks [54, 55, 65] although other functions such as Michaelis-Menten kinetics are viable options [51]. In the experiments we carried out, the microenvironmental conditions are growth factor and HGF. The growth factor (model variable x1) is always present, while HGF (model variable x2) is present in only some of the experimental conditions. In particular, the HGF is not present in the control condition. All experimental results are normalized to this control (no-HGF) condition. To represent these experimental conditions in the model, the input value x1 is set to a nonzero constant value (e.g., x1 ≡ C), while the variable x2 is set to be 0 or nonzero (e.g., x2 ≡ C) if HGF is present. We chose to set the parameter constant C to be 10. N(i) represents the neighborhood of a protein node i (a set of nodes connected to the node i), and α indicates a tendency to return to the untreated state. The transfer function T accounts for saturation effects, and the constants ε and β modulate amplitude and slope. In the model, we set ε to be 4.5 and β to be 0.5 to model a smooth sigmoidal behavior.
Drug inhibition is modeled by knocking out a corresponding protein activity. For example, if a drug inhibits a protein xi, then the variable xi is set to be a very small number (xi=log2(xtreatedxuntreated)=−μ,μ≫1). It is worth noting that we modeled the effect of drugs tyrosine-protein kinase Met inhibitor (METi), EGFRi, and AKTi by inhibiting pMET, pEGFR, and pAKT activities, respectively. Two other drugs, mitogen-activated protein kinase kinase (MEKi), and ERKi, are modeled to inhibit pERK and pRSK, respectively (Fig 2B).
We combine the pathway model to a logistic growth ODE model to simulate the growth of in silico cells as a well-mixed population in a resource-limited environment. The model is defined as follows,
dyidt=ri(1−∑j=1NyjK)yi,i=1,2,…N,
(2)
where yi represents the number of in silico cell i, ri is the cell intrinsic growth rate of the cell i, N is the total number of cell types, and K is the carrying capacity (set to be 1 billion). To model influence of the signaling pathway on cell population growth, we formulate a cell population growth rate, ri, as a function of cell viability solutions and the treatment-naïve growth rate (ri ≡ f (θi, r0); θi: cell viability solution in linear scale not in log2 scale; r0: growth rate in an untreated control condition). Specifically, the cell viability for each cell type is obtained by fitting pathway model to our experimental data, as described in the Results section. The cell viability of cell type i (θi) represents the number of cell type i that survived after being given therapy relative to an untreated control condition (i.e., θi = Pi(t)/P0(t), θi: cell viability, Pi,0(t): number of cell type i at time t in a treated and untreated condition, respectively). To obtain a functional form for the growth rate, we make the following assumptions. We assume that a cell population initially grows exponentially (Pi(t)=Pi(0)erit,P0(t)=P0(0)er0t,t<T for some time T). We also assume that the number of initial cell population is the same in a treated condition and in an untreated control condition (Pi(0) = P0(0)). Then, cell viability is presented as a function of growth rates and time (θi=Pi(t)/P0(t)=erit/er0t). Solving the function for growth rate ri, we obtain a functional form for the growth rate. We use the doubling time of A549 cells (22 hours from [66]) to obtain the treatment-naïve growth rate (r0 = 0.76 per day). We use our cell viability assay experimental time point (t = 3 days, described in Results section). Now, we have a constant growth rate of cell type i for each treatment condition (500 cells x 28 treatment conditions, total 14,000 growth rates, ri). Then, the ODE Eq (2) is solved to simulate a given treatment response of cell i over time. All of the in silico cells are solved simultaneously competing for limited resource (carrying capacity K).
The pathway model is integrated into an HCA model [67, 68] to simulate treatment responses in a spatially heterogeneous microenvironment. The model has the following assumptions. In the HCA, the cells are defined as points on a two-dimensional lattice that also contains continuous concentration fields of microenvironmental factors, together representing a cross-section of tumor composed of cancer cells and stroma (50 cells x 38 cells). Here, we define the tumor and stroma region explicitly based on an image segmentation of lung adenocarcinoma tissue from a patient. The tumor region contains cancer cells. Each cancer cell contains the pathway model, as developed above (Fig 2A), that links the microenvironment to cell phenotypes. The model grid can contain any number of possible microenvironmental variables. For simplicity, however, we consider only growth factors and HGF. The growth factors are assumed to be constant in the domain. We explicitly model HGF dynamics in space and time using the following partial differential equation,
∂H(x,t)∂t=D∇2H(x,t)−λH(x,t),
(3)
where H(x,t) represents the concentration of HGF at a lattice point x in tumor region and at time t. D represents the diffusion rate, and λ is a decay rate. The parameter values used in a simulation are given in the corresponding figure legend. The concentration of HGF is fixed to be a constant value (H(x,t) = γ) in the stromal region. A Neumann boundary condition (∂H∂n(x)=0, normal derivative = 0) was imposed on the domain boundary. A Dirichlet condition (H(x,t) = γ) was imposed at the interface between tumor and stroma.
The steady state solution of Eq (3) is fed into one of the inputs (HGF) in the pathway model in each cell (Fig 2A). The pathway determines cell viability and controls three different phenotypes—proliferation, quiescence, and death—as defined by the rules summarized in the flowchart (S1 Fig). Each cell is allowed to execute only one phenotype per time step (day). Of note, the model considers only orthogonal neighbors (north, west, south, and east) for space to divide or move. Cells are not allowed to leave nor enter across the boundary and are thus confined within the domain. We assume that the distribution of HGF barely changes during HCA simulation. For example, if a cell divides into two daughter cells, this increased number of cells does not impact the HGF distribution because we assume the consumption of HGF by cells negligible and that cells do not produce HGF. Because HGF consumption is minimal and the HGF diffusion timescale (approximate seconds) is a lot shorter than cell division time scale (approximate day) and the model domain size is small (50 cells x 38 cells), any consumption would quickly be equilibrated to the steady state.
A549 lung adenocarcinoma cell line was maintained in RPMI 1640 medium supplemented with 10% FBS. Cells were confirmed to be free of mycoplasma using PlasmoTest (Invivogen, San Diego, CA).
Cells were washed with ice-cold PBS, and whole cell extracts were prepared using lysis buffer (0.5% NP-40, 50 mM Tris-Cl, pH 8.0, 150 mM NaCl, 1 mM EDTA) supplemented with protease inhibitor (Roche, Mannheim, Germany) and phosphatase inhibitor cocktail (Sigma-Aldrich, Carlsbad, CA). Whole-cell extracts were resolved on SDS-PAGE and transferred to nitrocellulose membrane. The membrane was blocked in 5% skim milk/PBST and then incubated in primary antibody at 4°C overnight. Bound antibodies were visualized by horseradish peroxidase-conjugated secondary antibodies and SuperSignal West Pico Chemiluminescent Substrate (Thermo Scientific, Waltham, MA). Primary antibodies used for our study were purchased from Cell Signaling Technology (Danvers, MA) (except for β-actin, which was from Sigma-Aldrich, St. Louis, MO).
Cells were plated on 96-well plate at 2,000 cells per well and then exposed to drugs for 72 hours. Cell viability was analyzed by CellTiter-Glo (Promega, Madison, WI) according to the manufacturer’s recommendations.
Proximity ligation assays were performed as described using Duolink Far Red kit (Sigma-Aldrich, Carlsbad, CA) with antibodies to the following: EGFR (clone B38, Cell Signaling, Danvers, MA), GRB2 (clone 81, BD, San Jose, CA), and AF488-conjugated pan cytokeratin (clone Ae1/Ae3, eBioSciences, San Diego, CA). Tissue specimen was from a de-identified patient treated at Moffitt Cancer Center and was obtained via an institutionally approved protocol.
We followed the steps described in Fig 1 to investigate the effects of signaling heterogeneity on targeted therapy outcomes. First, we develop an intracellular signaling pathway model. We construct a simplified cancer signaling pathway based on prior information about the pathway, and we experimentally perturb the pathway using various kinase inhibitors in two different microenvironmental conditions (Fig 1 Design & Experiments). Among key microenvironmental factors, HGF has been shown to contribute to resistance in multiple Food and Drug Administration (FDA)-approved targeted therapy drugs [35, 36]. Therefore, we consider HGF as an additional microenvironmental stimulus. Then, we build a mathematical model of the cancer signaling pathway to predict both signaling and a phenotypic response (cell viability change due to a given therapy) to different inhibitors that target the pathway (Fig 1 prior pathway information). It is important to note that the model includes two microenvironmental factors (i.e., growth factor or HGF) and cell viability (Fig 1 prior pathway information) in addition to intracellular proteins (Fig 1 prior pathway information). Model parameters are calibrated using experimentally measured protein expression levels and cell viability after different inhibitors are applied under different microenvironmental conditions (Fig 1 Pathway Modeling, model calibration). Now using this panel of in silico cells with calibrated signaling networks, we predict distributional responses to different targeted therapies, reveal possible mechanisms of drug response and resistance, and propose combination therapy strategies that could deal with heterogeneity. The model predictions are then tested experimentally.
Next, we develop a logistic cancer cell population growth model to describe tumor growth in a homogeneous but resource-limited microenvironment (Fig 1 Multiscale Mathematical Modeling). The intrinsic growth rate of each cancer cell is estimated based on the treatment-naïve growth rate and cell viability obtained from the signaling pathway model calibration. We predict post-treatment cell population heterogeneity and average efficacy after continuous application of various inhibitors (both mono and combination therapies). Finally, we develop an HCA model to investigate the effects of spatially heterogeneous microenvironments on targeted therapy outcomes (Fig 1 Multiscale Mathematical Modeling). The model couples continuous microenvironmental factors with a discrete cell–based model. Each individual cell contains the trained network model that links microenvironment to its phenotype, which determines cell fate in a given condition. As a proof of concept, we simulate the response to an inhibitor in a section of tissue composed of both tumor cells and stroma. The model predicts complex cancer cell signaling responses and treatment outcomes, driven by both cell-intrinsic and -extrinsic mechanisms.
Kirsten rat sarcoma (KRAS)-driven cancer treatment is an important clinical need that remains largely unmet due to limited targeted drug efficacy of key downstream effectors, including MAPK and PI3K-AKT pathways. We therefore choose the KRAS mutant non–small-cell lung cancer (NSCLC) cell line (A549 cell line) as our experimental model system. Using our simplified oncogenic KRAS signaling pathway (Fig 2A, Mathematical modeling section for an explanation of how we obtained this simplified pathway) as a guide, we pharmacologically inhibited individual proteins in the MAPK pathway (MET, EGFR, MEK, ERK inhibitors) and PI3K-AKT (AKT inhibitor) pathway in A549 cells in the both absence and presence of HGF. Drug-induced changes in the phosphorylation of pathway proteins, surrogates for protein activity, were measured by western blotting (Fig 2B). Cell viability was also assessed after 72 hours of drug treatment (Fig 2C). These experimental data were quantified using ImageJ (Fig 2D), and all data were normalized to the control experimental condition (treatment-naïve condition).
The quantified changes (Fig 2D) and the Monte Carlo simulation were then employed as an optimization procedure to estimate model parameters (weights, Wij) that minimize our cost function. A network with lower cost represents the experimental data more accurately. The cost function is defined as follows,
C(W)=∑in∑dM(xi,d¯−yd)2+∑j∈N(r)χ1+exp(ηwjr)
(4)
where xi,d¯ is the steady state activity of protein or cell viability xi in treatment condition d, yd represents experimental data, and M is the total number of treatment conditions. The weight wjr indicates the weight between RAS_m (r) and its neighbors (N(r)), and the constants (χ, η) modulate the magnitude of the penalty. The first term explains the difference between model prediction and experimental data for a network W. The second term is directed at the activating RAS mutation and incorporates a penalty for estimated weights from the RAS_m node (mutant RAS) that are too small. We included this penalty because our model is based on empirical data of a KRAS mutant cancer cell line (A549 cell), where the resulting KRAS protein is constitutively active. We aimed to capture this activating mutation by penalizing small weights from RAS_m to its neighbors.
We used the following method to implement Monte Carlo simulations:
The model calibration resulted in more than 5,000 weight matrices that fit to the experimental data. We selected the best 500 (top 10%) weight matrices and used these to define our 500 in silico cells. The distributions of in silico cells are presented as box plots in Fig 2E along with the experimental measures. Errors (root-mean-squared-error [RMSE] formula given in S1 Text) are in the range of (0.03–0.56, except ERK: 0.96). The fit of ERK was poor because of unexpected inhibition of pERK by the drug (ERK inhibitor, SCH772984) [69]. The trained networks (weights) are quite heterogeneous (S1 Table). The distribution for each weight is different (S1 Table, skewed, normal, bimodal distributions, with a range of heterogeneity [Shannon] index values).
The weights here may represent relative protein abundance or protein-binding activity. There is ample evidence for differential abundance of protein species across cellular populations. An excellent example was recently published showing that variations in adaptor protein abundance are a major source of regulation of the EGFR-MAPK pathway [70]. There are several examples of differential binding activity of proteins in cell signal transduction. It is well established that adaptors such as GRB2, SHC1, and GAB1 can be recruited to receptor tyrosine kinases (RTKs) either directly or indirectly. Therefore, stochastic variation in multiprotein complex composition at individual receptors exists, and this will vary both within and between cells. It is also accepted that activation-induced receptor degradation and phosphatase activity will affect not only RTK adaptor interactions but also downstream signaling molecules such as rapidly accelerated fibrosarcoma (RAF), MEK, and ERK. Additionally, we have previously shown that, in EGFR mutant cell lines, only a fraction of the receptor is phosphorylated, and cell lines harboring the same oncogenic mutation have different levels of phosphorylated Tyrosine (Tyr) and Serine/threonine (Ser/Thr) residues [71]. Collectively, these examples indicate that protein–protein interactions in response to growth factors are not simply on–off states and that multiple factors independent of protein abundance control final signaling output.
We simulated responses of in silico cells to seven different inhibitors (EGFRi, METi, RAS_mi, AKTi, RAFi, MEKi, and ERKi). Of note, cell viability in untreated conditions is set to be a single constant value (cell viabilityuntreated ≡ 1). It is also worth noting that RAS_mi is assumed to inhibit only RAS_m (RAS mutant), not RAS_w. Distributions of relative cell viability (log2 scale, log2 (treated/untreated)) of all in silico cells are presented in Fig 3A. Similar to the experimental results (Fig 2B–2D), MEKi and ERKi reduced average cell viability significantly, whereas mean effects of EGFRi, METi, and AKTi are marginal. These results reveal quite heterogeneous responses to drugs, which could be assessed by experimental approaches [72, 73]. For example, the distribution of EGFRi treatment is bimodal due to a bimodal distribution of trained MEK-ERK weights (S1 Table and S2 Fig), suggesting the presence of a subpopulation that responds significantly differently to drug from the rest of the population (S2 Fig, green versus pink).
With our calibrated in silico cell lines, we next examine which drug combinations significantly reduce cell viability and which show marginal effects. For example, what should be cotargeted with AKTi to decrease cell viability significantly? The model predicts that activation of alternative pathways under therapy may provide an escape route to therapy. For example, AKT inhibitor induced increased activity of ERK and ribosomal S6 kinase (RSK) (S3 Fig). Cells with high ERK and RSK activity display resistance (cell viability >0) under the inhibitor, which implies that simultaneous inhibition of this alternative pathway would overcome resistance. We reasoned that combination targeting of proteins that are highly correlated with relative cell viability under a given treatment would be beneficial. In order to identify such protein nodes, scatter plots between predicted protein activity (phosphorylation) and predicted relative cell viability were considered (S4 Fig). Then, the Pearson’s coefficient was calculated for all cases. We observed that, under multiple treatment conditions—including inhibition of EGFR, MET, RAS_m, AKT, and RAF—ERK and RSK showed the highest correlation with relative cell viability (S4 Fig, pink box). This suggests that co-inhibition of ERK (by MEKi) or RSK (by ERKi) activity with other therapies would decrease cell viability most significantly. Further simulation revealed an additional application of either MEKi or ERKi to each—EGFRi, METi, RAS_mi, RAFi, and AKTi—significantly decreased cell viability compared to monotherapy of EGFRi, METi, RAS_mi, RAFi, and AKTi (Fig 3B). We tested some of these model predictions experimentally and validated, to some degree, the model’s predictive ability (Fig 3C).
We next systematically assessed cell viability reduction to all mono and combination therapies using an unsupervised hierarchical clustering approach to classify cell populations on the basis of their treatment response (relative cell viability). The treatments were categorized into multiple groups (Fig 4A, a tree diagram on the right end of heat-map). The combination of AKTi with MEKi is uniformly effective to all the cells (see the red asterisk [*] row, dark blue across all in silico cells, with little variation between cells). This combination (MEKi/AKTi) has previously been shown to be effective in NSCLC both in vitro and in vivo [74]. A striking variation is observed in response to the treatments of AKTi, RAS_mi, and RAS_mi/AKTi (Fig 4A, red bars in the first two groups versus dark blue to yellow bars in the rest of the groups). The first two clusters (pink and black color on the top of the heat-map) are associated with poor responses (little to no reduction of cell viability after a given therapy), while others are correlated with good treatment outcomes (significant reduction of cell viability after a given therapy).
Why are some cells sensitive to a given therapy while others are resistant to the same therapy? We hypothesize that this differential drug sensitivity, at least within the context of our model, must be attributed to differential protein activity as modulated by protein–protein interactions (i.e., the weights). To highlight possible mechanisms, we visualized the weights between protein nodes (Wij in our model) using both circular chord diagrams [75] and network diagrams with weighted edges (Fig 4, bottom panels; left: chord diagram; right: weighted network with different edge widths). In the circular diagrams, protein nodes are arranged around a circle with the weight between protein nodes connected to each other through the use of arcs. The width of each arc is determined proportionally by the weight between two protein nodes. To illustrate differences in signaling, we selected two representative in silico cells (Fig 4A and 4B), where cell a is resistant to AKTi, RAS_mi, and RAS_mi/AKTi and cell b is more sensitive to these therapies. In addition, we compared ranges of all weights of all in silico cells between clusters defined by the hierarchical clustering (S5A Fig). We observe heterogeneity of weights within each cluster and between clusters. Differences between clusters are significant for some weights such as weights of growth factor EGFR, EGFR-RAS_w, MET-RAS_w, RAS_w-RAF, and MEK-ERK (S5A Fig).
We next asked how an additional microenvironmental stimulation would modulate the responses to targeted treatments. To address this question, all in silico cells were treated in the presence of HGF, a significant stromal factor that contributes to drug resistance [35, 36]. An unsupervised hierarchical clustering on the basis of cell viability changes, from the no-HGF condition, separated the treatments into several groups (Fig 5A, clustering of treatments on the right tree diagram). The analysis also classified the in silico cells into several groups based on cell viability changes due to HGF stimulation (Fig 5A, tree diagram on the top of heat-map).
To test some of these model predictions, we considered three different combination therapies, AKTi/MEKi, EGFRi/MEKi, and EGFRi/AKTi. Of note, the model predicted that the treatments AKTi/MEKi and EGFRi/MEKi are not affected by HGF stimulation (Fig 5B, first two red bar graphs). The experimental data matched well with these predictions (Fig 5B, first two gray bar graphs). The model also predicted that the effect of combination therapy EGFRi/AKTi is significantly modulated by HGF stimulation (Fig 5B, the third red bar). This was also corroborated by experiment (Fig 5B, the third gray bar).
The suite of in silico cells is differentially affected by HGF stimulation. For example, some cells are not affected by the HGF stimulation (e.g., gray bars in EGFRi/RAS_mi in Fig 5A), while others are significantly affected by stimulation (e.g., red bars in EGFRi/RAS_mi in Fig 5A). Why are some cells affected by HGF stimulation, while others are not? To understand why this is the case, we selected two representative cells (a and b) and visualized the weights between protein nodes using both chord diagrams and weighted network diagrams (Fig 5C, bottom panels). In cell a, the influence of MET on RAS_w is relatively small (Fig 5C, thin blue chords from MET → RAS_w). In contrast to cell a, the influence of MET on RAS_w in cell b is much stronger (Fig 5C, cell b, thick blue chord from MET → RAS_w). This may explain why cell b increased its viability significantly upon HGF stimulation compared to no-HGF condition. In addition, distributions of all weights in each cluster show differential activity of proteins and its effects across all six different clusters (S5B Fig).
So far, treatment responses were assessed as if all in silico cells were treated in isolation. To address the effects of cell competition on treatment outcomes, we simulated all cells growing together under treatment with various inhibitors in a homogeneous, resource-limited microenvironment (Methods section). The entire in silico cell population responds in a similar way on some therapies (Fig 6, e.g., EGFRi, METi, RAFi, EGFRi/RAS_mi, etc.), but under other therapies (Fig 6, e.g., MEKi, ERKi, MEKi/ERKi, AKTi/RAFi, etc.)—due to differential viability—some cells became dominant. Interestingly, some treatments simultaneously selected for the same dominant in silico cell (Fig 6, color-shaded boxes). Addition of HGF to some treatments (e.g., EGFRi, RAS_mi, RAFi) significantly changes the fitness of in silico cells and thus drives a different dominant in silico cell (S6 Fig, dotted lines in no-HGF vs solid lines in HGF).
We assessed post-treatment population heterogeneity by measuring the Shannon index (H(x) = −∑ipi(x)log(pi(x)), where pi(x) is the probability of finding an in silico cell i after a given therapy). We compared the index with average cell viability change (Fig 6B). The two are linearly correlated (ρ = 0.65), implying that the less effective a treatment is in controlling the average in silico population growth, the more heterogeneous the post-treatment population would be (Fig 6B). Combination therapies not only display a better average treatment response but also a less diverse post-treatment population (Fig 6B, boxes versus circles). Among all combination therapies, those that combined either with ERKi or MEKi display much better average therapeutic responses (Fig 6B, small cell viability). These treatments not only effectively decrease average cell viability but also lead to a less diverse post-treatment population (e.g., Fig 6B, EGFRi/[ERKi or MEKi] vs EGFRi/[METi,AKTi,RAFI] in red-color circles). HGF stimulation minimally affected the linear relationship between post-treatment heterogeneity and average cell viability reduction (Fig 6C, ρ = 0.63 vs Fig 6B, ρ = 0.65). However, a few treatments did elicit significant changes in both average response and heterogeneity due to HGF stimulation (Fig 6D).
Because activated receptors require multiple protein interactions to activate downstream signaling, we have utilized proximity ligation assays that measure the functional association of RTKs and adaptor proteins. Using NSCLC patient specimens and xenograft models, we have previously identified an association of EGFR:GRB2 complexes and response to EGFR inhibition [76] and more recently identified a correlation between MET:GRB2 complexes and response to MET kinase inhibitors [77]. Using this approach, we consistently observe spatial heterogeneity in abundance of RTKs binding to adaptor complexes. The abundance is often highest at tumor regions that are adjacent to stromal regions (Fig 7A).
Motivated by this experimental observation (Fig 7A), we developed an HCA model to investigate the effects of microenvironmental heterogeneity on treatment outcomes (See Methods section and S1 Fig). A steady state configuration of HGF is considered throughout the whole simulation (See Methods section and Fig 7B). We randomly initialized in silico cells that contain calibrated signaling networks (Fig 2) in the domain to mimic a slice of tumor tissue (Fig 7C first, time step = 0; domain size: 50 cells × 38 cells). As proof of concept, we simulated RAS_m inhibition. After 180 days of the inhibition, distinct cells emerge near stroma (high HGF) in contrast to the nonstroma region (Fig 7C second, light- and dark-blue cells near stroma versus orange, red cells elsewhere). A clearer separation among the cell population emerges as therapy continues (S1 Movie). We also observed heterogeneous protein activity across the tissue (Fig 7D and S2 Movie). The signaling responses of some cells are affected by HGF (Fig 7D, e.g., high MET and RAF, black arrows near stroma), while signaling in other cells is not (Fig 7D, e.g., low activity of MET and RAF near stroma, red arrow).
The treatment selects for cells with high MET and RAF activity (phosphorylation), especially residing near the stroma (high HGF), which suggests that a therapy of METi or RAFi in addition to RAS_mi may be more effective. To test this suggestion, we simulated two sequential therapies of RAS_mi and RAFi and one concurrent therapy (Fig 7E and S3–S6 Movies). Depending on the order of the sequence (RAS_mi first versus RAFi first), different patterns of cells emerge after 400 days of treatment (Fig 7E, first versus second). Importantly, a concurrent combination of the two inhibitors was effective enough to eradicate all cells in this small region just after 30 days of treatment (S5 Movie and Fig 7E third, number of cells). The direct spatial competition of each cell within the tissue directly facilitated this result. To be more relevant to the clinical timeframe, we also simulated a shorter treatment schedule (e.g., 60 days) and observed similar cell behavior (S7 Fig).
Until now, we assumed that the initial states of the protein activities in our in silico cells were all zero. In order to examine the impact of changing this on the above treatment outcomes, we randomly seeded in silico cells in a slice of tissue (Fig 7C), and for each in silico cell, a random number was assigned to each initial protein activity. Then, we simulated RAS_m inhibitor for 30 days using our HCA model (S1 Fig) and repeated this process 100 times. The resulting configurations display some degree of heterogeneity (S8A Fig, shows representative results for three different initial configurations) due to cell–cell spatial competition as well as variable HGF modulation of the in silico cells. The HGF selects for cells whose cell viability is significantly modulated by HGF (violet to red cells near stroma). To illustrate this more accurately, we quantified the total number of surviving cells at time step 30 for each simulation (see S8B Fig for distributions) and classified them in terms of HGF modulation (S8C Fig). The treatment consistently selected for certain cells (S8C Fig), some of which are more influenced by HGF than others (S8C Fig).
We implemented an integrated mathematical and experimental approach to develop a panel of in silico cells that readily reproduced average kinase inhibitor responses in two different microenvironments (HGF and no-HGF). The mathematical model of a simplified oncogenic RAS-driven MAPK and AKT-PI3K pathway describes weighted interactions between proteins in the pathway. The weights here may represent relative protein abundance or protein activity. The calibrated model predicted heterogeneous responses to kinase inhibitors due to differential activities of proteins from the in silico cells under a given therapy condition. In addition, the model identified a combination therapy that effectively reduced cell viability across the entire in silico cell population (Fig 4, AKTi/MEKi). Critically, the effects are not modulated by HGF stimulation (Fig 5). This combination has been shown to be effective in NSCLC both in vitro and in vivo [74]. Integrating the pathway model into a two-dimensional lattice-based model allowed us to take a significant step toward modeling the multiscale behavior of cancer by bridging the signaling, cell, and multicellular scales with feedback from the microenvironment. We were also able to simulate the impact of an inhibitor on a tissue structure composed of tumor and stroma, showing complex signaling responses and selection for distinct in silico cells by the stroma (Fig 7 and S1 and S2 Movies), which highlighted a novel combination therapy (Fig 7 and S3–S6 Movies).
We are only just beginning to understand the importance of nongenetic heterogeneity [19, 78]. Much more needs to be done in teasing apart the different scales of heterogeneity (genetic, cellular, microenvironmental), how they interact and modulate one another, and how this might alter our current combination treatment strategies. Variable protein activity has been observed in previous studies of isogenic cancer cell lines, revealing that single-cell heterogeneity and protein–protein interaction strength is different [73]. A more recent study, of an isogenic cancer, showed striking variation in genetic, cellular, and phenotypic heterogeneity [79]. Our own experiments and simulations showed heterogeneous cancer cell signaling in a section of lung adenocarcinoma (Fig 7) [76, 77].
There are other modeling approaches for analysis of signaling pathways including logical, Boolean, and artificial neural network (reviewed in [37]). In particular, several recent studies developed integrated approaches of mathematical modeling with systematic perturbation experiments applying various kinase inhibitors to cancer cells. Some of these studies proposed novel combination therapies [43, 51, 54], just as we have. However, in addition to predicting average cell viability, we also consider post-treatment heterogeneity and, critically, the impact of the microenvironment. Historically, Boolean models have been used in understanding cancer cell–signaling responses. To investigate applicability of such a Boolean approach, we constructed an equivalent Boolean model of the signaling pathway (S2 Text and S9 Fig). This model predicted the treatment combinations (AKTi/MEKi and AKTi/RAFi) consistent with experimental data (Fig 3C, two out of seven different combination therapies). However, the other five combination therapies were not consistent with experimental data (yellow asterisks in S9C Fig). Taken together, these results suggest that the Boolean model (S9A Fig) is insufficient to predict combination therapies.
To obtain a tractable model, several simplifications have been made in this study. The model considered only two microenvironmental variables (HGF and a generic growth factor) on a two-dimensional square lattice, where the heterogeneous population of in silico cells mimicked a slice of lung tissue. Although three-dimensional models will describe these dynamics in a more realistic way, the key predictions—for example, the effect of stroma—will be consistent in a three-dimensional setting. Other simplifying aspects of the HCA approach included not considering tissue mechanical properties as well as constraining cell movement to orthogonal neighbors with discontinuous displacement. A potential alternative approach would be to consider off-lattice models that allow force-driven interactions that better describe mechanical aspects of tumor growth [80–83] and tumor morphology [84–86]. We are also acutely aware that the signaling network considered is only a fragment of a much larger and far more complex signaling cascade that turns external signals into phenotypic decisions. Because not all proteins in cancer cells are directly measurable due to experimental limitations, many players and intermediate proteins in the pathway are not included in our study. Therefore, an interaction between two proteins represents diverse influence of one entity on other entity in steady state, such that an entity can be a microenvironmental variable, an intracellular protein, or cell viability. We emphasize that, by definition, all models are but simplifications of reality and the true utility of a model is not that it can mimic reality but that it provides useful insight into the system. Despite this simplicity, our integrated approach provided multiple testable hypotheses for the complex KRAS NSCLC cell signaling network, proposed possible drug resistance mechanisms, and suggested better treatment strategies.
Finally, while we relied on western blots for protein activity (phosphorylation) readouts, alternative approaches such as Reverse Phase Protein Array exist [87]. The important step, however, was combining this average protein activity with prior information about network connectivity, allowing us to generate a suite of in silico cell lines that not only reproduced this average behavior but also gave insights into potential single-cell variability. This highlights a key need to improve our understanding of heterogeneous cell signaling networks: single-cell profiling. Such data would ideally include intracellular, cellular, and phenotypic profiling in multiple, uniform microenvironments. However, our results also emphasize the importance of interactions between heterogeneous cell populations and spatially structured environments. Therefore, while quantifying single-cell dynamics will provide key information about intrinsic cell heterogeneity, to fully understand how these differences impact treatment responses, we must consider how interactions that change through space and time alter this heterogeneity and thus treatment outcomes.
|
10.1371/journal.pgen.1006204 | dachshund Potentiates Hedgehog Signaling during Drosophila Retinogenesis | Proper organ patterning depends on a tight coordination between cell proliferation and differentiation. The patterning of Drosophila retina occurs both very fast and with high precision. This process is driven by the dynamic changes in signaling activity of the conserved Hedgehog (Hh) pathway, which coordinates cell fate determination, cell cycle and tissue morphogenesis. Here we show that during Drosophila retinogenesis, the retinal determination gene dachshund (dac) is not only a target of the Hh signaling pathway, but is also a modulator of its activity. Using developmental genetics techniques, we demonstrate that dac enhances Hh signaling by promoting the accumulation of the Gli transcription factor Cubitus interruptus (Ci) parallel to or downstream of fused. In the absence of dac, all Hh-mediated events associated to the morphogenetic furrow are delayed. One of the consequences is that, posterior to the furrow, dac- cells cannot activate a Roadkill-Cullin3 negative feedback loop that attenuates Hh signaling and which is necessary for retinal cells to continue normal differentiation. Therefore, dac is part of an essential positive feedback loop in the Hh pathway, guaranteeing the speed and the accuracy of Drosophila retinogenesis.
| Molecules of the Hedgehog (Hh) family are involved in the control of many developmental processes in both vertebrates and invertebrates. One of these processes is the formation of the retina in the fruitfly Drosophila. Here, Hh orchestrates a differentiation wave that allows the fast and precise differentiation of the fly retina, by controlling cell cycle, fate and morphogenesis. In this work we identify the gene dachshund (dac) as necessary to potentiate Hh signaling. In its absence, all Hh-dependent processes are delayed and retinal differentiation is severely impaired. Using genetic analysis, we find that dac, a nuclear factor that can bind DNA, is required for the stabilization of the nuclear transducer of the Hh signal, the Gli transcription factor Ci. dac expression is activated by Hh signaling and therefore is a key element in a positive feedback loop within the Hh signaling pathway that ensures a fast and robust differentiation of the retina. The vertebrate dac homologues, the DACH1 and 2 genes, are also important developmental regulators and cancer genes and a potential link between DACH genes and the Hh pathway in vertebrates awaits investigation.
| Temporal and spatial coordination between cell proliferation and differentiation is essential for proper organ patterning. A way to ensure this coordination is through the use of regulatory signaling pathways that control both processes. Among those is the Hedgehog (Hh) signaling pathway that regulates organ growth and patterning in embryos and tissue homeostasis in adults, both in vertebrates and invertebrates [1–4]. Not surprisingly, mutations in components of the Hh signaling pathway cause a number of human disorders, including congenital abnormalities and cancer [2–4].
One of the processes in which Hh signaling plays an essential role is the patterning of the retina in vertebrates and invertebrates [5–7]. In Drosophila, Hh is responsible for organizing a moving signaling wave that patterns the primordium of the fly eye during the last larval stage (L3). The processes under Hh signaling control have been extensively studied and summarized in what follows. The front of the differentiation wave is marked by a straight indentation of the eye epithelium, called morphogenetic furrow (MF), that runs across the dorsoventral axis of the eye primordium, or “eye imaginal disc” [8–10]. hh, initially expressed along the posterior margin of the eye disc [11] and later by the differentiating photoreceptors (PRs), activates the expression of the BMP2 decapentaplegic (dpp) within the MF [12]. Hh instructs undifferentiated proliferating progenitor cells to synchronously undergo mitosis (First Mitotic Wave, FMW) and then stop temporarily their cell cycle in G1 phase through Dpp, which acts long range [13–16]. At a shorter range, Hh initiates the expression of the proneural gene atonal (ato) [17–23] and stabilizes the G1 state by activating the expression of the p21/p27 Cdk inhibitor homologue dacapo (dap) [24–27]. In addition, together with Dpp, Hh induces coordinated cell shape changes responsible for MF formation [18,19,23,28–31] by promoting the apical constriction, apical-basal contraction and basal nuclei migration of cells (Fig 1A). These cellular changes are mediated, at least in part, through the contraction of the acto-myosin cytoskeleton [32,33]. Immediately behind the MF, Ato expression is restricted to evenly spaced cells, which become the ommatidial founder photoreceptors (PR8s). Then, PR8s induce neuronal differentiation of the adjacent precursor cells. Precursor cells that did not start their differentiation program immediately after the MF suffer one last round of mitosis, the Second Mitotic Wave (SMW) [10]. Therefore, Hh secreted by differentiating PR cells drives the anterior propagation of the MF and its associated differentiation wave, while regulating the SMW locally [25,34–36]. Thus, the MF coincides spatially with the onset of differentiation. Interestingly, the MF state is transient: while anterior precursor cells are recruited to enter the MF state, the newly differentiating PRs and cells at the SMW exit this “furrowed” state.
The coordinated action of Hh has been shown to rely on dynamic changes of its signaling activity. In flies, Hh signaling regulates the post-translational proteolytic processing of the Gli-family transcription factor Cubitus interruptus (Ci). Hh binding to the receptor Patched (Ptc) relieves the inhibition exerted by unbound Ptc on the transducer Smoothened (Smo), and thus promotes the activation of the Fused (Fu) kinase [1–3]. In turn, activated Fu promotes the conversion of the full-length form of Ci (CiFL) to Ci activator (CiA) form [37]. As a result, CiFL is no longer phosphorylated by Protein Kinase A (PKA) [38–40] and other kinases. Otherwise, CiFL phosphorylation leads to the generation of the Ci repressor form (CiR) through partial CiFL degradation by the F-box-containing protein E3 ubiquitin ligase complex (SCFSlim-Cullin1). The relative amount of both CiR and CiA determines the transcriptional status of Hh-target genes, such as ptc, dpp and engrailed (en) [41–44].
In the developing Drosophila eye, while low levels of Hh signaling promotes cell shape changes associated with MF formation and concomitant dpp expression, high levels are required for the re-entry of the precursor cells in the cell cycle at the SMW and for the activation of roadkill (rdx) expression [45–48]. Rdx targets CiFL to full degradation through the recruitment of the Cullin3 (Cul3)-based E3 ligase complex [45,46]. Thus, posterior to the MF, high levels of Hh signaling attenuate its own activity by Rdx:Cul3 complex, allowing retinogenesis to occur properly [45,46]. Therefore, mutations that affect MF progression could be additional components of the machinery that regulates Hh signaling intensity and dynamics. Mutations in the retinal determination gene dachshund (dac) affect MF movement, without blocking differentiation [49]. dac expression depends on Hh signaling itself [25,50]. It localizes to all nuclei straddling the CiFL-expressing domain, from the progenitor domain to the SMW, where differentiating PR cells start expressing hh (Fig 1B and 1C–1C´´). Therefore, high Dac levels coincide with the major neuronal differentiation and morphogenetic processes controlled by Hh signaling. Altogether, these results indicate that dac exhibits the traits required for being a candidate modulator of Hh signaling intensity. Here, we show that indeed dac potentiates Hh signaling in the MF by promoting CiFL accumulation and CiA activity downstream or in parallel of Fu. Our observations argue that this mechanism is absolutely required to promote proper retinogenesis by controlling the timing of MF formation, accurate specification of the founder PR cell and to trigger the Rdx-dependent negative feedback, which turns Hh signaling off posterior to the MF. Thus, Hh signaling potentiation by Dac allows the fast building up of signaling that is required for the swift processes associated with the moving retinal differentiation wave in Drosophila.
To investigate a role of dac as a candidate modulator of Hh signaling, we reexamined the consequence of removing dac on MF-associated processes. Consistent with previous observations [49], all GFP-marked dac- clones larger than 6 cells straddling this region showed a delay in MF formation (Fig 1D, S1A and S1A´ Fig; n = 53). Accordingly, the onset of PR differentiation, detected by labeling with an antibody against the neuronal marker Elav, was also retarded in these clones (Fig 1E and 1E´, S1D and S1D´ Fig). This retardation was associated with a delay in the onset of ato expression and with an aberrant spacing of Ato-positive PR8 cells (Fig 1, compare 1G and 1G´ with 1F). These defects were not specific of any ommatidial cell type: in dac- clones posterior to the MF, we detected expression of the hh-Z enhancer trap, which marks PR2-5 cells (Fig 1H and 1H´), of the PR3/4 marker Spalt (Sal) (Fig 1I) and of the PR7 marker Prospero (Pros) (Fig 1J). However, the density of ommatidia (Fig 1E and 1E´, 1H and 1H´, S1D and S1D´ Fig) and the proper number of cell types per ommatidium were affected by the loss of dac function. For instance, some ommatidia only contained one Sal-expressing cell instead of two (yellow arrow in Fig 1I).
Concomitant with the delayed MF and differentiation onset, the SMW was also retarded and became asynchronous. Posterior to the SMW, dac- clones showed persistent reentry into the cell cycle, as detected by elevated expression of the G2/M CyclinB (CycB) (S2A and S2A´ Fig), increased number of cycling cells (S2B and S2B´, S2C and S2C´ Fig), as well as maintenance of the expression of the G1/S cyclin CyclinE (CycE) and loss of dap (S2D and S2D´ and S2E Fig). Taken together, we conclude that dac is required for three essential roles played by Hh signaling: MF movement, regulation of the cell cycle and proper retinogenesis.
Downregulating the function of the Hh-signal transducer smo (smo3 clones) also caused a delay in MF movement (seen by E-Cadherin (E-Cad) higher signal intensity) and affected apical constriction of cells within the MF (S1B and S1B´ Fig). In addition, the density of ommatidia was reduced in smo- clones (S1E and S1E´ Fig). Thus, smo- and dac-mutant clones shared similar phenotypes (Fig 1D, 1E and 1E´ and S2A and S2B´, S2D and S2E´ Fig). In agreement with dac and smo being part of the same signaling pathway, dac synergized with smo in MF formation and PRs differentiation. All smo, dac double-mutant cells failed to undergo the cell shape changes associated with the MF (S1C and S1C´ Fig, n = 33 discs) and to differentiate PRs (S1F and S1F´ Fig, n = 20 discs). As dac was expressed in cells that accumulated CiFL at high levels in the MF and at low levels posterior to the MF where Ci promotes the transcription of the rdxZ reporter (Fig 1B–1C´´ and 2A–2B´´´), we next analyzed if dac affected Hh signaling. Strikingly, 67% of discs containing GFP-marked dac- clones straddling the MF showed reduced CiFL levels (Fig 2C–2F; n = 24 discs) and lower transcription of dpp, monitored by the transcriptional reporter dppZ (Fig 2G–2J; 64% of discs, n = 14). In addition, all dac- clones failed to trigger high levels of a lacZ enhancer trap insertion in the rdx gene (rdxZ) posterior to the MF (Fig 2K–2N; n = 9 discs). All these results indicate that dac is required for a full activation of the Hh signaling pathway.
To determine if Dac is sufficient to potentiate Hh signaling, we analyzed the effect of overexpressing dac on Hh signaling activity in the wing imaginal disc, where endogenous Dac protein is expressed only in a few restricted patches [49]. In this tissue, Hh produced in the posterior (P) compartment signals to the anterior (A) compartment (Fig 3A). Thus, cells along the AP boundary compartment receive maximal Hh signaling, leading to the activation of rdx expression and consequently signaling attenuation through Rdx:Cul3-mediated CiFL degradation [45,46]. At these signaling levels, immediately adjacent to the P compartment, ptc expression is induced [41–44]. Next to this domain and further away from the AP boundary dpp is expressed [41–44]. Therefore, if dac potentiates Hh signaling, we expected that its expression along the AP boundary should enhance signaling levels and allow dpp transcription. Although overexpressing dac (HA::dac) along the AP boundary compartment using a ptc-Gal4 driver (Fig 3A) promoted CiFL accumulation (Fig 3, compare 3C–3C´´ with 3B–3B´´ and 3D´ with 3D) and increased expression of dppZ (Fig 3, compare 3G and 3G´ with 3E, 3F and 3F´ and 3H´ with 3H), these effects were relatively modest. Overactivation of Hh signaling would also be expected to potentiate rdx transcription, which would limit CiFL accumulation and Hh signaling activity. In agreement with this, overexpressing HA::dac in the dorsal compartment using the apterous-Gal4 (ap-Gal4) driver (Fig 3A) upregulated rdxZ expression in 85% of discs analyzed (n = 20) and drastically extended the CiFL-expression domain in the dorsal wing disc cells closed to the AP boundary in 82% of cases (n = 17) (Fig 3, compare 3J–J´´ with 3I and 3I´ and 3K´ with 3K). Taken together, we conclude that dac is necessary and sufficient to potentiate Hh signaling activity by promoting CiFL accumulation and the activation of Hh target genes.
The Hh pathway has built-in a negative feedback that serves to attenuate signaling following maximal activation. This feedback rests on the activation of rdx by high signal levels. Once expressed, Rdx drives a Cul3-dependent CiFL degradation posterior to the MF thus allowing the exit from the furrow state [45,46]. We therefore tested if the loss of dac function induces the persistence of Hh signaling posterior to the MF. Indeed, some dac- clones located in internal region of the disc primordium showed ectopic expression of CiFL (Fig 4B and 4B´) and of the Hh-target genes Ptc (Fig 4D and 4D´) and dppZ (Fig 4F–4H). Strikingly, all these clones dropped basally (Fig 4A, 4C and 4E). As sustained exposure to Hh signaling promotes MyOII-dependent cell ingression and groove formation [32,33], we analyzed the shape of dac- clones posterior to the MF using the apical marker E-Cad. We confirmed that the disappearance of dac- cells from the apical surface (Fig 5A) did not result from a loss of cell polarity, as dac- cells maintained E-Cad expression apically (Fig 5B). However, cross section through the eye disc showed that dac- clones ingressed within the epithelium, forming grooves (Fig 5C and 5C´). In addition, these clones accumulated activated MyOII, detected by phospho-Myosin Light Chain (pMLC) antibody at the apical surface of ingressed clones (Fig 5D, 5E and 5E´). These drastic changes in cell shape were associated with the presence of PR nuclei, expressing Elav and hhZ that were still localized close to the apical cell surface but appeared on basal focal planes compared to control GFP-positive neighboring nuclei (Fig 5F and 5F´ to 5I and 5I´). dac- clones spanning the disc margin that delayed MF progression and PRs differentiation also contained higher Ptc levels (S3A and S3A´´ Fig). However, those in which MF initiation and retinogenesis were compromised [49] showed reduced Ptc expression (S3B and S3B´ Fig). Thus, dac is required for the swift dynamic changes in Hh signaling associated to the passing MF. In its absence, Hh target genes activation and Hh-regulated processes suffer a general delay. Interestingly, one of the consequences is that Hh signaling persists for longer, as its attenuation mediated by rdx is also delayed.
To understand how dac promotes CiFL accumulation, we analyzed the requirement for dac to transduce Hh signaling in cells expressing constitutive active forms of Hh pathway components. We first expressed in clones a form of ci insensitive to phosphorylation by PKA (cipka+). Consistent with previous observations, cipka+ clones promoted precocious differentiation anterior to the MF, where cells express dac endogenously [20,51–56]. All cipka+ clones accumulated CiFL (Fig 6A and 6A´; n = 7 discs), while 89% displayed an enrichment of F-actin, reminiscent to the apical cell constriction in the MF (Fig 6C and 6C´; n = 27 discs) and 62% formed ectopic PRs (Fig 6E and 6E´; n = 8 discs). However, posterior to the MF, where CiFL degradation is independent of PKA [57], Cipka+ accumulation was reduced when compared to clones located in the MF and anterior to the MF (Fig 6A and 6A´). Thus, Cipka+ may suffer degradation in this domain. Removing dac function did not affect the ability of clones expressing ectopic cipka+ to accumulate CiFL (Fig 6B and 6B´; 100% of discs, n = 10), F-actin (Fig 6D and 6D´; 95% of discs, n = 22) or differentiate ectopic PRs (Fig 6F and 6F´; 22% of discs, n = 12) anterior to the MF. Therefore, dac acts upstream of CiFL.
We next investigated if dac was required for the activity of the upstream Ci regulator Fused (Fu). Overexpressing a constitutive active form of fu (fuEE+) anterior to the MF also triggered CiFL accumulation in 93% of discs with clones (Fig 6G and 6G´, n = 27 discs), an enrichment of F-actin or E-Cad, reminiscent to the apical cell constriction in the MF in 85% of cases (Fig 6I and 6I´; n = 13 discs), and ectopic PR differentiation in 11% of discs with clones (Fig 6K and 6K´; n = 27 discs). In contrast, when these clones were also mutant for dac, the accumulation of CiFL (Fig 6H and 6H´; 36% of discs, n = 11) and of F-actin (Fig 6J and 6J´; 41% of discs, n = 27) was severely reduced. In addition, these clones were no longer able to differentiate ectopic PRs (Fig 6L and 6L´, n = 6). Further, overexpressing fuEE+ did not rescue the MF delay of dac-mutant tissue (Fig 6H and 6H´ and 6J and 6J´)). Thus, dac potentiates Hh signaling by promoting CiFL accumulation downstream of, or parallel to fu.
dac encodes a nuclear protein, which has been shown to bind double-stranded nucleic acids [58] and to activate transcription of a reporter gene in yeast [59]. We therefore tested, using a genome-wide ChIP-seq approach, the possibility that Dac regulated the expression of some of the components along the Hh signaling pathway (see Materials and Methods). We analyzed specifically ChIP peaks in distal regions (i.e. excluding those falling in 5’UTRs and 1kb upstream of transcription start sites). Within the ChIP peaks, we identified a set of 352 distal regions in which we found significantly enriched an A-rich motif. This motif is similar to the DNA binding motif identified previously for the human DACH by protein structure, in-silico and ChIP-seq analyses (S4 Fig, [60]). This fact indicated that the set of 352 ChIP peaks were likely directly bound by Dac. However, among the nearby genes (S4 Fig), we could not identify any of the major components of the Hh signaling pathway (including hh itself plus smo, PKA, CKI, ci, cullin1, Slimb, skp-1, and cul3). This result suggested that the effect of dac on the activity of the Hh pathway was unlikely to be mediated by the transcriptional regulation of these signaling components. To test experimentally this point, we generated dac- clones and asked whether dac loss affected the expression of the hhZ and ciZ enhancer traps, which serve as transcriptional reporters. Although dac mutant clones located in internal region of the eye primordium appeared to contain reduced hhZ expression (Fig 1H and 1H´), this effect likely resulted from a reduction in the number of PR cells per ommatidia in dac- mutant tissue, as PR hhZ levels were not different than in wild type cells. Similarly, in clones straddling the MF, the absence of dac function delayed hhZ expression, but appeared normally as PR cells gained ELAV signal (S5A and S5A´´´ Fig). In addition, dac was neither necessary nor sufficient to control ci transcription, as ciZ expression was not affected in dac mutant in the eye discs (S5B and S5C´ Fig) or in wing discs overexpressing dac using the ptc-Gal4 driver (S5 Fig, compare S5H to S5J with S5D to S5F and S5K with S5G). We conclude that dac does not promote CiFL accumulation downstream of, or parallel to fu by affecting hh or ci expression and, most likely, neither affecting the transcription of other major pathway components.
The differentiation of the retina in Drosophila occurs both very fast and with high precision: every 2 hours, one new column of assembled ommatidia is added to the developing eye [61]. This phenomenon requires the coordination of gene expression, cell cycle and tissue morphogenesis. All these processes depend critically on the dynamics of Hh signaling. In this report, we show that Dac is an essential element in this dynamics. Dac potentiates Hh signaling in the MF, upstream of Ci and downstream of, or in parallel to Fu. By doing so, dac ensures proper retinogenesis by controlling the timing of MF formation, the accuracy of cell cycle control, the tissue changes associated to the MF, the correct specification of the founder PR8 cell and the attenuation of Hh signaling posterior to the MF, which allows the progression of the differentiation wave.
Our observations demonstrate that dac is required to strengthen Hh signaling. First, the absence of dac function recapitulates phenotypically a reduction of Hh signaling. Removing smo or dac function delays MF progression (Figs 1 and 2, S1 Fig, [23,28–31,49]) and the re-entry in S-phase in the SMW (S2 Fig; [25]). In addition, loss of dac (Fig 1) or smo [23,29,31] results in reduced ato expression ahead of the MF and affects the restriction of ato, first in proneural clusters and then in single PR8 posterior to the MF. Furthermore, neither loss of dac (Fig 1), nor smo [31,35] affects ommatidial cell fate. The cause of the reduced number of ommatidia and variable number of cells per ommatidium in dac mutant clones (Fig 1) is not totally clear. It could result from the effects in ato expression, including its abnormal spacing posterior to the MF and the singling of the ommatidial founder PR8 [62–67]; it could also arise from alterations in cell cycle control, as we detect persistent cell cycling posterior to the MF, which may affect cell recruitment into the ommatidium [25,36]; or from a combination of both. Second, smo and dac synergize to promote MF progression and PR differentiation (S1 Fig). Accordingly, removing one copy of hh enhances the dac mutant eye phenotype and fully suppresses PRs differentiation of dac mutant clones located in internal regions of the disc primordium [49]. Third, dac mutant clones show reduced levels of CiFL and lower expression of the Hh target gene dpp in the MF and rdx posterior to the MF (Fig 2). In addition, ptc expression, another Hh target, is reduced in marginal dac mutant clones that fail to differentiate photoreceptor cells (S3 Fig). Conversely, expressing dac ectopically in the wing disc is sufficient to enhance CiFL accumulation and rdx expression, and to induce dpp in the domain immediately adjacent to the compartment boundary (Fig 3).
Our observations argue that dac potentiates Hh signaling by promoting CiFL accumulation downstream of, or in parallel to Fu. First, loss of dac reduces CiFL levels in the MF (Fig 2), while expressing dac ectopically in the wing disc has the opposite effect (Fig 3). Second, an activated form of Fu (FuEE+), which inhibits Ci processing (into CiR) and promotes CiFL/CiA accumulation [68], requires dac function for this accumulation of CiFL and to induce MF-like features and precocious PR differentiation anterior to the MF. In contrast, removing dac function has no major effect on the ability of a PKA-insensitive form of Ci (Cipka+) to trigger features associated to ectopic Hh signaling in this domain (Fig 6). The precise molecular mechanism by which Dac affects this step along the Hh signaling pathway is unknown at the moment. dac encodes a nuclear protein, which has been shown to bind double-stranded nucleic acids [58] and to activate transcription of a reporter gene in yeast [59]. Our functional and ChIP-seq experiments indicate that hh and ci are not direct transcriptional targets of Dac. In addition, the ChIP-seq data suggest that neither are any of the major pathway components of the Hh pathway (S4 Fig)–although without further studies this possibility cannot be ruled out completely. This would shift the control of the Hh signaling activity to other Dac targets, which would exert this control directly or indirectly. Another, not mutually exclusive possibility is that Dac collaborates with CiA in enhancing the expression of CiA-dependent target genes, such as rdx or ptc (Figs 2 and 3). Although we have previously shown that Dac inhibits the transcriptional ability of the Homothorax/Yorkie (Hth/Yki) complex [69], Dac may act as an activator or repressor depending on the cellular context [70]. Therefore, Dac may contribute to the transcriptional activity of CiA promoting the expression of target genes responsible for proneural fate acquisition and differentiation. In agreement with this hypothesis, using the bioinformatic tool Clover [71], with standard parameters, we also found, besides the Dac motif, the Gli/Ci consensus binding motif significantly enriched (p<0.001) in the distal Dac-ChIP-peaks (S4 Fig), suggesting that Dac might indeed collaborate with CiA to enhancer expression of its targets. In addition, we cannot exclude other mechanisms of Dac action independent of its role in transcriptional regulation. In fact, although it was not reported in Drosophila tissues, the human Dac homologue DACH1 presents both nuclear and cytoplasmic localization in different tissues [72–75]. Moreover, DACH1 localization shifts from the nucleus in normal tissue to the cytoplasm in ovarian cancer [72]. Dac/DACH1 might be involved in the control of Hh signaling pathway in a subcellular localization manner. Further studies are required to elucidate the mechanism by which Dac affects CiFL accumulation and consequently potentiates Hh signaling.
In the developing eye, dac lies downstream of eye absent (eya) and the dpp signaling pathway [50,76–78] where it regulates multiple events that together coordinate cell proliferation and differentiation in time and space. We demonstrate here that dac potentiates Hh signaling. This together with dac’s role in the transition from proliferating progenitor cells to committed precursor cells together with Dpp signal [69] can account for the pleiotropic and essential roles played by dac (Fig 7). Retinogenesis starts with the formation of the MF and the triggering of two major Hh targets: dpp and ato. The weakening of Hh signaling together with the reduced dpp transcription can account for MF delay, as both signals are required for this morphogenetic event [18,19,23,28–31]. Right at the MF and immediately posterior to it, the cell cycle is tightly regulated. We observe that the expression of the p21/p27 Cdk inhibitor dap is lost in dac- cells (S2E Fig) and that this is accompanied by persistent cycling beyond the SMW (S2A and S2C´ Fig). This cell cycle misregulation, together with the aberrant restriction of ato expression, is the likely cause of the abnormal retinogenesis in dac-mutant tissues that includes ommatidia with variable number of cells. Posterior to the MF, high Hh signaling levels activates the expression of Rdx, which, together with Cul3, targets CiFL to full proteasomal degradation. In doing so, Hh signal becomes attenuated, and this attenuation allows cells to exit the furrowed state. In fact, dac-mutant clones posterior to the MF often remain as apically constricted inpouchings of the disc epithelium with high levels of P-MyOII, accumulated CiFL and expression of CiA target genes (Figs 4 and 5).
This report, together with our previous study [69], places dac as an essential regulator of retinal development, controlling the transitions from proliferating progenitor cells to committed precursor cells first [69] and then from precursors to differentiating retinal cells (this work). The dac expression profile spans the regions of the eye disc where it exerts its functions: the increasing Dac expression in precursor cells approaching the MF ensures that cells transit from proliferation to G1 arrest, while peak levels in the MF secure proper retinogenesis by potentiating Hh signaling here. High dac expression in the MF in addition to its downregulation in differentiating photoreceptors could both contribute to turn off Hh signaling in this domain. This could be achieved, at least in part, by inducing the activation of a negative feedback via rdx expression and by limiting CiFL accumulation downstream or in parallel to Fu, respectively. Whether this function is also carried out by the vertebrate dac homologues, the DACH1 and DACH2 genes, awaits further investigation.
Fly stocks used were dac3, smo3, hhP30 (hhZ), rdx03477 (rdxZ), P{dpp-lacZ.Exel.2}3 (dpp-Z),ciZ [79], UAS-HA:dacF [80], UAS-mCD8GFP, UAS-mCherry (TRIP #35787), UAS-fuEE [81–83], UAS-Cipka [82], ptc-Gal4 [84], ap-Gal4 [85]. Mutant clones for dac3 marked by the absence of GFP were generated through mitotic recombination [86]. The MARCM technique [87] was used to induce clones marked by the presence of GFP, mutant for dac3 or smo3 or smo3 and dac3 or expressing UAS-fuEE or UAS-cipka or mutant for dac3 and expressing UAS-fuEE or UAS-cipka. Larvae were heat-shocked for 1 hour at 37°C between 48 and 72h after egg laying. Gain of function experiments using UAS-HA:dacF were performed using the ptc-Gal4 or ap-Gal4 driver that drives expression in the AP boundary compartment or dorsal region of the wing disc, respectively. Crosses carrying UAS-HA:dacF and the corresponding controls were raised at 18°C, while others were raised at 25°C.
Imaginal discs were dissected and fixed according to standard protocols. Primary antibodies used were mouse anti-Dac (1:100; mAbdac2.3, DSHB); mouse anti-CycB (1:25; F2F4, DSHB); mouse anti-β-Galactosidase (1:200; Z378B, Promega), rabbit anti-β-Galactosidase (1:1000; 55976, Cappel); rat anti-E-Cad (1:50; DCAD2, DSHB); guinea pig anti-CycE (1:1000; gift from T. Orr-Weaver, Whitehead Institute, Cambridge, USA); rabbit anti-Ato (1:5000; [22]); rabbit anti-pMLC (1:10; 36715, Cell Signaling), which reveals pMyOII; rat anti-Elav (1:1000; 7E8A10, DSHB); mouse anti-Elav (1:100; 9F8A9, DSHB); guinea pig anti-Sens (1:1000; [88]); guinea pig anti-Pros (1:25; [89]); rabbit anti-Sal (1:200; [90]); rabbit anti-PH3 (1:200; 9701, Cell Signaling); mouse anti-Ptc (1:100; Drosophila Ptc (Apa1), DSHB); rat anti-Ci (1:10; 2A1, DSHB). Rhodamine-conjugated (Sigma) and C660–conjugated Phalloidin (Biotium) were used at a concentration of 0.3 μM and 5U/ml, respectively. DAPI was used at a concentration of 1ng/ml. Fluorescently labeled secondary antibodies were from Jackson Immunoresearch, (1:200). Imaging was carried out on Leica SP2 or SP5 confocal microscopy set ups. Plot profiles of fluorescent intensity were obtained using an NIH ImageJ program [91]. Fluorescent intensities were normalized to the maximum intensity for each channel.
Anti-mouse-HRP (Sigma) was used for immunoperoxidase staining. Digoxigen labelled dap RNA probe (Roche) was produced from the cDNA clone LP07247 (BDGP). For BrdU incorporation assay, eye-antennal discs were dissected and incubated in 10μM BrdU in PBS for 30min. BrdU was detected with an anti-BrdU antibody (1:400; Roche) after treatment with DNase.
Wandering 3rd instar larvae (Dac:GFP, Bloomington stock 42269) were dissected in cold PBS and imaginal discs were fixed with formaldehyde for 25 minutes. Chromatin was fragmented by sonication till it reached an average size of 500 bp. 20 μl of protein A/G magnetic beads (Merck, Millipore) was added to pre-clean the samples. The anti-GFP Ab (ab290, Abcam) was added to a fixed chromatin aliquot and incubated at 4°C overnight. Immunocomplexes were recovered by adding protein A/G magnetic beads to the sample and incubating for 3 hours at 4°C. Beads were resuspended in elution buffer, RNase was added to the immunoprecipitated chromatin and incubated for 30 minutes at 37°C. ChIP libraries were prepared with the Truseq DNA library prep kit (Illumina) and the samples were sequenced on a HiSeq 2000 (Illumina). The reads were cleaned using fastq-mcf and mapped with bowtie2 to the Drosophila melanogaster genome (Flybase version 5). Dac-ChIP peaks (minus input) were called using macs2 (dac-ChIP-vs-ChIP_Input_peaks.bed). From this bed file all regions lying in a 5’UTR or 1kb upstream of a TSS were removed using bedtools intersectBed, to retain only distal Dac-ChIP-peaks (dac-ChIP-vs-ChIP_Input_peaks_not-in-5UTR_not-in-1kb-up.bed). The distal Dac-ChIP-peaks were loaded into i-cistarget [92], a motif enrichment tool, to obtain a final table (S4 Fig) of potential Dac targets. Access to the whole dataset can be found in the GEO database with the accession number GSE82151 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE82151), used as reference for all subsequent manuscripts referring to these data.
|
10.1371/journal.pntd.0005234 | Naloxonazine, an Amastigote-Specific Compound, Affects Leishmania Parasites through Modulation of Host-Encoded Functions | Host-directed therapies (HDTs) constitute promising alternatives to traditional therapy that directly targets the pathogen but is often hampered by pathogen resistance. HDT could represent a new treatment strategy for leishmaniasis, a neglected tropical disease caused by the obligate intracellular parasite Leishmania. This protozoan develops exclusively within phagocytic cells, where infection relies on a complex molecular interplay potentially exploitable for drug targets. We previously identified naloxonazine, a compound specifically active against intracellular but not axenic Leishmania donovani. We evaluated here whether this compound could present a host cell-dependent mechanism of action. Microarray profiling of THP-1 macrophages treated with naloxonazine showed upregulation of vATPases, which was further linked to an increased volume of intracellular acidic vacuoles. Treatment of Leishmania-infected macrophages with the vATPase inhibitor concanamycin A abolished naloxonazine effects, functionally demonstrating that naloxonazine affects Leishmania amastigotes indirectly, through host cell vacuolar remodeling. These results validate amastigote-specific screening approaches as a powerful way to identify alternative host-encoded targets. Although the therapeutic value of naloxonazine itself is unproven, our results further demonstrate the importance of intracellular acidic compartments for host defense against Leishmania, highlighting the possibility of targeting this host cell compartment for anti-leishmanial therapy.
| Leishmaniasis is a poverty-related disease threatening 350 million people throughout the world. It is caused by the protozoan parasite Leishmania, a digenetic organism that switches from an extracellular stage in the sand fly vector to an intracellular stage in phagocytes of the vertebrate host. Drugs currently available to treat leishmaniasis are toxic to the patient and drug-resistant parasites are emerging, urging for new therapeutics. A novel strategy to tackle intracellular pathogens entails targeting the host cell, in order to indirectly interfere with pathogens growth. Here we analysed the mechanism of action of naloxonazine, a compound previously shown to specifically affect the intracellular amastigote stage of Leishmania. We show that this compound affects acidic compartments of macrophages and that these naloxonazine-induced modifications are responsible for Leishmania intracellular growth inhibition. Even though the therapeutic potential of naloxonazine itself is not proven, our results reveal the possibility of targeting host cell intracellular acidic compartments for anti-leishmanial therapy.
| Protozoan parasites of the genus Leishmania are the causative agents of a wide variety of diseases ranging from self-healing or severe mucocutaneous lesions to a visceral disease which is lethal in the absence of treatment. Leishmaniasis is one of the most significant neglected tropical diseases, with an estimated 12 million people infected. Leishmania parasites have a digenetic life cycle; switching from an insect vector in which parasites dwell as extracellular promastigotes, to a mammalian host, where parasites reside exclusively intracellulary (intramacrophage amastigote stage).
Pentavalent antimonials (SbV) like sodium stibogluconate (SSG) have been the first-line treatment against leishmaniasis for several decades but their clinical value has become compromised by increasing treatment failure and the emergence of resistant parasites. This concern is particularly important in the Indian subcontinent where visceral leishmaniasis (VL) caused by Leishmania donovani is endemic and where most VL cases occur [1]. Current treatment alternatives consist of amphotericin B, miltefosine or paromomycin (in mono- or combination therapy) but these compounds also have drawbacks including cost, toxicity or decreased efficacy after a few years of use [2].
Although the mechanism of action of these compounds is not fully understood, they are all known to target Leishmania components, therefore directly interfering with parasite growth: amphotericin B forms a complex with ergosterol, the main sterol of Leishmania cellular membrane, leading to formation of aqueous pores and increased membrane permeability [3]; miltefosine has been shown to inhibit the parasite cytochrome c oxidase and to cause apoptosis-like processes in L. donovani [4]; and paromomycin is an aminoglycoside antibiotic that inhibits protein synthesis in Leishmania with low host cell toxicity [5]. SbV on the other hand, has been shown to target both the parasite and the host cell: SbV is reduced to trivalent antimony (SbIII), which directly alters the parasite redox metabolism and antioxidant defense system, but SbV itself also indirectly affects parasite survival by increasing host cell production of toxic oxygen and nitrogen intermediates, thereby creating additional oxidative and nitrosative stress upon SbIII-sensitized parasites [6]. Antimonial anti-leishmanial activity is thus partly indirect, targeting host cell pathway(s) that consequently affect Leishmania intracellular development.
Targeting host cell pathways to interfere with the intracellular development of pathogens is a strategy increasingly investigated for antimicrobial therapy that might bring novel therapeutic approaches in a context of increased treatment failure and poor alternatives [7,8]. Following this line, a recent high-throughput screening campaign against kinetoplastids at GlaxoSmithKline identified several compounds associated with human proteins with no known homologs in kinetoplastids, highlighting the possibility of targeting host-pathogen interactions[9].
Here we report the host-dependent anti-leishmanial activity of naloxonazine, a mu-opioid receptor (MOR) antagonist. This compound was first identified in a high-throughput screen against Leishmania donovani intracellular amastigotes [10]. We now show that it affects host cell intracellular compartments thereby inhibiting Leishmania establishment in the phagolysosomal vacuole.
Parasite strains used in this study included L. donovani 1S2D (MHOM/SD/62/1S-cl2D), L. donovani 1S2D expressing the enhanced green fluorescent protein (eGFP) and two L. donovani clones of clinical isolates from the Terai endemic region in Nepal (MHOM/NP/02/BPK282/0cl4 and MHOM/NP/03/BPK275/0cl18 respectively susceptible and resistant to SSG and further designated SSG-S BPK282 and SSG-R BPK275). Promastigotes were maintained at 26°C in hemoflagellate modified Eagles’s medium (HOMEM) supplemented with 20% Foetal Bovine Serum (FBS). Differentiation of promastigotes into axenic amastigotes was achieved as described previously [11]. THP-1 cells (human acute monocytic leukemia cell line–ATCC TIB202) were grown in RPMI supplemented with 10% FBS and 50 μM 2-mercaptoethanol at 37°C in 5% CO2. For Leishmania infections, THP-1 cells were treated with 0.1 μM phorbol myristate acetate (PMA, Sigma) at 37°C for 48 h to achieve differentiation into adherent, non-dividing macrophages. Cells were washed and incubated with complete RPMI medium containing stationary phase L. donovani promastigotes at a macrophage/promastigote ratio of 1/10. After 4 h incubation at 37°C, non-internalized promastigotes were removed by 3 successive washes with PBS and incubated with naloxonazine, naloxone, β-funaltrexamine, CTOP, endomorphine, DAMGO, sinomenine, concanamycin A (all purchased from Sigma) or imatinib (Cell Signaling Technology) for 24 to 72 h. Half maximal inhibitory concentrations (GI50) were determined using a high-content imaging assay as described previously [10]. Briefly, compounds were serially diluted 3-fold in DMSO, with final assay concentrations ranging from 50 μM to 0.02 μM (1% final concentration of DMSO), 2 μM amphotericin B and 1% DMSO were used as positive and negative controls respectively. For confocal microscopy, infected cells were washed with PBS, fixed for 30 minutes with 4% formaldehyde, rinsed again with PBS and stained with 4’,6’-diamidino-2-phenylindole (DAPI 300 nM). Images were acquired with an LSM 700 Zeiss confocal microscope.
20 μM naloxonazine was incubated for 50 h in RPMI 10% FBS- 50 μM 2-mercaptoethanol with or without THP-1 cells at a concentration of 106 cells/ml. 100 μl of culture media were collected at different time points (T0; 0,5; 1; 5; 10; 20; 30; 45 and 50 h) and kept frozen. 20 μl of these samples were then mixed with 40 μl cold acetonitrile containing either 2 μM naloxone or 2 μg/mL K777 (N-methylpiperazine-PhehomoPhe-vinylsulfone-phenyl), centrifuged, and 3 μL per sample injected into an API4000 (AB Sciex) LC-MS/MS system and analysed with positive-ion-mode electrospray ionization. A binary mobile phase (A,15% methanol:water; B, 100% methanol:water; both containing 0.1% formic acid, 0.1% ACN and 160 mg/L NH4OAc) was pumped at 0.5 mL/min through a 4.6 x 50 mm, 5 μm, 100 Å pore Kinetex C18 column (Phenomenex). The gradient used was: 0–0.5 min, 0% B; 0.5–3.0 min, linear ramp to 100% B; 3.0–4.0 min, 100% B; 4.0–4.5 min, linear ramp to 0% B; 4.5–7.0 min, 0% B. MS settings were as follows: common settings were temperature = 600°C, GS1 (ion source nebulizer gas) = GS2 (ion source heater gas) = 50 lbf in-2; CUR (curtain gas) = 35 lbf in-2; CAD (collision gas) = 12 lbf in-2; IS (ion spray voltage) = 5500 V; analyte-specific settings for naloxonazine, naloxone and K777, repectively, were DP (declustering potential) = 101 V, 76 V and 56 V; EP (entrance potential) = 13.2 V, 10 V and 10 V; CE (collision energy) = 47 eV, 37 eV and 57 eV; CXP (collision cell exit potential) = 18 V, 14 V and 18 V. The MS/MS transitions used were naloxonazine, m/z 651.5 → 325.3; naloxone, 328.3 → 253.1; K77, 575.5 → 101.3, and retentions were 3.16, 3.03 and 4.43 min, respectively.
MOR-specific siRNA was purchased from Qiagen (Hs_OPRM1_7 FlexiTube siRNA). THP-1 cells were transfected with 1 μM of siRNA using the Amaxa nucleofector kit V following the manufacturer’s instructions. Control cells were mock transfected in parallel (“mock” control). After nucleofection, THP-1 cells were resuspended in complete RPMI medium (5. 105 cells/ml), treated with PMA for 24 h and infected with stationary phase L. donovani 1S2D promastigotes as described above. Parasite infectivity was assessed 48 h after infection. Down-regulation of MOR mRNA level was analysed 24 h after nucleofection and 48 h after infection by qRT PCR as described below using the following primer sets: MOR fwd: 5’ GGTACTGGGAAAACCTGCTGAAGATCT, rev: 5’ GGTCTCTAGTGTTCTGACGAATTCGAGTGG and 18S rRNA: Fwd 5' ACCGATTGGATGGTTTAGTGAG, Rev 5' CCTACGGAAACCTTGTTACGAC. The relative expression level of MOR was determined based on the Ct value normalized to the Ct value of the reference 18S rRNA. siRNA treated cells were compared to the mock transfected cells.
Non-infected, PMA-activated THP-1 cells (5.105 cells/ml, 10 ml) were treated with 10 μM of naloxonazine or 10 μM of naloxone. After 4 h of treatment, compounds were removed by 2 washes with PBS and cells were further incubated 20 h in compound-free RPMI medium. This time-point was chosen to maximize the chances of detecting naloxonazine-induced transcriptional changes while limiting the observation of downstream effects. Total RNA was extracted using TRIzol (Invitrogen) and amplified with the Amino Allyl MessageAmp™ II aRNA Amplification Kit (Ambion) following the manufacturer’s protocol. The monofunctional NHS-ester Cy3 and Cy5 dyes (GE Healthcare Life Sciences) were coupled with 10 μg amplified RNA. The two aRNA pools to be compared were mixed and applied to the Human Exonic Evidence Based Oligonucleotide (HEEBO) array (Stanford Functional Genomics Facility). HEEBO oligonucleotide set consists of 44,544 70mer probes that were designed using a transcriptome-based annotation of exonic structure for genomic loci. Four samples (two from naloxonazine-treated and two from naloxone-treated cells) were competitively hybridized on two individual chips (further called “array1” and “array2”). The hybridization was performed at 63°C for 16 h in a humidified slide chamber containing the labeled probe, 3X SSC, and 0.2% SDS. After hybridization, the hybridization chamber was removed from the 63°C water bath, washed with 0.6X SSC, 0.03% SDS, and then 0.06X SSC. Microarrays were scanned using a GenePix Pro Axon 4000B scanner, data were analysed with the Acuity software (Molecular Devices). Fluorescent data were background adjusted and the ratios of naloxonazine-treated to naloxone-treated data were calculated for each probe set. Sets of genes showing a ratio > 2 were functionally clustered using DAVID [12,13].
RNA was extracted as described above, from non-infected THP-1 cells treated with 10 μM of naloxonazine for 4 h and further incubated 20 h in compound-free RPMI medium. cDNA synthesis was done with Transcriptor Reverse Transcriptase (Roche) and a 15-mer oligo(dT) primer from 1 μg of total RNA. qPCRs were run with the SensiMix SYBR no-ROX kit (Bioline) on a LightCycler 480 (Roche). The following primer sets were used: vATPase subunit c (ATP6V0C): Fwd 5' ATGTCCGAGTCCAAGAGC, Rev 5' CTACTTTGTGGAGAGGATGAG; vATPase subunit a (TCIRG1): Fwd 5’ ATCTGGCAGACTTTCTTCAG, Rev 5’ AAGATGCTGGTGGCGCGACT; B-Actin (ACTB): Fwd 5' TCCCTGGAGAAGAGCTACGA, Rev 5' AGCACTGTGTTGGCGTACAG; 18S rRNA: Fwd 5' ACCGATTGGATGGTTTAGTGAG, Rev 5' CCTACGGAAACCTTGTTACGAC. The relative expression levels of vATPase and actin were determined based on the Ct value of each gene normalized to the Ct value of the reference 18S rRNA. Naloxonazine treated cells were compared to untreated cells.
Total protein extracts of THP-1 cells infected with L.d. 1S2D, treated or not with 10 μM of naloxonazine for 24 or 48 h, were prepared in Laemmli sample buffer (Bio Rad) and analysed by SDS-PAGE and western blotting. The equivalent of 105 cells were loaded per well. Membranes were first incubated with an anti-vATPase subunit a3 (rabbit polyclonal anti-TCIRG1, abcam, 1:1000), and an anti-rabbit HRP (1:5000), then stripped with Restore western blot stripping buffer (Thermofisher) and further incubated with an anti α-tubulin (mouse monoclonal, abcam, 1:1000) and an anti-mouse HRP (1:5000). Proteins were detected by chemoluminescence following the manufacturer’s instructions (PierceECL western blotting substrate, Thermofisher). Quantitative densitometry was performed using Image J.
THP-1 cells infected with eGFP-expressing L. donovani were treated or not with 10 μM of naloxonazine, 10 μM of naloxone or 80 nM of concanamycin A for 24 h, or co-treated with either naloxonazine (10 μM) and concanamycin A (80 nM) or naloxone (10 μM) and concanamycin A (80 nM) for 24 h, then stained with 180 nM of the Lysotracker red DND-99 (Life Technologies) for 1 h at 37°C. For microscopy, cells were further stained with 500 nM of the nucleic acid stain Hoechst 33342 (Life Technologies) and images were acquired with an LSM 700 Zeiss confocal microscope. For flow cytometry, Lysotracker red DND-99 stained cells were first trypsinised with TrypLE Select (Invitrogen), washed with PBS and analysed with a BD FACSVerse flow cytometer and the BD FACSSuite software.
GraphPad Prism 5 software was used to determine the statistical significance (Two-way ANOVA or t-test as specified in the figure legends).
Clinical samples were from an already existing collection (B.P. Koirala Institute of Health Sciences in Dharan). All samples were anonymized and their use was approved by the review boards of the Nepal Health Research Council, Kathmandu, the Institute of Tropical Medicine, Antwerp and the Antwerp University.
The activity of naloxonazine was tested in vitro against three stages of Leishmania donovani: insect-stage promastigotes, intracellular amastigotes (within the macrophage host cell) and host cell-free axenic amastigotes (an amastigote-like stage obtained from differentiation of promastigotes in vitro in the absence of a host cell). Naloxonazine was shown to be active against the intracellular amastigote stage with a half maximal inhibitory concentration (GI50) of 3,45 μM. It exhibited a reasonable selectivity, with a GI50 of 34 μM against the THP-1 host cell. Remarkably, the compound was inactive against L. donovani promastigotes or axenic amastigotes, indicating the importance of the host cell microenvironment for compound activity (Fig 1A). Naloxonazine was also tested against two L. donovani clinical isolates, one showing susceptibility, the other showing resistance to antimonials (SSG-S and SSG-R strains). The activity of naloxonazine against these isolates was comparable, suggesting that the mechanism of resistance developed against SSG does not affect naloxonazine activity (Fig 1B).
The necessity of the host cell presence for naloxonazine’s anti-leishmanial activity might be hypothesized to be linked to the metabolic properties of macrophages, i.e. naloxonazine could be a prodrug dependent on host cell metabolism to gain anti-leishmanial activity. In order to define the exact chemical moiety endowed with anti-leishmanial activity and to evaluate whether the macrophage would metabolize naloxonazine into an active compound, naloxonazine stability during incubation in THP-1 cell medium was evaluated by LC-MS/MS in the presence or absence of THP-1 host cells. Naloxonazine had a half-life of 15 h and was shown to be degraded into naloxone, another MOR antagonist, regardless of the presence of THP-1 macrophages (Fig 1C). Interestingly, naloxone was shown to be inactive against all stages of L. donovani, including the intracellular amastigotes (Fig 1D). Naloxonazine is thus not a prodrug activated by the macrophage host cell but its activity seems inherent, associated with its unperturbed chemical identity.
The kinetics of naloxonazine activity showed that parasite growth was already inhibited by 70% after 24 h of compound incubation; 95% of growth inhibition was achieved after 72 h incubation (Fig 1E). Remarkably, exposure of infected macrophages to naloxonazine for 4 h, followed by a washing step to remove the compound from the cells and an additional incubation of 70 h, led to the same level of parasite growth inhibition as a 72 h-incubation with the compound (Fig 1F). This observation is in accordance with the degradation time of naloxonazine (Fig 1C). Moreover, delaying addition of naloxonazine to 24 or 48 h after infection reduced its anti-leishmanial effect by 75%, suggesting that naloxonazine is most active at early stages of infection.
We hypothesized that naloxonazine anti-leishmanial activity is dependent on its antagonistic effect towards MOR of macrophages. siRNA-mediated knock-down of MOR was therefore carried out in THP-1 cells to evaluate the importance of these receptors for parasite intracellular growth. Fifty percent down-regulation of MOR mRNA was obtained, but this was not accompanied by changes in infection levels (Fig 2A and 2B). The amount or function of the MOR protein was not assessed in the siRNA-treated cells; however, phenocopy of naloxonazine’s effect on L. donovani growth could not be observed with a set of other antagonists or agonists of opioid receptors (Fig 2C), supporting the conclusion that the activity of naloxonazine on Leishmania intracellular growth is independent of opioid receptors.
Microarray profiling of THP-1 cells treated with naloxonazine or naloxone was performed to pinpoint host cell pathways differentially affected by the drugs and identify pathways that could be important for Leishmania intracellular growth. A 4 h-compound incubation followed by an additional compound-free incubation of 20 h was chosen, to maximize the chances of detecting naloxonazine-induced transcriptional changes while limiting the observation of downstream effects. Two percent of the probes showed at least two-fold differential gene expression in naloxonazine versus naloxone treated THP-1 cells. These upregulated genes were functionally clustered with the Database for Annotation, Visualization and Integrated Discovery (DAVID [12]). Both the vacuolar H+ ATPase gene family and actin and actin-related genes clusters were perturbed, pointing to a possible effect of naloxonazine on phagolysosome formation or maturation (Table 1). The expression level of two vATPase subunits and actin was also analysed by qRT PCR in naloxonazine-treated compared to untreated cells at the same time point as the one chosen for the microarray experiment. Upregulation of these genes after naloxonazine treatment was confirmed (Fig 3A). Upregulation of vATPase subunit a3 was also established at protein level after 24 and 48 h of treatment (Fig 3B). To test whether naloxonazine could affect the phagolysosome, L. donovani-infected THP-1 cells treated or not with naloxonazine were stained with Lysotracker, a fluorescent acidotropic probe that accumulates in cellular compartments with low internal pH. Stained cells were analysed by confocal microscopy and flow cytometry. The intensity of the Lysotracker signal was increased after naloxonazine treatment, indicating an increased combined volume of acidic vacuoles (Fig 3C, 3D and 3E). These results suggested that naloxonazine influenced the intracellular acidic compartments of the host cell.
In order to determine if naloxonazine-induced changes in intracellular compartments were responsible for the effect on L. donovani intracellular growth, L. donovani-infected THP-1 cells were treated with both naloxonazine and the vATPase inhibitor concanamycin A. Remarkably, concanamycin A was sufficient to restore normal infection levels in naloxonazine-treated cells, confirming the importance of host cell-acidic compartments for controlling L. donovani intracellular growth (Fig 4A and 4B). To further establish the importance of acidic compartments for Leishmania intracellular growth inhibition, we evaluated the activity of imatinib, an inhibitor of Abelson tyrosine kinase previously shown to trigger intracellular acidification in monocyte-derived macrophages [14, 23]. In agreement with our previous observations, imatinib exhibited anti-leishmanial activity in the low micromolar range (GI50 of 4 μM; Fig 4C).
We showed that naloxonazine does not directly target L. donovani but rather interferes with intracellular acidic compartments of the host cell. Upon infection, Leishmania parasites are recognized by phagocytic cells and internalized by phagocytosis in a phagosome/parasitophorus vacuole. During the process of phagocytosis, the phagosome matures through fusion with endosomes and lysosomes, ultimately leading to a highly microbicidal environment. One component of this microbicidal response is the acidification of the phagosome due to the recruitment of the vATPase proton pump to the mature phagosomal membrane [15]. Intracellular pathogens can survive these extreme conditions by arresting phagosomal maturation at an early non-microbicidal stage, developing resistance to the microbicidal arsenal of the phagolysosome, or escaping from the phagosome into the cytosol.
It is well established that Leishmania amastigotes are adapted to the acidic pH found in the parasitophorus vacuole and are able to proliferate under these conditions [16,17]. In contrast, it has been proposed that promastigotes, the parasite stage of the insect vector, delay phagosome maturation to avoid destruction before differentiation into amastigotes [18,19]. Although this hypothesis has been challenged by the observation that promastigote-containing parasitophorus vacuoles do fuse with lysosomes [20], the importance of acidic pH for controlling intracellular Leishmania growth is well recognized. Infection studies in Stat1 deficient mice for instance showed an increased Leishmania intracellular growth that was associated with an increase in phagosomal pH [21].
Our study demonstrated a naloxonazine-induced increased expression of the vATPase transporter as well as an increase of Lysotracker-positive intracellular compartments which was associated to an enhanced capacity of the host cell to control infection. Whether naloxonazine influences the pH of the parasitophorus vacuole or the amount of acidic vacuoles is unclear at this stage. Time-course analysis supported a naloxonazine anti-leishmanial effect at early stages of infection, in accordance with previous observations showing the importance of acidic compartments in the early steps of infection [18, 19]. The microarray analysis performed in this study also showed upregulation of actin and some actin-related genes in naloxonazine-treated cells. However, further investigation is required to assess the importance of actin for the anti-leishmanial activity of naloxonazine.
The molecular pathways affected by naloxonazine that lead to modification of the phagosome are yet to be determined. Naloxonazine is a MOR antagonist [22]; however, albeit such receptors are expressed by macrophages and more specifically by the THP-1 cell line, they could not be linked to Leishmania growth inhibition. Naloxone, β-funaltrexamine or CTOP, other MOR antagonists, were inactive against L. donovani and knock-down of MOR in THP-1 cells did not affect L. donovani intracellular growth. Moreover, naloxonazine is a very potent MOR antagonist (Kd < 2 nM) while its activity against L. donovani is in the micromolar range. These data suggest that the cellular target of naloxonazine in this case is independent of MOR.
Targeting phagolysosome acidification to fight against intracellular pathogens is a strategy that has previously been validated for Mycobacterium tuberculosis, a pathogen that infects macrophages through delayed phagosome maturation. Imatinib, an Abelson tyrosine kinase inhibitor used to treat early chronic myeloid leukemia, was shown to decrease the pH of intracellular compartments which in turn reduced M. tuberculosis intracellular growth in vitro and in vivo (14,23). Whether Abelson tyrosine kinases are involved in the naloxonazine-induced pH decrease of intracellular compartments deserves further investigation.
Host-directed therapies (HDT) are considered an innovative strategy for infectious diseases, given the concern over parasites evolving resistance to current treatments, combined with the recognition of the importance of host determinants for progression of infections. HDTs are receiving increasing attention in treatment of tuberculosis for instance, and are expected to improve treatment outcomes against drug-susceptible as well as multi-drug-resistant strains [24]. Examples of HDTs under evaluation for tuberculosis treatment include anti-inflammatory compounds, statins and imatinib [25]. In antiviral therapy, HDTs have also been raised as interesting alternatives and possible solutions for limiting the emergence of drug resistance [26].
Treatment against leishmaniasis is also jeopardized by increasing treatment failure and drug resistance. Of importance, treatment failure does not necessarily correlate with parasite drug resistance (at least against SSG or miltefosine), highlighting the importance of the host background for treatment outcome in this case [27]. This observation would therefore also argue in favor of HDTs against leishmaniasis. In this context, immunotherapy has received considerable interest in recent years, with the idea of modulating the immune system to achieve a protective response and parasite elimination [28,29]. Combination of immuno- and chemo-therapy is believed to be synergistic, allowing infection control while reducing the threat of drug resistance.
Whether HDTs would be less likely to induce resistance remains an interesting and important question, and for Leishmania at least, the example of SbV raises concern. Indeed, although SbV targets host cell defense pathways, SbV-R parasites have been isolated in the Indian subcontinent, and notably these parasites showed increased infectivity [30]. This ought to raise concern of possibly generating more virulent strains should parasites become resistant to immuno-therapy.
Targeting host-encoded functions unrelated to the immune response but important for parasite invasion or intracellular development provides additional options for HDTs for leishmaniasis and intracellular pathogens in general. This is exemplified by naloxonazine or imatinib and their interference with endocytic components of the host cell.
Although the potential of naloxonazine itself as a therapeutic anti-leishmanial drug is unproven, targeting pathways linked to phagosome acidification remains of great interest. In addition, naloxonazine was as potent against both SSG-R and SSG-S clinical isolates, highlighting the possible benefits of such a drug target for parasites resistant to classic chemotherapy.
|
10.1371/journal.pgen.1004383 | Bayesian Test for Colocalisation between Pairs of Genetic Association Studies Using Summary Statistics | Genetic association studies, in particular the genome-wide association study (GWAS) design, have provided a wealth of novel insights into the aetiology of a wide range of human diseases and traits, in particular cardiovascular diseases and lipid biomarkers. The next challenge consists of understanding the molecular basis of these associations. The integration of multiple association datasets, including gene expression datasets, can contribute to this goal. We have developed a novel statistical methodology to assess whether two association signals are consistent with a shared causal variant. An application is the integration of disease scans with expression quantitative trait locus (eQTL) studies, but any pair of GWAS datasets can be integrated in this framework. We demonstrate the value of the approach by re-analysing a gene expression dataset in 966 liver samples with a published meta-analysis of lipid traits including >100,000 individuals of European ancestry. Combining all lipid biomarkers, our re-analysis supported 26 out of 38 reported colocalisation results with eQTLs and identified 14 new colocalisation results, hence highlighting the value of a formal statistical test. In three cases of reported eQTL-lipid pairs (SYPL2, IFT172, TBKBP1) for which our analysis suggests that the eQTL pattern is not consistent with the lipid association, we identify alternative colocalisation results with SORT1, GCKR, and KPNB1, indicating that these genes are more likely to be causal in these genomic intervals. A key feature of the method is the ability to derive the output statistics from single SNP summary statistics, hence making it possible to perform systematic meta-analysis type comparisons across multiple GWAS datasets (implemented online at http://coloc.cs.ucl.ac.uk/coloc/). Our methodology provides information about candidate causal genes in associated intervals and has direct implications for the understanding of complex diseases as well as the design of drugs to target disease pathways.
| Genome-wide association studies (GWAS) have found a large number of genetic regions (“loci”) affecting clinical end-points and phenotypes, many outside coding intervals. One approach to understanding the biological basis of these associations has been to explore whether GWAS signals from intermediate cellular phenotypes, in particular gene expression, are located in the same loci (“colocalise”) and are potentially mediating the disease signals. However, it is not clear how to assess whether the same variants are responsible for the two GWAS signals or whether it is distinct causal variants close to each other. In this paper, we describe a statistical method that can use simply single variant summary statistics to test for colocalisation of GWAS signals. We describe one application of our method to a meta-analysis of blood lipids and liver expression, although any two datasets resulting from association studies can be used. Our method is able to detect the subset of GWAS signals explained by regulatory effects and identify candidate genes affected by the same GWAS variants. As summary GWAS data are increasingly available, applications of colocalisation methods to integrate the findings will be essential for functional follow-up, and will also be particularly useful to identify tissue specific signals in eQTL datasets.
| In the last decade, hundreds of genomic loci affecting complex diseases and disease relevant intermediate phenotypes have been found and robustly replicated using genome-wide association studies (GWAS, [1]). At the same time, gene expression measurements derived from microarray [2] or RNA sequencing [3] studies have been used extensively as an outcome trait for the GWAS design. Such studies are usually referred to as expression quantitative trait locus (eQTL) analysis. While GWAS datasets have provided a steady flow of positive and replicable results, the interpretation of these findings, and in particular the identification of underlying molecular mechanisms, has proven to be challenging. Integrating molecular level data and other disease relevant intermediate phenotypes with GWAS results is the natural step forward in order to understand the biological relevance of these results. This strategy has been explored before and allowed the identification of the genes and regulatory variations that are important for several diseases (reviewed in [4]).
In this context, a natural question to ask is whether two independent association signals at the same locus, typically generated by two GWAS studies, are consistent with a shared causal variant. If the answer is positive, we refer to this situation as colocalised traits, and the probability that both traits share a causal mechanism is greatly increased. A typical example involves an eQTL study and a disease association result, which points to the causal gene and the tissue in which the effect is mediated [5]–[7]. In fact, looking for overlaps between complex trait-associated variants and eQTL variants has been successfully used as evidence of a common causal molecular mechanism (e.g., [5], [8]). The same questions can also be considered between pairs of eQTLs [9], [10], or pairs of diseases [11].
However, identifying the traits that share a common association signal is not a trivial statistical task. Visual comparison of overlaps of association signals with an expression dataset is a step in this direction (using for example Sanger tool Genevar http://www.sanger.ac.uk/resources/software/genevar/), but the abundance of eQTLs in the human genome and across different tissues makes an accidental overlap between these signals very likely [2]. Therefore visual comparison is not enough to make inferences about causality and formal statistical tests must be used to address this question.
Nica et al. [5] proposed a methodology to rank the SNPs with an influence on two traits based on the residual association conditional on the most associated SNP. By comparing the GWAS SNP score with all other SNPs in the associated region, this method accounts for the local LD structure. However, this is not a formal test of a null hypothesis for, or against, colocalisation at the locus of interest. A formal test of colocalisation has been developed in a regression framework. This is based on testing a null hypothesis of proportionality of regression coefficients for two traits across any set of SNPs, an assumption which should hold whenever they share causal variant(s) [12], [13]. No assumption is made about the number of causal variants, although the method does assume that in the case of multiple causal variants, all are shared. Both the ranking method and proportionality testing share the drawback of having to specify a subset of SNPs to base the test on, and Wallace [14] shows that this step can generate significant biases. The main sources of bias are overestimation of effect sizes at selected SNPs (termed “Winner's curse”), and the fact that, owing to random fluctuations, the causal variant may not always be the most strongly associated one. These factors lead to rejection of colocalisation in situations where the causal SNP is in fact shared. Although this can be overcome in the case of proportionality testing by averaging over the uncertainty associated with the best SNP models [14], perhaps the greatest limitation is the requirement for individual level genotype data, which are rarely available for large scale eQTL datasets.
The success of GWAS meta-analyses has shown that there is considerable benefit in being able to derive association tests on the basis of summary statistics. With these advantages in mind, He et al. [7] developed a statistical test to match the pattern of gene expression with a GWAS dataset. This approach, coded in the software Sherlock, can accommodate p-values as input. However, their hypothesis of interest differs from the question of colocalisation, with the focus of the method being on genome-wide convergence of signals, assuming an abundance of trans eQTLs. In particular, SNPs that are not associated with gene expression do not contribute to the test statistic. Such variants can provide strong evidence against colocalisation if they are strongly associated with the GWAS outcome.
These limitations motivate the development of novel methodologies to test for colocalisation between pairs of traits. Here, we derive a novel Bayesian statistical test for colocalisation that addresses many of the shortcomings of existing tools. Our analysis focuses on a single genomic region at a time, with a major focus on interpreting the pattern of LD at that locus.
Our underlying model is closely related to the approach developed by Flutre et al. [10], which considers the different but related problem of maximising the power to discover eQTLs in expression datasets of multiple tissues. A key feature of our approach is that it only requires single SNP p-values and their minor allele frequencies (MAFs), or estimated allelic effect and standard error, combined with closed form analytical results that enable quick comparisons, even at the genome-wide scale. Our Bayesian procedure provides intuitive posterior probabilities that can be easily interpreted. A main application of our method is the systematic comparison between a new GWAS dataset and a large catalogue of association studies in order to identify novel shared mechanisms. We demonstrate the value of the method by re-analysing a large scale meta-analysis of blood lipids [15] in combination with a gene expression study in 966 liver samples [16].
We consider a situation where two traits have been measured in two distinct datasets of unrelated individuals. We assume that samples are drawn from the same ethnic group, i.e. allele frequencies and pattern of linkage disequilibrium (LD) are identical in both populations. For each of the two samples, we consider for each variant a linear trend model between the outcome phenotypes Y and the genotypes X (or a log-odds generalised linear model if one of the two outcome phenotypes Y is binary):We are interested in a situation where single variant association p-values and MAFs, or estimated regression coefficients and their estimated precisions , are available for both datasets at Q variants, typically SNPs but also indels. We make two additional assumptions and discuss later in this paper how these can be relaxed. Firstly, that the causal variant is included in the set of Q variants, either directly typed or well imputed [17]–[19]. Secondly, that at most one association is present for each trait in the genomic region of interest. We are interested in exploring whether the data support a shared causal variant for both traits. While the method is fully applicable to a case-control outcome, we consider two quantitative traits in this initial description.
SNP causality in a region of Q variants can be summarised for each trait using a vector of length Q of (0, 1) values, where 1 means that the variant is causally associated with the trait of interest and at most one entry is non-zero. A schematic illustration of this framework is provided in Figure 1 in a region that contains 8 SNPs. Each possible pair of vectors (for traits 1 and 2, which we refer to as “configuration”) can be assigned to one of five hypotheses:
In this framework, the colocalisation problem can be re-formulated as assessing the support for all configurations (i.e. pairs of binary vectors) in hypothesis .
Our method is Bayesian in the sense that it integrates over all possible configurations. This process requires the definition of prior probabilities, which are defined at the SNP level (Methods). A probability of the data can be computed for each configuration, and these probabilities can be summed over all configurations and combined with the prior to assess the support for each hypotheses . The result of this procedure is five posterior probabilities (PP0, PP1, PP2, PP3 and PP4). A large posterior probability for hypothesis 3, PP3, indicates support for two independent causal SNPs associated with each trait. In contrast, if PP4 is large, the data support a single variant affecting both traits. An illustration of the method is shown in Figure 2 for negative (Figure 2A–B, FRK gene and LDL, PP3 >90%) and positive (Figure 2C–D, SDC1 gene and total cholesterol, PP4 >80%) colocalisation results.
While the method uses Approximate Bayes Factor computations (ABF, [20], and Methods), no iterative computation scheme (such as Markov Chain Monte Carlo) is required. Therefore, computations are quick and do not require any specific computing infrastructure. Precisely, the computation time behaves as , where Q is the number of variants in the genomic region and d the number distinct associations (typically d = 2, assuming two traits and at most one causal variant per trait).
Importantly, the use of ABF enable the computation of posterior probabilities from single variant association p-values and MAFs, although the estimated single SNP regression coefficients and their variances or standard errors are preferred for imputed data.
Given the well-understood requirements for large sample size for GWAS data, we used simulations to investigate the power of our approach. We generated pairs of eQTL/biomarker datasets assuming a shared causal variant. We varied two parameters: the sample size of the biomarker dataset and the proportion of the biomarker variance explained by the shared genetic variant. We set the proportion of the eQTL variance explained by the shared variant to 10% and we used the original sample size of the liver eQTL dataset described herein [16]. Text S1 contains a description of the simulation procedure.
Results are shown in Figure 3. We find that given a sample size of 2,000 individuals for the biomarker dataset, the causal variant needs to explain close to 2% of the variance of the biomarker to provide reliable evidence in favour of a colocalised signal (lower percentile for PP4 >80%).
Until recently the assumption that, for a given GWAS signal, the causal variant in that interval had been genotyped was unrealistic. However, the application of imputation techniques [17]–[19] can provide genotype information about the majority of common genetic variants. Therefore, in situations where a common variant drives the GWAS signal, it is now plausible that, in imputed datasets, genotype information about this variant is available. Nevertheless, limited imputation quality can invalidate this hypothesis. This prompted us to investigate the implication of not including the causal variant in the genotype panel.
To address this question, we used Illumina MetaboChip data and imputed the genotyped regions using the Minimac software ([19] and Methods). We then selected only the subset of variants present in the Illumina 660K genotyping array. We simulated data under the assumption of a shared causal variant, with 4,000 individuals in the biomarker dataset. We then computed the PP4 statistic with and without restricting the SNP set to the Illumina 660K Chip SNPs (Figure 4). We also considered two different scenarios, with the causal SNP included/not included in the Illumina 660W panel (Figures S1 and S2 for more exhaustive simulations).
Our results show that when the causal variant is directly genotyped by the low density array, the use of imputed data is not essential (Figure 4A). However, in cases where the causal variant is not typed or imputed in the low density panel, the variance of PP4 is much higher (Figure 4B). In this situation, the resulting PP4 statistic tends to decrease even though considerable variability is observed. Inspection of simulation results in Figure 5 (bottom row for tagging SNP, leftmost graph for shared causal variant) shows that while PP4 tends to be lower than for its counterpart with complete genotype data (top row, leftmost graph), PP3 remains low. This indicates that more probability is given to PP0, PP1 and PP2, which can be interpreted as a loss of power rather than misleading inference in favour of distinct variants for both traits.
Statistical power may also be affected by the mode of inheritance of the causal variant. To address this, we simulated cases under a recessive pattern of inheritance. Our results show that if the true model is recessive, but the eQTL signal is nonetheless analysed using the trend test, then we will often also successfully detect a colocalised signal (Figure S9).
We compared the behaviour of our proposed test with that of proportional colocalisation testing [12], [14] in the specific case of a biomarker dataset with 10,000 samples (Figure 5, and also Figures S3 and S4). Broadly, in the case of either a single common causal variant or two distinct causal variants, our proposed method could infer the simulated hypotheses correctly (PP4 or PP3 >0.9) with good confidence, and PP3 >0.9 slightly more often than the proportional testing p-value <0.05. A key advantage in our Bayesian approach is the ability to distinguish evidence for colocalisation (i.e. high PP4) from a lack of power (i.e. high PP0, PP1 or PP2). In both of these cases (high PP4 or high PP0/PP1/PP2), the use of the proportional approach leads to failure to reject the null even though the interpretation of these situations should differ.
It has been proposed that gene expression may be subject to both global regulatory variation which acts across multiple tissues and secondary tissue specific regulators [21]. Neither approach covers this case explicitly in its construction, but it is instructive to examine their expected behaviour. The proportional approach tends to reject a null of colocalisation, suggesting that a single distinct causal variant can be sufficient to violate the null hypothesis of proportional regression coefficients. In contrast, the Bayesian approach tends to favour the shared variant in the cases covered by our simulations (median PP4 > median PP3), and either hypotheses H3 or H4 can potentially have strong support (PP4 >0.9 in close to 50% of simulations, and PP3 >0.9 in around 25% of simulations). Of course, the ultimate goal should be to extend these tests to cover multiple causal variants, but in the meantime, it can be useful to know that a high PP4 in our proposed Bayesian analysis indicates strong support for “at least one causal variant” and that rejection of the null of proportionality of regression coefficients indicates that the two traits do not share all causal variants, not that they cannot share one.
We have so far assumed that each trait is associated with at most one causal variant per locus. However, it is not unusual to observe two or more independent associations at a locus for a trait of interest [22]. In the presence of multiple independent associations, the assumption of a single variant per trait prompts the algorithm to consider only the strongest of these distinct association signals. Hence, the presence of additional associations that explain a smaller fraction of the variance of the trait, for example additional and independently associated rare variants, have a negligible impact on our computations.
To illustrate this situation, we simulated datasets with two causal variants: one colocalised eQTL/biomarker signal plus a secondary independent “eQTL only” signal (Figure S8). These simulations confirm that the PP4 statistic is only affected in the presence of two independent associations that explain a similar proportion of the variance of the trait (Figure S8).
The natural and statistically exact modification of our approach would compute, for each trait, Bayes factors for sets of SNPs rather than single SNPs (up to N SNPs jointly to accommodate for N distinct associations per trait). However, this approach has two drawbacks. Firstly, the interpretation of the resulting posterior probabilities is more challenging in situations where some but not all of the variants are shared across both traits. More importantly, the typical approach consists of publishing single variant summary statistics, which would prevent the use of standard summary statistics, a key feature of our approach.
Owing to the focus of our algorithm on the strongest association signal, an alternative approach to deal with multiple associations consists of using a stepwise regression strategy, which would then reveal the secondary association signals. Our colocalisation test can then be run on using the conditional p-values. We find this approach to be the most practical and illustrate below an application for a locus that contains several independent eQTL associations (Figure 6). In situations where only single SNP summary statistics are available, the approximate conditional meta-analysis framework proposed by Visscher et al. [23] can be used to obtain conditional p-values.
Teslovich et al. [15] reported common variants associated with plasma concentrations of low-density lipoprotein cholesterol (LDL), high-density lipoprotein cholesterol (HDL) and triglyceride (TG) levels in more than 100,000 individuals of European ancestry. They then reported the correlations between the lead SNPs at the loci they found and the expression levels of transcripts in liver. For the lipid dataset we have access only to summary statistics. The liver expression dataset used in this analysis is the same as the one used in [15]. In Teslovich et al., regions are defined within 500 kilobases of the lead SNPs, and the threshold for significance is . At this threshold, they found 38 SNP-to-gene eQTLs in liver (Supplementary Table 8 of [15]). Table S1 shows our results for these 38 previously reported colocalisations. A complete list of all our identified colocalisations (independently of previous reports) is provided in Tables S2, S3, S4, S5 (broken down by lipid traits). Using the coloc web server for this analysis with a PP4 >75, it took 1 minute to complete chromosome 1 and approximately 7 minutes to analyse the entire imputed genome-wide data on a laptop.
The majority of our results are consistent with the findings of Teslovich et al., with 26 out of 38 loci having PP4 . To assess the role of the prior, we varied the critical parameter , which codes for the prior probability that a variant is associated with both traits. Here we report the results using the . The complete list of results is provided in Table S1.
Table 1 lists the previously reported lipid-eQTL for which we find strong support against the colocalisation hypothesis (PP3 >75%). The LocusZoom association plots for each of these loci can be found in Figure S5. In addition to the loci listed in Table 1, we found strong evidence of distinct signals between HLA-DQ/HLA-DR and TC (Table S1) but these results must be interpreted with caution owing to the extensive polymorphism in the major histocompatibility complex region.
For only one locus (CEP250), we did not find a significant eQTL signal, pointing to potential differences in bioinformatics processing and/or imputation strategy. In such a situation, both PP3 and PP4 are low and PP0, PP1 and PP2 concentrate most of the posterior distribution.
Three loci (TMEM50A, ANGPTL3, PERLD1/PGAP3) do not have enough evidence to strongly support either colocalisation or absence of colocalisation (Table S1) and these should remain marked as doubtful.
One of these genes, ANGPTL3 is noteworthy. Examining this locus (Figure S6), it is clear that the pattern of association p-values is consistent between LDL and ANGPTL3 expression. However, the extent of LD is strong, with 98 strongly associated variants. In such a situation, there is uncertainty as to whether the data support a shared causal variant for both traits, or two distincts variants for eQTL/LDL. Because the data are consistent with both scenarios, the choice of prior becomes determinant. Accordingly, PP4 drops from 91% to 49% if one uses instead of .
Table 2 lists the 14 colocalised loci (15 genes) that were not reported by Teslovich et al. (or in Global Lipids Genetics Consortium [24] for the gene NYNRIN), but for which our method finds strong support for colocalisation (PP4 >75%). Figure S7 shows the LocusZoom plots for these colocalisation results. Eleven of these 15 genes are strong candidates for involvement in lipid metabolism and/or have been previously suggested as candidate genes: SDC1, TGOLN2, INHBB, UBXN2B, VLDLR, VIM, CYP26A1, OGFOD1, HP, HPR, PPARA. See Text S2 for a brief overview of the function of these genes. Four others genes have a less obvious link: CMTM6, C6orf106, CUX2, ENSG00000259359.
Three previously reported genes (SYPL2, IFT172, TBKBP1) which, based on our re-analysis, do not colocalise with the lipid traits, have a nearby gene with a high probability of colocalisation (respectively, SORT1, GCKR, KPNB1). This suggests that these genes are more likely candidates in this region. To explore the possibility that secondary signals may colocalise, we applied the stepwise regression strategy described above to deal with several independent associations at a single locus. We performed colocalisation test using eQTL results conditional on the top eQTL associated variant. Two of the loci (SYPL2/LDL or TC, APOC4 and TG) showed evidence of colocalisation with expression after conditional analysis (Table 1).
An example of this stepwise procedure for the gene SYPL2 and LDL is provided in Figure 6. We find that the top liver eQTL signal is clearly discordant with LDL association (Table 1 and Figure 6). However, conditioning on the top eQTL signal reveals a second independent association for SYPL2 expression in liver. This secondary SYPL2 eQTL colocalises with the LDL association (PP4 >90%, Figure 6).
We developed a web site designed for integration of GWAS results using only p-values and the sample size of the datasets (http://coloc.cs.ucl.ac.uk/coloc/). The website was developed using RWUI [25]. Results include a list of potentially causal genes with the associated PP4 with their respective plots and ABF, and can be viewed either interactively or returned by email.
Researchers can request a genome-wide scan of results from a genetic association analysis, and obtain a list of genes with a high probability of mediating the GWAS signals in a particular tissue. The tool also allows visualisation of the signals within a genetic region of interest.
The database and browser currently include the possibility of investigating colocalisation with liver [15] and brain [26], [27] expression data, however the resource will soon be extended to include expression in different tissues. This method, as well as alternative approaches for colocalisation testing [12], [14], are also available with additional input options in an R package, coloc, from the Comprehensive R Archive Network (http://cran.r-project.org/web/packages/coloc).
We have developed a novel Bayesian statistical procedure to assess whether two association signals are colocalised. Our method is best suited for associations detected by GWAS, which are likely to reflect common, imputable, variations with small effects, or a rare variants with large effect sizes. Our aim differs from a typical fine-mapping exercise in the sense that we are not interested in knowing which variant is likely to be causal but only whether a shared causal variant is plausible. The strength of this approach lies in its speed and analytical forms, combined with the fact that it can use single variant p-values when only these are available.
Our results show that to provide an accurate answer to the colocalisation problem, high-density genotyping and/or accurate use of imputation techniques are key. The quality of the imputation is another important parameter. Indeed, while the variance of the regression coefficient can be estimated solely on the basis of the minor allele frequency for typed SNPs and sample size (and the case control ratio in the case of a binary outcome) [17], [28], this ignores the uncertainty due to imputation. Filtering out poorly imputed SNPs partially addresses this problem, with the drawback that it may exclude the causal variant(s). Hence, providing estimates of the variance of the MLE, together with the effect estimates, will result in greater accuracy. This additional option is available on the coloc package in R (http://cran.r-project.org/web/packages/coloc).
We currently assume that each genetic variant is equally likely a priori to affect gene expression or trait. A straightforward addition to our methodology would consider location specific priors for each variant, which would depend for example on the distance to the gene of interest, or the presence of functional elements in this chromosome region [29]. Our computation of the BF also assumes that, under , the effect sizes of the shared variant on both traits are independent. This could be modified if, for example, one compares eQTLs across different tissue types, or the same trait in two different studies. [30] has proposed a framework to deal with correlated effect sizes, and these ideas could potentially be incorporated in our colocalisation test.
Another related issue is the choice of prior probabilities for the various configurations. For the eQTL analysis, we used a prior probability for a cis-eQTL. A more stringent threshold may be better suited for trans-eQTLs where the variants are further away from the gene under genetic control. We also used a prior probability of for the lipid associations. Although our knowledge about this is still lacking, this estimate has been suggested in the literature in the context of GWAS [20], [31], [32]. We assigned a prior probability of for , which encodes the probability that a variant affects both traits. It has been shown that SNPs associated with complex traits are more likely to be eQTLs compared to other SNPs chosen at random from GWAS platforms [33], and a higher weighting for these SNPs has been proposed when performing Bayesian association analyses [34], [35]. Also, eQTLs have been shown to be enriched for disease-associated SNPs when a disease-relevant tissue is used [9], [36]. Our sensitivity analysis for the parameter showed broadly consistent results (Table S1). In cases where GWAS data are available for both traits, [10] show that it is possible to estimate these parameters from the data using a hierarchical model. This addition is a possible extension of our approach.
The interpretation of the posterior probabilities requires caution. For example, a low PP4 may not indicate evidence against colocalisation in situations where PP3 is also low. It may simply be the result of limited power, which is evidenced by high values of PP0, PP1 and/or PP2. Moreover, a high PP4 is a measure of correlation, not causality. To illustrate this, one can consider the relatively common situation where a single variant appears to affect the expression of several genes in a chromosome region (as observed, for example, in the region surrounding the SORT1 gene). Several eQTLs will be colocalised, both between them and with the biomarker of interest. In this situation one would typically expect that a single gene is causally involved in the biomarker pathway but the colocalisation test with the biomarker will generate high PP4 values for all genes in the interval.
We show that we can use conditional p-values to deal with multiple independent associations with the same trait at one locus. While we found this solution generally effective, Wallace [14] points out that this top SNP selection for the conditional analysis can create biases, although the bias is small in the case of large samples and/or strong effects. For difficult loci with multiple associations for both traits and available genotype data, it may be more appropriate to estimate Bayes factors for sets rather than single variants in order to obtain an exact answer. This extension would avoid the issue of SNP selection for the conditional analysis.
Importantly, GWAS signals can be explained by eQTLs only when the causal variant affects the phenotype by altering the amount of mRNA produced, but not when the phenotype is affected by changing the type of protein produced, although the former seems to be the most common [33]. Furthermore, since many diseases manifest their phenotype in certain tissues exclusively [2], [21], [37], [38], colocalisation results will be dependent on the expression dataset used. In addition to identifying the causal genes, the identification of tissue specificity for the molecular effects underlying GWAS signals is a key outcome of our method. We anticipate that building a reference set of eQTL studies in multiple tissues will provide a useful check for every new GWAS dataset, pointing directly to potential candidate genes/tissue types where these effects are mediated.
While this report focuses on finding shared signals between a biomarker dataset and a liver expression dataset, we plan to utilise summary results of multiple GWAS and eQTL studies, for a variety of cell types and traits. In fact, our method can utilise summary results from any association studies. Disease/disease, (cis or trans) eQTL/disease or disease/biomarkers comparisons are all of biological interest and use the same statistical framework. We expect that the fact that the test can be based on single SNP summary statistics will be key to overcome data sharing concerns, hence enabling a large scale implementation of this tool. The increasing availability of RNA-Seq eQTL studies will further increase the opportunity to detect isoform specific eQTLs and their relevance to disease studies. Owing to the increasing availability of GWAS datasets, the systematic application of this approach will potentially provide clues into the molecular mechanisms underlying GWAS signals and the aetiology of the disorders.
This paper re-analyses previously published datasets. All samples and patient data were handled in accordance with the policies and procedures of the participating organisations.
We used in our analysis gene expression and genotype data from 966 human liver samples. The samples were collected post-mortem or during surgical resection from unrelated European-American subjects from two different non-overlapping studies, which have been described in [16]. The cohorts were both genotyped using Illumina 650Y BeadChip array, and 39,000 expression probes were profiled using Agilent human gene expression arrays. All of the expression data has been normalised as one unit even though they were part of different studies, since high concordance between data generated using the same array platforms has been previously reported. Probe sequences were searched against the human reference genome GRCh37 from 1000 Genomes using BLASTN. Multiple probes mapping to one gene were kept in order to examine possible splicing. The probes were kept and annotated to a specific gene if they were entirely included in genes defined by Ensembl ID or by HGNC symbol using the package biomaRt in R [39]. After mapping and annotating the probes, we were left with 40,548 mapped probes covering 24,927 genes.
Quality control filters were applied both before and after imputation. Before imputation, individuals with more than 10% missing genotypes were removed, and SNPs showing a missing rate greater than 10%, a deviation for HWE at a p-value less than 0.001 were dropped. After imputation, monomorphic SNPs were excluded from analyses.
To speed up the imputation process, the genome was broken into small chunks that were phased and imputed separately and then re-assembled. This was achieved using the ChunkChromosome tool (http://genome.sph.umich.edu/wiki/ChunkxChromosome), and specifying chunks of 1000 SNPs, with an overlap window of 200 SNPs on each side, which improves accuracy near the edges during the phasing step. Each chunk was phased using the program MACH1 with the number of states set to 300 and the number of rounds of MCMC set to 20 for all chunks. Phased haplotypes were used as a basis for imputation of untyped SNPs using the software Minimac with 1000 Genomes European ancestry reference haplotypes (phase1 version 3, March 2012) to impute SNPs not genotyped on the Illumina array. Variants with a MAF less than 0.001 were also excluded post-imputation. The data was then collated in probability format that can be used by the R Package snpStats [39].
eQTL p-values, effect sizes, and standard errors were obtained by fitting a linear trend test regression between the expression of each gene and all variants 200 kilobases upstream and downstream from each probe. After filtering out the variants with MAF <0.001, monomorphic SNPs, multi-allelic SNPs (as reported in 1000 Genomes or in the Ensembl database) and variants not sufficiently well imputed (Rsq <0.3, as defined by minimac http://genome.sph.umich.edu/wiki/minimac) between both datasets, we applied our colocalisation procedure. We conducted conditional analysis on SNPs with p-values for the expression associations, and repeated the colocalisation test using expression data conditioned on the most significant SNP. The aim of this analysis is to explore whether additional signals for expression other than the main one are shared with the biomarker signal.
The biomarker p-values from the meta-analyses (with genomic control correction) were obtained from a publicly available repository (http://www.sph.umich.edu/csg/abecasis/public/lipids2010/).
The regional association plots for the eQTL and Biomarker datasets were created using LocusZoom [40] (http://csg.sph.umich.edu/locuszoom/).
We call a “configuration” one possible combination of pairs of binary vectors indicating whether the variant is associated with the selected trait. We can group the configurations into five sets, , , , , , containing assignments of all SNPs Q to the functional role corresponding to the five hypothesis , , , , . We can compute the posterior probabilities given the data for each of these 5 hypothesis by summing over the relevant configurations:(1)where P(S) is the prior probability of a configuration, is the probability of the observed data D given a configuration S, and the sum is over all configurations S which are consistent with a given hypothesis , where h = (1,2,3,4). Thus, the probability of the data given a configuration is weighted by the prior probability of that configuration.
Next, to avoid computing the proportionality constant in Equation 1, we can reformulate the posterior probability for each hypothesis by writing this quantity as a ratio. For example, the posterior probability under hypothesis 4, dividing each of these terms by the baseline , is:(2)
The ratios in the numerator and denominator of equation 2 are:(3)The first ratio inside the sum in this equation is a Bayes Factor (BF) for each configuration, and the second ratio is the prior odds of a configuration compared with the baseline configuration . The BF can be computed for each variant from the p-value, or estimated regression coefficient and variance of , using Wakefield's method. By summing over all configurations in we are effectively comparing the support in the data for one alternative hypothesis versus the null hypothesis. An in-depth description of the method making use of the current assumptions can be found in Text S1.
A Bayes Factor for each SNP and each trait 1 and 2 was computed using the Approximate Bayes Factor (ABF, [20]). Wakefield's method yields a Bayes factor that measures relative support for a model in which the SNP is associated with the trait compared to the null model of no association.
The equation used is the following:(4)where is the usual Z statistic and the shrinkage factor r is the ratio of the variance of the prior and total variance (). Assuming a normal distribution, the p-value of each SNP can be converted to standard one-tailed Z-score by using inverse normal cumulative distribution function. So for a SNP, all that it is needed are the p-values from a standard regression output, and , the standard deviation of the normal prior N(0,W) on . The variance of the effect estimate, V, can be approximated using the MAF and sample size. However for imputed data it is preferable to use the variance outputted in standard regression analysis directly in the ABF equation. For the expression dataset used here, the variance and effect estimates from the regression analysis were used for computation of ABFs (see Text S1 for more details).
Prior probabilities are assigned at the SNP level and correspond to mutually exclusive events. We assigned a prior of for and , the probability that a SNP is associated with either of the two traits. Since all SNPs are assumed to have the same prior probability of association, this prior can be interpreted as an estimate for the proportion of SNPs that we expect to be associated with the trait in question. We also assigned a prior probability of for , the probability that one SNP is associated with both traits. This probability can be better understood when it is re-expressed as the conditional probability of a SNP being associated with trait 2, given that it is associated with trait 1. So assigning a probability of means that 1 in 100 SNPs that are associated with trait 1 is also associated with the other. As a sensitivity analysis, we ran the comparison with Teslovich et al. using two other prior probabilities for , which means 1 in 50 SNPs that are associated with one trait is also associated with the other; and which means 1 in 10 SNPs.
To compute the ABF, we also needed to specify the standard deviation for the prior, and we set this to 0.20 for binary traits and 0.15 for quantitative traits (more details in Text S2).
|
10.1371/journal.pntd.0000231 | Human TLR1 Deficiency Is Associated with Impaired Mycobacterial Signaling and Protection from Leprosy Reversal Reaction | Toll-like receptors (TLRs) are important regulators of the innate immune response to pathogens, including Mycobacterium leprae, which is recognized by TLR1/2 heterodimers. We previously identified a transmembrane domain polymorphism, TLR1_T1805G, that encodes an isoleucine to serine substitution and is associated with impaired signaling. We hypothesized that this TLR1 SNP regulates the innate immune response and susceptibility to leprosy. In HEK293 cells transfected with the 1805T or 1805G variant and stimulated with extracts of M. leprae, NF-κB activity was impaired in cells with the 1805G polymorphism. We next stimulated PBMCs from individuals with different genotypes for this SNP and found that 1805GG individuals had significantly reduced cytokine responses to both whole irradiated M. leprae and cell wall extracts. To investigate whether TLR1 variation is associated with clinical presentations of leprosy or leprosy immune reactions, we examined 933 Nepalese leprosy patients, including 238 with reversal reaction (RR), an immune reaction characterized by a Th1 T cell cytokine response. We found that the 1805G allele was associated with protection from RR with an odds ratio (OR) of 0.51 (95% CI 0.29–0.87, p = 0.01). Individuals with 1805 genotypes GG or TG also had a reduced risk of RR in comparison to genotype TT with an OR of 0.55 (95% CI 0.31–0.97, p = 0.04). To our knowledge, this is the first association of TLR1 with a Th1-mediated immune response. Our findings suggest that TLR1 deficiency influences adaptive immunity during leprosy infection to affect clinical manifestations such as nerve damage and disability.
| Mycobacterium leprae (ML) causes a disabling and stigmatizing disease that is characterized by distinct immune responses. ML produces a spectrum of illness in humans, and several lines of evidence indicate that host genetic factors influence susceptibility and clinical manifestations. Leprosy can occur as the lepromatous or tuberculoid forms, which are associated with different clinical manifestations, histopathology, T cell cytokine profiles, and bacterial burden in affected sites. Leprosy is also associated with unique immunologic reactions, such as reversal reaction, which is characterized by the rapid development of a Th1 T cell cytokine response that can cause substantial morbidity. We and others recently discovered a common human polymorphism in TLR1 (T1805G, I602S) that regulates cytokine production in response to lipopeptide stimulation, influences the cellular innate immune response to Mycobacteria, is associated with altered localization, and is present in 50% of individuals worldwide. Here, we show that in humans the 1805G variant does not mediate an inflammatory response to ML in vitro and that this polymorphism is associated with protection from reversal reaction. These data suggest that a common variant of TLR1 is associated with altered adaptive immune responses to ML as well as clinical outcome.
| Mycobacterium tuberculosis (MTb) has established latent infection in one-third of the world's population and, among those with progressive disease, causes about 2 million deaths per year [1]. Mycobacterium leprae (ML), a related organism, is the etiologic agent of leprosy, an ancient scourge that still causes illness in several regions of the world [2]. Both MTb and ML produce a spectrum of illness in their hosts, yet, aside from frank immunodeficiency, the host factors that underlie the various clinical manifestations of MTb and ML infection are largely unknown. One possible explanation for this diversity of outcomes is common, subclinical variation in host defense genes.
Several lines of evidence suggest that genetic factors influence susceptibility to leprosy and other Mycobacteria [1]–[5]. Rare individuals with primary immunodeficiency syndromes are highly susceptible to certain mycobacterial species due to Mendelian disorders associated with highly penetrant phenotypes [2]. However, in most individuals, susceptibility to mycobacterial infection is associated with complex inheritance patterns that are determined by the combined effects of variation across many genes, with a modest contribution from each polymorphism. Evidence that commonly inherited gene variants influence susceptibility to mycobacterial infection comes from twin studies, genome-wide linkage studies, and candidate gene association studies [2]. Studies of leprosy infection in twins have shown a three-fold greater concordance for type of leprosy disease in monozygotic compared to dizygotic twins [6]. Genome-wide linkage studies have identified two single nucleotide polymorphisms (SNPs) in the shared promoter region of the PARK2 and the PARCG gene, several HLA-DR2 alleles, and a non-HLA region near chromosome 10p13 that are associated with leprosy or leprosy subtypes [2],[7],[8]. Candidate gene association studies have also shown associations between leprosy and polymorphisms in several genes, including lymphotoxin-a (LTA) [9], the vitamin D receptor [10], TNF-α [11], laminin-2 [12], and mannose binding lectin [13],[14].
Human infection with M. leprae presents a unique opportunity to link innate and adaptive immune responses to host genetic factors. Leprosy's divergent clinical forms reflect two distinct immune responses to the same pathogen. Lepromatous leprosy (defined as polar lepromatous (LL) or borderline lepromatous (BL)) is characterized by a Th2 immune response and poor containment of the infection. At the opposite pole, tuberculoid leprosy (defined as polar tuberculoid (TT) or borderline tuberculoid (BT)) features a Th1 cytokine response, vigorous T cell responses to ML antigen, and containment of the infection in well-formed granulomas [15],[16]. Reversal reactions (RR) represent the sudden activation of a Th1 inflammatory response to ML antigens. They often occur after the initiation of treatment in patients towards the lepromatous pole of the leprosy spectrum (LL, BL, or borderline borderline (BB) categories) and reflect a switch from a Th2-predominant cytokine response toward a Th1-predominant response [15],[16]. Risk factors for RR intrinsic to the host include age [17] and gene variants, although the latter have not been intensively investigated [18],[19].
Toll-like receptors (TLRs) are a family of highly conserved, type 1 transmembrane proteins that orchestrate the innate immune response to microbial motifs, also known as pathogen associated molecular patterns (PAMPs) [20]–[22]. The TLR pathway regulates the innate immune response to mycobacteria through several TLRs, including TLR1, 2, 4, 6, and 9 [23]–[26]. TLR2 (Online Mendelian Inheritancce in Man (OMIM):603028), as a heterodimer with TLR1 (OMIM: 601194) or TLR6 (OMIM: 605403), mediates recognition of several mycobacterial motifs, including lipopeptides, the 19 kDa protein, lipoarabinomannan (LAM) [1],[27],[28]. Interaction of these ligands with the extracellular domain of TLRs leads to activation of a signaling pathway, which results in expression of chemokines and cytokines [20]. Functional work by many investigators has shown that TLR2 is a critical mediator of the innate immune response to ML and MTb [29],[30]. In addition, several TLR2 polymorphisms have been reported to be associated with susceptibility to MTb [31]–[33].
By contrast, very little is known about the effect of TLR1 variation on the innate response to mycobacteria or clinical susceptibility to mycobacterial disease. It also remains controversial whether and how the innate immune response mediated by any individual TLR shapes adaptive immunity [34]–[37]. We recently characterized a non synonymous SNP, T1805G (I602S), in the transmembrane domain of TLR1 that regulates signaling in response to PAM3, a synthetic ligand of TLR1 [38]. Johnson et al. also found that this polymorphism was associated with decreased signaling as well as protection from leprosy in Turkey [39]. Intriguingly, it appears that the TLR1 signaling defect is due to a complete absence of TLR1 on the surface of monocytes in GG individuals [39]. Here, we investigate an association of this SNP with different clinical forms of leprosy in Nepal and examine the effect of this SNP on leukocyte signaling in response to ML stimulation.
RPMI Medium 1640, L-glutamine, penicillin-streptomycin, and DMEM were from GIBCO/Invitrogen (Carlsbad, CA). Ultrapure lipopolysaccharide (LPS) was from Salmonella minnesota R595 (List Biological Labs, Inc.). Lipopeptides PAM2Cys-SKKKK (diacylated, PAM2) and PAM3Cys-SKKK (triacylated, PAM3) were from EMC Microcollections (Tuebingen, Germany). Macrophage-activating Lipopeptide-2 S-[2,3-bis(Palmityloxy)-(2R)-propyl-cysteinyl-GNNDESNISFKEK] (diacylated lipopeptide from Mycoplasma fermentens, Malp-2) was obtained from Alexis Biochemicals (Lausen, Switzerland). M. leprae reagents were obtained from J. Spencer (Colorado State University) through NIH, NIAID Contract No. NO1-AI-25469, entitled “Leprosy Research Support.” HEK293 cells (ATCC#CRL-1573) were grown in DMEM (GIBCO cat. no. #11995), supplemented with 10% fetal bovine serum, 10 units/ml penicillin, and 10 µg/ml streptomycin.
Study participants in Seattle were healthy adults with no known history of unusual susceptibility to infections [40]. Study participants in Nepal included 933 leprosy patients referred for treatment at Anandaban Hospital in Katmandu, Nepal and later recruited to a study of genetic factors influencing susceptibility to reactional episodes in leprosy. The study population comprised more than 8 different ethnic and religious groups that included Brahmin (25.6%), Chhetri (22.3%), Tamang (14.3%), Newar (7.3%), Magar (5.4%), Muslim (3.3%), Sarki (3.5%), and Kami (2.7%), with 15.5% having unrecorded ethnicity A diagnosis of leprosy and determination of leprosy type was made by clinical symptoms, skin smears and biopsy reports. Assignment of leprosy category followed the Ridley/Jopling classification scheme [41]. Each patient had a minimum of three years of regular clinic follow-up prior to recruitment. In accord with guidelines of the US Department of Health and Human Services, protocols were approved by the Nepal Health Research Council, the University of Washington, the University of Medicine and Dentistry of New Jersey, and the Western Institutional Review Board. Written informed consent was obtained from all patients or from their relatives if the patient could not provide consent.
DNA from subjects in Nepal was obtained by extraction from whole blood using Nucleon BACC2 Genomic DNA (Amersham Lifesciences) and Roche High-Pure PCR template preparation extraction kits. DNA from subjects in Seattle was extracted from whole blood using QIAamp DNA Blood Midi kits (Qiagen, Valenica, CA). Genotyping was carried out with a MassARRAY technique (Sequenom) as previously described [42],[43]. For functional studies, the coding region of TLR1 was amplified from genomic DNA and cloned into the pEF6/V5-His-TOPO vector (Invitrogen, Carlsbad, CA) as previously described [38]. To obtain the polymorphic variants of TLR1, a whole plasmid PCR strategy with mutant primers was used as previously described [30].
PBMCs were derived from whole blood separated by centrifugation on a Ficoll-Hypaque gradient, plated at a density of 1×105 cells per well in 96-well plates in RPMI (supplemented with 10% fetal bovine serum), and incubated overnight. PBMC cytokine assays were then performed by stimulating with various TLR ligands or extracts of M. leprae for 18 hours. Each sample was assayed in triplicate. Cytokine levels were determined with a sandwich ELISA technique (Duoset, R&D Systems, Minneapolis, MN) or with a multiplex kit for the luminex platform (Human Fluorokine MAP Base Kit, Panel A, R&D systems, Minneapolis, MN). Levels of contaminating LPS as determined by the chromogenic Limulus amebocyte lysate test (Cambrex, MD) were 0.05–0.27, and 0.03–0.19 endotoxin units/ml, in wells incubated with whole irradiated ML (ML) and ML cell wall (MLcw), respectively, depending on the dose of reagent used. These values correspond to 4.5–27.2 and 3.04–19.4 pg/ml of endotoxin, respectively, in wells treated with whole irradiated ML and ML cw. All wells that received ML reagents were additionally treated with polymyxin B at a concentration of 10 µg/ml.
HEK293 cells were transfected with Polyfect (Qiagen, Hilden, Germany) per the manufacturer's instructions with 2–5×104 cells per well in a 96-well plate with pRL-TK (to control for transfection efficiency), ELAM-luciferase (NF-κB reporter), one of two TLR1 variants, TLR2, and CD14. After an overnight transfection, cells were stimulated with TLR ligands or extracts of ML for 4–6 hours, and then lysed and processed for luciferase readings per the manufacturer's instructions for the Dual Luciferase Reporter Assay System (Promega, Madison, WI).
Univariate analysis was performed for categorical variables with a Chi-Square test; Fisher's exact test was used when the number of samples in a group was less than 5. The Mann-Whitney U-test was used to make comparisons of the cytokine production between groups, as small sample sizes precluded an assumption of normal distribution. Student's t-test was used to compare results in the luciferase assay. Two-sided testing was used for all comparisons to evaluate statistical significance. A P value (p) of ≤0.05 was considered significant. Statistics were calculated with Prism version 4.03software. For genetic analysis, allelic, genotypic, and haplotypic frequencies were compared between groups. Haplotypes were constructed with an Expectation/Maximization (EM) algorithm with the program HAPIPF in IC Stata (version 10.0) [44]. Except for minor deviations, the observed allelic frequencies of SNPs were consistent with expected frequencies under Hardy-Weinberg equilibrium.
We investigated the effect of SNP T1805G (I602S) on NF-κB responses to ML in HEK293 cells transfected with 1805T (602I) or 1805G (602S), firefly luciferase conjugated to an NF-κB promoter (ELAM), TLR2, CD14, and Renilla luciferase conjugated to thymidine kinase to control for transfection efficiency [38]. Cells were then stimulated with media, ML extracts or TLR ligands: PAM3, a ligand for the TLR2/1 heterodimer or Malp-2, a ligand for the TLR2/6 heterodimer (Fig 1). HEK293 cells transfected with TLR2+1805T and stimulated with 50 µg/ml of whole, irradiated ML had significantly greater NF-κB activity than HEK293 cells transfected with TLR2 alone (600.7 vs. 159.8 relative luciferase units (RLU), p = 0.001), or cells transfected with TLR2+1805G (600.7 vs. 157.1 RLU, p = 0.000004) (Fig 1). Responses to ML were dose-dependent in cells transfected with either TLR1 variant and the signaling difference between 1805T and 1805G transfectants persisted over a range of doses (comparison for ML 5 µg/ml: 511.9 vs. 75.8, p<0.0001; comparison for ML 250 µg/ml: 673.7 vs. 262.8, p = 0.0005). We then investigated whether T1805G influenced signaling in response to MLcw, which contains lipopeptide moieties known to stimulate through TLR2/1 [29]. TLR2+1805T-transfected cells stimulated with 1 or 10 µg/ml of MLcw had significantly greater NF-κB activity than cells transfected with TLR2+1805G (Fig 1, comparison for MLcw 1 µg/ml: 444.4 vs. 53.8 RLU, p = 0.00005; comparison for MLcw 10 µg/ml: 562.8 vs. 238.1 RLU, p = 0.004). As a control, we also compared baseline signalling activity and response to tri-acylated lipopeptide (PAM3) in the two 1805 variants. Consistent with previous observations [38], the 1805T variant, when co-transfected with TLR2, mediated greater constitutive NF-κB activity compared to TLR2 alone (Fig 1, stimulation with media alone: 329.7 vs. 3.9 RLU, p = 0.000003). The TLR2+1805T transfectants were also readily distinguished from the TLR2+1805G transfectants by a significantly higher level of basal signaling (329.7 vs. 24.2 RLU, p = 0.000002). In addition, TLR2+1805T-transfected cells stimulated with PAM3 had significantly greater NF-κB activity compared to cells transfected with TLR2+1805G (841.4 vs. 383.7 RLU, p = 0.0009). In contrast, responses to Malp-2 did not significantly differ between the two variants (783.1 vs. 650.6 RLU, p = 0.37). Together, these results suggest that TLR1 variant 1805G leads to impaired innate immune responses to ML because of a defect in basal signaling of the TLR2/1 heterodimer.
We next examined whether TLR1_T1805G regulates innate immune responses to ML in human primary immune cells. We obtained PBMC from whole blood from 28 healthy individuals whose genotypes for TLR1_T1805G had previously been determined [38]. The PBMCs were then stimulated with whole, irradiated ML, MLcw, and a variety of TLR ligands, including PAM3, PAM2, and LPS. We compared the high (1805TT) and medium (1805TG) responding genotypes with the low responding genotype (1805GG) (Fig 2). When stimulated with MLcw, PBMCs from 1805TT or 1805TG (1805TT/TG) donors showed significantly greater IL-6 responses compared to 1805GG PBMCs (Fig 2B, for MLcw at 2 µg/ml: 5,950 vs. 2,198 pg/ml, p = 0.0005; for MLcw at 10 µg/ml: 6,115 vs. 3,320 pg/ml, p = 0.0076). Similarly, responses to whole, irradiated ML were significantly higher in 1805TT/TG PBMCs compared to 1805GG PBMCs (Fig 2B, for whole, irradiated ML at 20 µg/ml: 3,310 vs. 1,649 pg/ml, p = 0.0005; for whole, irradiated ML at 100 µg/ml: 6,183 vs. 2,246 pg/ml, p = 0.0017). As previously observed, after stimulation with PAM3, significantly higher levels of IL-6 were seen in PBMCs heterozygous or homozygous for 1805T compared to1805GG PBMCs (Fig 2A, for TT/TG genotypes vs. GG genotypes stimulated with 75 µg/mL PAM3: 3,966 vs. 1,491 pg/ml, p = 0.0007). Stimulation with LPS and PAM2, ligands with specificity for TLR4 and TLR2/6, respectively, produced no significant differences in IL-6 production between the two groups (Fig 2A). We also assessed the levels of other cytokines important in the monocyte immune response to mycobacteria and found that IL-1β production in PBMCs from 1805TT/TG donors stimulated with PAM3 or whole irradiated ML was significantly higher than in 1805GG PBMCs (Fig 3A). IL-1β levels did not differ between the two groups when PBMCs were stimulated with PAM2 or with LPS (TLR2/6 and TLR4 ligands, respectively) controls. Production of TNF-α, similarly, was significantly higher in 1805TT/TG PBMCs stimulated with PAM3, whole irradiated ML, or MLcw compared to 1805GG PBMCs (Fig 3B), but did not differ between the two groups after stimulation with Pam2 or LPS (TLR2/6 and TLR4 controls) (Fig 3B). Interestingly, there were no differences in IL-1β levels between TT/TG and GG groups after PBMC stimulation with MLcw, in contrast to the pattern seen with other cytokines (Fig 2B, 3A).
Although recently published data suggests that TLR1_T1805G is associated with susceptibility to leprosy, associations with different types of leprosy or immunologic reactions have not been previously examined. To determine whether the TLR1_T1805G polymorphism was associated with different forms of leprosy or leprosy immune reactions, 933 patients from Anandaban Hospital in Kathmandu, Nepal were enrolled in a retrospective study. Of this total, 581 had polar lepromatous (LL), borderline leprosy (BL) or borderline borderline (BB) and 343 had tuberculoid leprosy (including borderline tuberculoid (BT) and polar tuberculoid (TT)). A total of 344 patients experienced immune reactions during 3 years of regular visits to a leprosy clinic, of whom 238 had RR and 108 had erythema nodosum leprosum (ENL) and 2 had both reactions. The baseline characteristics of this population are described in Table 1. When individuals with lepromatous leprosy were compared to those with tuberculoid leprosy, there were no significant associations between SNP T1805G and either form of leprosy, either at the allelic or genotypic level of analysis (Table 2). However, there was a trend toward an association of the TG or GG genotypes with lepromatous leprosy (OR [odds ratio] 4.76, 95% CI [95% confidence interval] 0.58–38.87, p = 0.11) in comparison to tuberculoid leprosy that did not reach significance. We also genotyped five additional TLR1 SNPs that are contained in common TLR1 haplotypes. We did not find associations of any of these SNPs with leprosy type (Table 2). Analysis of TLR1 haplotypes generated from these six SNPs similarly yielded no association with leprosy type (data not shown).
We next investigated whether T1805G (I602S) might be associated with ENL or RR, a Th1-mediated immune event clinically manifested by inflamed skin lesions, fever, and neuritis. There was no association of T1805G or any other TLR1 polymorphisms with ENL. In contrast, the 1805G allele was associated with a reduced risk of developing RR in comparison to the 1805T allele (Table 3). The allele frequency of 1805G was 3.9% in those with RR versus 7.4% in those without (unadjusted OR 0.51, 95% CI 0.29–0.87, p = 0.01). The distribution of the genotype frequencies was also significantly different with a p value of 0.05 (Table 3). We next examined whether this association was affected by population admixture. This cohort contains representatives of more than 8 different ethnic groups with the majority belonging to one of four groups (Brahmin, Chhetri, Tamang, and Newar). There were no significant differences in frequencies of leprosy type or immunologic reactions among the different ethnic groups (Table 1). We performed a multivariate logistic regression, adjusting for ethnicity, and found that the odds ratio remained significant (OR 0.52, 95% CI 0.30–0.92, p = 0.03). We also adjusted for ethnicity, sex and age as a continuous variable, and again found that the odds ratio remained significant (OR 0.54, 95% CI, 0.30–0.96, p = 0.04).
Previously, we found that TG individuals are intermediate between TT and GG individuals in responses to TLR1 stimulation [38]. We therefore investigated the influence of TG heterozygotes with a recessive (assumes T is recessive to G and compares TT versus TG/GG frequencies) or a dominant model (assumes T is dominant over G and compares TT/TG versus GG frequencies). In the recessive model, we found that the 1805 TG/GG genotypes were associated with a lower likelihood of RR compared to the TT genotype (OR 0.55, 95% CI 0.31–0.97, p = 0.04). In the dominant model, the GG genotype was associated with a non-significant reduction in risk when compared to the TT/TG genotypes (OR 0.15, 95%CI 0.01–2.62, p = 0.12).
We next examined other TLR1 polymorphisms to determine whether there were any additional associations between individual SNPs or haplotypes and RR. SNP rs5743592, located in intron 2 adjacent to the 5′ UTR of TLR1, was found to be associated with a modestly increased risk of RR (Table 3, OR for allelic comparison: 1.29, 95% CI 1.02–1.64, p = 0.04). When haplotypes of 6 TLR1 SNPs were examined (Table 4), one haplotype, TATTAG, was associated with protection from RR (OR 0.55, 95% CI 0.31–0.97, p = 0.05). None of the other five haplotypes had any association with RR. Haplotype TATTAG was the only haplotype occurring with a frequency greater than 1% that contained the 1805G allele. Lastly, we examined whether haplotypes formed from SNPs rs5743592 and 1805 were associated with altered risk of RR. The haplotype associations were consistent with the individual effect of each SNP on the risk RR (Table 4), without any additive or synergistic effects. Together, these genetic data demonstrate that TLR1 SNP 1805G is associated with protection from RR.
In this manuscript, we demonstrate that a human TLR1 SNP regulates the innate immune response to ML and is associated with protection from RR. This is the first study to describe an association of a TLR1 SNP with a Th1-mediated adaptive immune response. One weakness of our study is the low frequency of the 1805G variant in Nepal, which limited our power to detect associations with leprosy type. However, the relevance of the 1805G SNP to leprosy pathogenesis is supported by work from Johnson and colleagues, who recently reported an association of this variant with protection from leprosy in a Turkish cohort [39]. Although these authors do not mention whether or not T1805G was associated with different forms of leprosy in Turkey or with immune reactions, this may be due to the small size of their patient cohort (57 individuals). Genetic association studies that utilize cohorts of multiple ethnicities are also open to the criticism that associations are due to the effects of population admixture rather than the variant of interest. However, when we adjusted for ethnic composition of the comparision groups, we still found that the 1805G variant was associated with significant protection against RR.
There are several possible mechanisms by which TLR1 might affect the pathogenesis of RR. At the cellular level, TLR1 might exert its influence through control of innate immune functions, such as the capacity of dendritic cells (DCs) and macrophages to control bacillary replication. Alternatively, or in addition, SNP 1805 may regulate DC maturation and/or antigen presentation and thereby influence the activation and maintenance of T cell responses to M. leprae antigens. Interestingly, recent work by other investigators demonstrates that the differentiation of monocytes into mature antigen presenting cells (APCs) is mediated by TLR signaling [35],[36]. For example, the 19 kDa protein of MTb signals through TLR2/1 to downregulate MHC class I and II antigen presentation by macrophages, leading to impaired T cell activation [45]. We have previously shown that ML also exerts an inhibitory effect on APC activation and maturation, through an as-yet unidentified mechanism [46]. In LL patients who are clinically stable, macrophages within LL lesions contain numerous bacilli that are seemingly resistant to host killing [47],[48]. However, this inhibition of phagocyte function seems to be overturned during RR. When patients undergo RR, the bacilli within these macrophages are rapidly cleared. This clearance coincides with the influx into the lesion of CD1b+ DC, which activate M. leprae-specific T cells and thereby promote intracellular killing. Importantly, the generation of CD1b+ DCs appears to be dependent on signaling through TLR2/1 [48]. Here, we show that individuals carrying the 1805G SNP are protected against reversal reactions. Our data suggests that TLR1 may be an important regulator of these effects on DCs.
At the molecular level, the defect in 1805G signaling is likely due to a failure to express or retain TLR1 on the cell surface. Johnson recently showed that monocytes from 1805GG individuals completely lack surface TLR1, although total levels of this receptor are normal [39]. In Nepal, we observed a trend toward an association of the 1805G variant with lepromatous rather than tuberculoid leprosy. This trend, although not statistically significant, is consistent with impaired Th1 immunity, which is required for the tuberculoid form of the disease. An association of 1805G with lepromatous leprosy could explain the intriguing earlier observation by Krutzik and coworkers, who examined TT and LL lesions and were unable to detect any TLR1 staining in LL lesions [29]. Collectively, these findings suggest that TLR1 biology is different in lepromatous leprosy than in other forms of the disease. The absence of membrane-inserted TLR1 in 1805GG individuals may be associated with a Th2 immune response that arises by default in the absence of robust Th1 cytokine responses. This Th2 bias may in turn permit continued replication of the M. leprae bacillus and result in the clinical phenotype of LL.
In our initial characterization of TLR1_T1805G, we found that this polymorphism is present in up to 76% of Caucasian Americans and is associated with a defect in innate responses to bacterial lipopeptide [38]. Worldwide, the T1805G polymorphism has variable frequency across ethnic groups. In Turkey, the allele frequency of this variant is 43% [39], while among African Americans and Vietnamese individuals, it has a frequency of 25% and 1%, respectively [38]. A broad array of pathogens are sensed by TLR2, and consequently by TLR2/1 or TLR2/6 heterodimers. These microorganisms include gram-positive and gram-negative bacteria, fungi, parasites, and mycobacteria [49]. Given the association of TLR1 1805G with Th1-mediated immune events, this SNP may influence the pathogenesis of any number of inflammatory conditions, including chronic mycobacterial infection, autoimmune disorders, sepsis and allergic reactions. |
10.1371/journal.pntd.0007768 | The association of depression, anxiety, and stress with caring for a child with Congenital Zika Syndrome in Brazil; Results of a cross-sectional study | Zika virus (ZIKV) infection in pregnancy can cause microcephaly and a wide spectrum of severe adverse outcomes, collectively called “Congenital Zika Syndrome” (CZS). Parenting a child with disabilities can have adverse mental health impacts, but these associations have not been fully explored in the context of CZS in Brazil.
A cross-sectional study was undertaken in Recife and Rio de Janeiro, including 163 caregivers of a child with CZS (cases) and 324 caregivers with an unaffected child (comparison subjects), identified from existing studies. The primary caregiver, almost always the mother, was interviewed using a structured questionnaire to collect information on: depression, anxiety, and stress (Depression, Anxiety, and Stress Scale—DASS-21), social support (Medical Outcomes Study Social Support Scale—MOS-SSS), and socio-demographic data. Data was collected May 2017-January 2018. Ethical standards were adhered to throughout the research.
A high proportion of mothers reported experiencing severe or extremely severe levels of depression (18%), anxiety (27%) and stress (36%). Mothers of children with CZS were more likely to experience symptoms of depression, anxiety andstress, compared to mothers of comparison children. These associations of were more apparent among mothers living in Rio de Janeiro. These differences were reduced after adjustment for socio-economic status and social support. Among mothers of children with CZS, low social support was linked to higher levels of depression, anxiety and stress, but there was no association with socio-economic status.
Depression, anxiety and stress were very common among mothers of young children in Brazil, regardless of whether they were parenting a child with disabilities. Mothers of children with CZS may be particularly vulnerable to poor mental health, and this association may be buffered through better social support.
| The 2015 Zika epidemic in Brazil gained international attention with the birth of thousands of babies with severe adverse outcomes, collectively called “Congenital Zika Syndrome” (CZS). Parenting a child with disabilities can be extremely stressful, and in other settings is linked to depression, anxiety and other mental health conditions. However, these associations have not been fully explored in the context of CZS in Brazil. We conducted a cross-sectional study in Rio de Janeiro and Recife, two cities in Brazil. We identified 163 caregivers (mostly the mother) of a child with CZS, and 324 caregivers of children of similar ages, but without obvious disabilities. The caregivers were asked a series of questions to gauge whether they were experiencing depression, anxiety or stress. We found that a high proportion of the mothers reported experiencing severe or extremely severe levels of depression (18%), anxiety (27%) and stress (36%). Levels of depression and anxiety were even higher for mothers of children with CZS compared to mothers of unaffected children, especially for those mothers with low social support and those living in Rio de Janeiro.
| Brazil hit international headlines in late 2015 with the birth of thousands of babies with microcephaly, which was soon linked to congenital infection with the Zika Virus (ZIKV). It was quickly apparent that babies born after congenital ZIKV infection experience a range of severe conditions beyond microcephaly alone, which is now collectively called “Congenital Zika Syndrome” (CZS). [1] Phenotypic characterization is still ongoing, but to date CZS is indicated by: microcephaly, other patterns of brain damage (including subcortical calcifications), damage to the back of the eye, congenital contractures, and hypertonia restricting body movement soon after birth. Children with CZS therefore have multiple and broad-ranging impairments, and as a result have complex care needs, which lie mostly with parents, in particular the mother. By early 2018, 3,149 cases of CZS were registered in Brazil, with many more potential cases not officially confirmed. [2]
Evidence is growing that carers of children with disabilities are more likely to experience depression, anxiety and stress, [3–5] and this may make it more difficult to meet the care needs of their child. Mental health impacts arise through different potential pathways. It may be distressing and overwhelming to meet the complex needs of a child with disabilities, which will continue for the lifetime of the child. Prevalence of depression and anxiety may therefore be particularly high among the carers of children with greater functional difficulties. [3, 4] Parents of children with disabilities may also experience negative attitudes and discrimination, and consequent social isolation, including marital breakdown, [6, 7] which can contribute to poor mental health. Caring for a severely disabled child often has cost implications, both direct costs (e.g. treatment) and indirect costs (e.g. foregone wages as the mother does not return to work), [8] creating financial problems, [7] and thereby increasing the risk of depression, anxiety and stress. [9] On the positive side, good social support and access to adequate resources may buffer the negative mental health impacts for carers.[10]
Carers of children with CZS may be particularly vulnerable to poor mental health. [11] These children experience severe impairments, with high healthcare needs. Affected children may be highly irritable and difficult to care for—as one mother described: “For every ten minutes of sleep, she cries for an hour”.[12] Another source of distress in relation to CZS is the uncertainty about the long-term trajectory of the condition, given its newness, and a lack of specialized professional knowledge. Families of children with CZS are also on average poorer and therefore may be less resilient to these challenges. [13] However, limited evidence is available on the impacts of CZS on the mental health of carers, despite the strong case for why a link is likely. One small study, conducted in Sergipe State in Brazil compared levels of anxiety and depression in the first 24 hours after birth between 9 mothers of newborns with microcephaly and twenty mothers with healthy newborns. [14] Mothers of babies with microcephaly had lower scores in the psychological domain of quality of life than those with healthy newborns, but there was no difference in anxiety (which was high in both groups) or depression. After ten months of follow-up, the mothers of children with CZS reported high levels of anxiety and low quality of life.[15] However, these results must be interpreted with caution due to the very small sample size, and other relevant studies are lacking.
In order to address these gaps in the literature, the current study aimed to explore the association of having a child affected by CZS with depression, anxiety, and stress, and to assess whether these relationships are buffered by social support and socio-economic status.
The methods of the study have been described in full elsewhere. [16] Briefly, a cross-sectional study was undertaken to explore the differences between caregivers with a child affected by CZS to those of typically developing children, in terms of social and economic indicators. Two contrasting sites were selected. The first was metropolitan areas in Recife, in the State of Pernambuco in Northeast Brazil, which had a high number of suspected and confirmed cases of CZS. [2] The second site was Rio de Janeiro City, where symptomatic ZIKV was less prevalent and reports of CZS far lower.
We aimed to recruit 100 cases of children with CZS and 100 comparison subjects without CZS per setting, which would provide the power to detect an OR of 2.6 in each site for the association between depression and CZS, assuming 95% confidence, 80% power and a prevalence of depression of 15% in unaffected caregivers. [17] Across the two samples (i.e. 200 cases and 200 comparison subjects), the sample size would be adequate to detect an OR of 2.05 for the same association.
In Recife, the source of most of the cases and all the comparison subjects was an existing case-control study, initiated in January 2016 to identify causes of CZS. [18] Cases were children born with microcephaly (head circumferences < 2 SD than the mean) in eight public maternity hospitals in Recife. Controls were children born in the same hospitals, but without microcephaly and without apparent neurological or other health problems (determined from transfontanellar ultrasonography, and through physical examination by the study neonatologist), with both examinations performed soon after birth. Controls were matched to cases on the basis of expected date of delivery and place of mother’s residence (by Health Region). At follow-up, caregivers of controls were asked whether there were any developmental delays (using the Denver test) and if the response was positive, the control child was excluded from the study and referred for further investigation. Additional cases with CZS were identified from an ongoing “cohort of children”. These children were identified as potentially having CZS from those born to a cohort of pregnant women who presented with a rash (a common symptom of ZIKV infection), and from outpatient clinics of children with CZS. Suspect cases were examined by a pool of specialists to confirm CZS. At follow-up, the participants were are classified as “cases” and “comparison subjects” (rather than cases and controls), to reflect the fact that we were assessing cross-sectional association of mental health conditions and CZS, rather than identifying exposures which may be aetiologically linked to CZS.
In Rio de Janeiro, the source of the cases and comparison subjects was the Vertical Exposure to Zika Virus and Its Consequences for Child Neurodevelopment: Cohort Study in Fiocruz/IFF (ClinicalTrials.gov Identifier: NCT03255369). Cases were children born to mothers known to be ZIKV positive, who had microcephaly. Comparison subjects were born to mothers without a history of symptoms and without developmental delay, as shown by: 1) a composite Bayley Score ≥85 conducted between 6 and 36 months following the recommended guidelines and/or 2) assessment by two paediatricians based on the child’s medical records. [19]
The primary caregiver, usually the mother, was interviewed using a structured questionnaire.
The Depression, Anxiety, and Stress Scale (DASS-21) was used to assess symptoms of depression, anxiety, and stress among caregivers. [20] It is a 21-item questionnaire with a four-point (0–3) answer scale that focuses on the extent participants had experienced certain symptoms over the previous week. Items are arranged into three subscales (depression, anxiety and stress), with seven items for each subscale. DASS-21 is a reliable tool to assess psychological distress and has been adapted and validated for Brazilian Portuguese. [21] The Medical Outcomes Study Social Support Scale (MOS-SSS) was used to measure social support. It is a 19-item questionnaire with each item scored on a Likert scale of 1 to 5, which includes five scales covering different aspects of social support (affection, positive social interaction, emotional, informational, and material). It has also been validated for Brazilian Portuguese. [22] Data was also collected on: the parents’ socio-demographic characteristics (e.g. asset ownership, parental marital status).
Data collection was undertaken between May 2017 and January, 2018. In Recife, the caregivers were interviewed in their homes, at the Primary Health Centre or occasionally, in their workplace. In Rio de Janeiro the interview was undertaken in person at attendance at The Fernandes Figueira Institute (IFF). In Recife, four quantitative interviewers were included (all female) and three were included in Rio de Janeiro (two female one male), all had a degree in social science or public health.
Variables were created from the standardized questionnaires. For DASS-21, sub-scales were calculated for Depression, Anxiety, and Stress Scale with each subclass’s score equal to the sum of seven corresponding questions. The sum scores were multiplied by 2 in order to match the original scale score in DASS-42 so that each subscale score ranges from 0 to 42. [20] Categories were created for:
Depression (normal: <9; mild: 10–13, moderate: 14–20, severe: 21–27, extremely severe: >28).
Anxiety (normal: <7; mild: 8–9, moderate: 10–14, severe: 15–19, extremely severe: >20).
Stress (normal: <14; mild: 15–18, moderate: 19–25, severe: 26–33, extremely severe: >= 34).
For MOS-SSS, an overall social support index was calculated, ranging from 0 to 100, with higher score indicating better availability of social support. In addition, four separate social support functional subscales were generated measuring: 1) emotional/informational social support, 2) tangible social support, 3) affectionate social support, and 4) positive social interaction.
Socio-demographic characteristics were regrouped into categories, to reflect the distribution of the data among the participants (groupings shown in Table 1). The Social-economic strata were estimated following the Brazilian Criteria, based on household assets from a pre-defined list, head of household education as well as the household’s access to piped water and a paved street. The household income estimates in each strata are approximations based on the 2013 National Household Sample Survey. [23]
A chi-square test was used to compare the socio-demographic characteristics of cases and comparison subjects, to identify potential confounders of the association of CZS with mental health outcomes. The mean depression, anxiety and stress levels were compared for carers of children with CZS and comparison subjects using a Wilcoxon rank sum test.
Multivariable logistic regression analyses were undertaken using Stata (version 15) to compare the odds of study outcomes (i.e. maternal depression, anxiety and stress) among carers of children with CZS to those of unaffected children, adjusted for age, socio-economic status (SES) variables and location (Rio de Janeiro/Recife). We tested associations for effect-modification by, in turn, location, social support and economic status. There were few missing data, and analyses were conducted for participants where complete data for relevant indicators were available.
Ethical approval for the full study was received from LSHTM and the Fiocruz ethics committee (CAAE 60682516.2.1001.5269). The original case-control study in Recife was approved by the Research Ethics Committees of the Pan American Health Organization (PAHO-2015-12-0075) and Fiocruz Pernambuco (CAAE: 51849215.9.0000.5190) and the Cohort study in Rio de Janeiro was approved by the IFF Ethics Committee (CAAE 52675616.0.0000.5192). The “Cohort of children” in Recife was approved by the ethics Committee of the Oswaldo Cruz University Hospital, University of Pernambuco (CAAE: 52803316.8.0000.5192). All interviewees were adults and all provided written informed consent.
The study included 81 cases and 112 comparison subjects from Recife, and 82 cases and 155 comparison subjects from Rio de Janeiro (Table 1). Among those interviewed, 97.5% were the mother, 1.4% the father and 1.0% the grandmother, and so the carer is referred to as the “mother” throughout for convenience. In both locations, mothers of cases were significantly younger than mothers of comparison subjects. Furthermore, in Recife the case children were older than the comparison children, whereas in Rio de Janeiro the reverse was true. In Rio de Janeiro, the case mother was less likely to be living in a union (71% versus 81% for comparison subjects), was less likely to be educated beyond the primary level (64% versus 74% for comparison subjects), and was in a lower economic stratum as compared to the comparison subjects. These differences were not apparent in the Recife sample, and in both locations the cases and comparison subjects were similar in terms of the number of people living in the household and gender of the child.
Median depression (p = 0.01), anxiety (p = 0.03) and stress scores (p = 0.05) were higher among cases than comparison subjects (Table 2). After stratification by location it was apparent that in Rio de Janeiro, scores were significantly higher for cases than comparison subjects for depression (p = 0.02), anxiety (p = 0.006) and stress (p = 0.02), while no differences were apparent in Recife between cases in comparison subjects.
Overall, a high proportion of mothers (across cases and comparison subjects) reported experiencing severe or extremely severe levels of depression (18%), anxiety (27%) and stress (36%) (Table 3). Unadjusted analyses showed that mothers of children with CZS showed higher levels of “any” depression (OR = 1.5, 95% CI = 1.0–2.2) or “any” anxiety (1.4, 1.0–2.1) compared to comparison subjects. However, these associations were not apparent after adjustment for socio-economic status and social support. There did not appear to be a trend in this association by severity of depression or anxiety. No differences in stress levels were identified between cases and comparison subjects.
The association between depression, anxiety and stress (in turn) and components of social support were investigated among the mothers of children with CZS (Table 4). Mothers of children with CZS were more likely to experience symptoms of depression if they had low levels of social support, and this was consistent across all domains of social support. For anxiety and stress, this pattern was apparent in relation to overall social support, positive social interaction, and emotional/informational support, but not for the other domains of social support.
The association between parenting a child with CZS and depression, anxiety and stress (in turn) was stratified by socio-economic status, social support and location (Table 5). No significant effect modification was detected for any of these comparisons Parents of children with CZS living in Rio de Janeiro appeared more likely to experience depression, anxiety and stress than parents of children with CZS living in Recife, but the p-value for interaction did not reach statistical significance.
This cross-sectional study compared the mental health of mothers of children with CZS to those with developmentally typical children in Brazil. The study found that a high proportion of mothers reported experiencing severe or extremely severe levels of depression, anxiety and stress, even among mothers of children without apparent disabilities. Mothers of children with CZS were more likely to experience symptoms of depression and anxiety, but not stress, and these differences disappeared after adjustment for socio-economic status and social support. Among mothers of children with CZS, low social support was linked to higher levels of depression, anxiety and stress. After stratification by location, anxiety and depression were associated with parenting a child with CZS in Rio de Janeiro, but not in Recife.
Previous studies have consistently demonstrated an association between parenting a child with neurodevelopmental disabilities and the experiences of depression, anxiety and stress.[3–5, 9] We can speculate on the reason for the disparity with the results of the current study, where these associations were less strong and consistent. ZIKV was reported about widely in the media in Brazil, and so there was a high awareness of the condition, and potentially less stigma experienced by parents than occurs for other types of childhood disability. The Brazilian government committed to the support of children with CZS and their families, and consequently many families may be able to access to health services and social benefits, reducing financial stress. There are also some social support structures in Brazil for parents of CZS, ranging from informal Whatsapp groups to established NGOs, which may help to buffer the negative mental health impacts of caring for a child with disabilities. Furthermore, poor mental health appeared to be extremely common among the comparison subjects, making it more difficult to detect a difference with cases, perhaps because of the pervasive social issues experienced in Brazil currently. The negative mental health impacts of having a child with CZS was more pronounced in parents living in Rio de Janeiro than in Recife. A potential explanation may be that the comparison mothers in Rio de Janeiro were generally older, more likely to be living in a union, and had higher levels of education and socio-economic status than the mothers in the other groups, which may influence the lower levels of depression and anxiety. In Recife, the cases and comparison subjectswere matched by location, and as women living in poverty and vulnerable situations are more likely to experience mental health concerns, differences between cases and comparison subjectswould have been reduced.
The mental health of carers of children with disabilities remains an important concern. However, in this context the high prevalence of depression, anxiety and stress among mothers of unaffected children was disturbing and needs urgent attention. Depression, anxiety and stress reduce the quality of life of those affected and are also linked to poorer physical health and social functioning. [24, 25] Parental depression can also result in negative parenting behaviours and reduced interaction with the child, so that the child may not receive the stimulation needed to thrive,[24, 25] which is particularly of concern in the case of children with disabilities, who are already facing challenges in their development. It is therefore important to identify the mental health needs of parents of children with disabilities, so that appropriate interventions can be implemented. A recent Cochrane review reported that psychological therapies improved parental mental health for parents of children with cancer post- treatment, [26] although evidence was lacking of positive impacts for parents of children with other chronic conditions. Improving social support may also help to buffer the negative impacts of caring on mental health, for instance, through establishing parent support groups, offering relationship counselling, or developing informal support networks (e.g. Whatsapp groups). [10, 11, 27] Improving mental health and social support of carers, may also help to improve adaptive parenting behavior, [26] and thereby benefit the child. Currently, more research is needed on which interventions work to improve mental health and social support of carers, particularly in low resource settings.
There are important strengths and limitations of the study, which must be taken into account when considering the findings. This was a relatively large study, conducted across two contrasting sites, and including comprehensive questionnaires using validated scales for depression, anxiety, stress and social support. This study also fills an important gap, as currently robust quantitative data on the mental health impacts of carers of children with CZS is lacking. In terms of limitations, clinical diagnosis of mental health conditions in the parents was lacking, and so we relied on screening questionnaires for assessing symptoms. The number of cases recruited was smaller than planned (though the number of comparison subjectswas higher), and the association between depression and CZS was weaker than anticipated, and so some of the analyses may have been under-powered. In particular, we were under-powered to identify effect modifiers.
This study found that depression, anxiety and stress were very common among mothers of young children in Brazil. Mothers of children with CZS may be particularly vulnerable to poor mental health, and this impact may be buffered by improving social support.
|
10.1371/journal.pntd.0000798 | Interactions between Global Health Initiatives and Country Health Systems: The Case of a Neglected Tropical Diseases Control Program in Mali | Recently, a number of Global Health Initiatives (GHI) have been created to address single disease issues in low-income countries, such as poliomyelitis, trachoma, neonatal tetanus, etc.. Empirical evidence on the effects of such GHIs on local health systems remains scarce. This paper explores positive and negative effects of the Integrated Neglected Tropical Disease (NTD) Control Initiative, consisting in mass preventive chemotherapy for five targeted NTDs, on Mali's health system where it was first implemented in 2007.
Campaign processes and interactions with the health system were assessed through participant observation in two rural districts (8 health centres each). Information was complemented by interviews with key informants, website search and literature review. Preliminary results were validated during feedback sessions with Malian authorities from national, regional and district levels. We present positive and negative effects of the NTD campaign on the health system using the WHO framework of analysis based on six interrelated elements: health service delivery, health workforce, health information system, drug procurement system, financing and governance. At point of delivery, campaign-related workload severely interfered with routine care delivery which was cut down or totally interrupted during the campaign, as nurses were absent from their health centre for campaign-related activities. Only 2 of the 16 health centres, characterized by a qualified, stable and motivated workforce, were able to keep routine services running and to use the campaign as an opportunity for quality improvement. Increased workload was compensated by allowances, which significantly improved staff income, but also contributed to divert attention away from core routine activities. While the campaign increased the availability of NTD drugs at country level, parallel systems for drug supply and evaluation requested extra efforts burdening local health systems. The campaign budget barely financed institutional strengthening. Finally, though the initiative rested at least partially on national structures, pressures to absorb donated drugs and reach short-term coverage results contributed to distract energies away from other priorities, including overall health systems strengthening.
Our study indicates that positive synergies between disease specific interventions and nontargeted health services are more likely to occur in robust health services and systems. Disease-specific interventions implemented as parallel activities in fragile health services may further weaken their responsiveness to community needs, especially when several GHIs operate simultaneously. Health system strengthening will not result from the sum of selective global interventions but requires a comprehensive approach.
| Prevention of neglected tropical diseases was recently significantly scaled up in sub-Saharan Africa, protecting entire populations with mass distribution of drugs: five different diseases are now addressed simultaneously with a package of four drugs. Some argue however that, similarly to other major control programs dealing with specific diseases, this NTD campaign fails to strengthen health systems and might even negatively affect regular care provision. In 2007, we conducted an exploratory field study in Mali, observing how the program was implemented in two rural areas and how it affected the health system. At the local level, we found that the campaign effects of care delivery differed across health services. In robust and well staffed health centres, the personnel successfully facilitated mass drug distribution while running routine consultations, and overall service functioning benefitted from programme resources. In more fragile health centres however, additional program workload severely disturbed access to regular care, and we observed operational problems affecting the quality of mass drug distribution. Strong health services appeared to be profitable to the NTD control program as well as to general care.
| Since 2000, global health initiatives (GHIs) have become a dominant international aid strategy, drawing on effective methods to control specific diseases, and accounting for a substantial increase of resources for global health [1]. Very soon, however, concerns emerged that, beyond anticipated benefits for targeted diseases, GHIs might erode health systems' capacity to respond to general health needs [2]–[7]. Early criticisms of GHI included the distortion of national policies as well as the creation of parallel bodies and processes burdening the health system [8]. Conversely, GHIs realized rapidly that their intervention capacity was limited by countries' weak health systems [6].
While it is now acknowledged that GHIs and health systems are influencing each other [8], [9], health systems and GHIs advocates still tend to have divergent views, partly framed in the long-standing horizontal-vertical discussion [10]. A WHO collaborative group was assigned in 2008 to assess the interactions between health systems and GHIs [9], and its findings were discussed during a policy dialogue meeting in Venice in June 2009 [11]. The report draws attention to the paucity of evidence to help understanding the interactions between GHIs and health systems. So far, most studies have dealt with global interventions in the field of HIV/AIDS control [8]. These research results may however not be applicable to other GHIs, as distinct objectives, policies, structures and operational processes of GHIs are likely to produce distinct effects on health systems [9]. Another limitation is that most studies focus on national level, while empirical evidence at point-of-delivery level remains particularly narrow [8], [9].
In recent years, growing awareness of NTDs, coupled with the availability of relatively low-cost control strategies, have led to important new global initiatives, including the WHO Neglected Tropical Disease Program, the Schistosomiasis Control Initiative (SCI), the Global Network for Neglected Tropical Diseases (GNNTD), the Neglected Tropical Disease Initiative (NTDI), and others [12]–[14]. The emphasis of the current global NTD control strategy is on mass drug administration. Based on geographic overlap and co-endemicity of NTDs, it addresses simultaneously up to five NTDs (lymphatic filariasis, onchocerciasis, schistosomiasis, soil-transmitted helminthiasis, and trachoma) with a package of 4 drugs (ivermectin or diethylcarbamazine, praziquantel, albendazole or mebendazole, and azithromycin). Mass chemotherapy conducted over successive years is expected to eliminate or reduce NTDs to a prevalence rate at which they no longer pose a threat to public health. The case for efficiency of the intervention is based on the “integration” of 5 diseases, but also on the fact that drugs are donated by pharmaceutical companies or available as generics, and distributed by community volunteers [15].
Mali, which already had experience with distinct campaigns for trachoma and schistosomiasis, was the first country to implement this integrated NTD control program in 2007 with USAID financial support. Since the knowledge on possible side effects of drug combinations was deemed insufficient, drugs were distributed serially for precautionary reasons, in a total period of 7 weeks between April and June 2007. Each single drug distribution was followed by a 2 weeks break (week 1: azitromicine, week 4: albendazole and ivermectine, week 7: praziquantel). For temporary financial constraints, the campaign was launched only in 3 regions, the plan for the rest of the country being postponed to end 2007. The campaign occurred simultaneously in the 3 regions and concerned 24 districts; in some of these districts, this coincided with a Vitamin A campaign.
The present study analyzes interactions between the NTD Control Program and Mali's health system, with a special focus on the district and service delivery level. Health districts in Mali are based on a network of health centers, each covering a defined health area; a typical health centre is staffed by a qualified nurse and 2 to 4 auxiliaries. A district executive team runs the referral hospital and provides technical support to the health centers.The aim of this exploratory study was to assess the program implementation processes on the field, and to identify plausible positive and negative effects for the health care system.
To gain more insight into the interactions between the NTD campaign and the local health system in Mali, we carried out an exploratory qualitative study, which is a common approach to study situations with little prior knowledge and few definite hypotheses [16].
We chose to observe NTD campaign interactions with health services in reasonably well functioning settings, in order to avoid that our findings could be attributed to major failures of the local health care system. Using ‘purposeful’ sampling [16], [17], we identified two rural districts from two different Regions that were typical of “good” rural districts in terms of output indicators (utilization rate for curative services and coverage rates for preventive activities).
Our approach combined three standard qualitative data collection methods, i.e. participant observation, in depth interviews with key informants and document analysis [16]. Interviews are valuable to gain information on feelings, thoughts and opinions, but less useful to describe events, behaviour or settings, as responses tend to be distorted by personal biases, lack of awareness, recall errors or selective accounts [16]. Participant observation is more appropriate to understand context issues, events and processes, but also has limitations including atypical behaviour of those being observed and selective perceptions of the observer [16]. Official documents provide a range of helpful information, including on planned processes and their rationale, but are selective and do not necessarily reflect actual processes. By using different methods we intended to compensate for limitations of each of them and to cross- check and triangulate our findings [16], [17].
Participant observation was conducted during two weeks in May – June 2007 by a researcher (AC) with a public health background, who was familiar with health systems in various sub-Saharan countries. In each of the two districts, she accompanied the district medical officer and/or health centre staff in their follow up of campaign activities at health centre and community level. This allowed for observations of mass drug administration in 16 health areas (8 areas per district); the selection of these areas was opportunistically determined by the district's agenda. Observation focused primarily on contextual issues, on procedures (e.g. task allocation, place of distribution, contents of information, dosage) and on behaviour of staff, drug distributors and community members. During observations, the researcher also used situational conversations, i.e. asking on-the-spot questions and discussing with district authorities and health centre staff in a naturalistic and informal way [18], [19]. Besides routine drug administration, participant observation including situational conversations also entailed a “training of trainers” session for district health teams at Regional level, two district executive team meetings, and a community meeting.
Structured interviews were not conducted at local level due to heavy time constraints for staff during the campaign. But in depth interviews were conducted with key informants, including ten Ministry of Health officers and ten representatives of support agencies acquainted with the Malian health system. Key informants were identified through snowball sampling [16]–[17]. These interviews sought to elicit information on campaign processes which could not be directly observed - such as decision making at national level, planning and financing - and to explore informants' views on interactions between the NTD control program and local health services.
Further information was collected by consulting Malian official documents, website search and literature review. We sought to triangulate the data as much as possible and to check the same information at different sources.
Observation and interview data were recorded in field note transcripts that were reviewed independently by two researchers. The study being exploratory, categories for analysis were largely inductive, though some were based on interactions of other GHIs with health systems reported in the literature [2]–[7]. They included implementation processes (training, drug procurement and distribution, monitoring), task distribution, positive and negative effects on service delivery, and decision making processes. Throughout these issues we looked for emerging recurrent patterns and variations among situations and/or informants.
Results were validated and complementary information was generated during three feedback sessions held in November 2007 with health authorities at peripheral, regional and national level.
Most data are based on field observation which had no potential harmful effects on patients or other vulnerable persons. Data collection was complemented by in depth interviews with twenty key informants. We asked about their opinions which were mostly public. Nevertheless we asked for oral consent after explaining the purpose and methods of our study, and ensured confidentiality, as some of the key informants who were high officials of the MOH or international organizations preferred not to be identified. This consent was witnessed by at least one person other than the principal investigator. No written consent of key informants was asked, as this would have been unusual in the context and might have biased the interview process, otherwise quite informal. Another ethical concern for this health systems research was governance: the national authorities of the Ministry of Health in Mali gave permission, regional and district authorities were supportive of this research, and provisional results were shared and discussed with authorities before broader dissemination. Potential ethical issues were examined with the head of the IRB of ITM (Antwerp) and it was decided that there was no need for a full review. Neither was it required under the Malian legislation (see law n°2009/63/4L), which rules biomedical research but not health systems research, and was not applicable at the time of the study.
Our results are presented according to the conceptual framework as presented by the WHO Maximizing Positive Synergies Collaborative Group [9]. We examine both positive and negative effects of the NTD control program on health service delivery, health workforce, health information system, supply management system, financing, and governance. These effects are summarized in Table 1.
Access to mass chemotherapy for targeted NTDs clearly improved according to all interviewees. Some informants however regretted that the control program only included drug distribution and some health education on NTDs, and did not address other NTD disease control strategies, such as curative care (e.g. eye surgery for trachoma) or sanitation.
Several informants also criticized the high priority given to targeted diseases, while more common health problems received little attention; they worried about the campaign mobilising energy and diverting staff's attention from routine care delivery. These interview results were in line with our direct field observations: routine care was cut down or even totally interrupted in most health centres observed, as a consequence of health centre nurses being absent from their station and not being replaced by other staff. During the first round of the NTD control program, head nurses were required to devote 10 full working days for program-related training and supervision, in addition to monitoring and drug supply activities (Table 2). The mass distribution schedule also required health centre staff to postpone or reorganise planned routine outreach immunisation sessions.
Some informants thought that, in a context of low service utilisation, the campaign was at least a way to bring services closer to underserved populations. Observation however did not suggest that the campaign had positive effects on nontargeted services. We observed missed opportunities for curative care: children queuing for NTD prophylactic drugs and presenting obvious other illnesses and need for care (e.g. abscesses or trauma) were not identified as such by the attending staff.
Not all health centers responded similarly to these interferences: 2 of the 16 health centers managed to keep their curative consultations and immunization services running normally and to use the campaign to support the overall development of their health centre. For instance, one nurse passed his NTD training on to other team members, another took advantage of NTD training of community volunteers to discuss other health issues than the targeted NTDs, and supervision of the campaign at village level became an opportunity for health education on other topics. These 2 health centres differed from the others in terms of human resources: both were well staffed and had no vacant positions. They were managed by a qualified nurse, who was on the job for over 5 years, and reputed for dynamism, professionalism and leadership, both at regional and central level. Moreover, utilisation rates in these centres were above the national average (>0.20 new cases/inhabitant/year), preventive coverage rates were considered good (>75%), and they benefited from a supportive community organisation.
Disparities between health centres were also seen with respect to the health centre capacity to implement the NTD control program: operational problems were observed in all health centres, except for the two more robust ones. These problems included errors in the population census, poor community mobilisation, drug dosing errors and omission of side effect monitoring. While these problems were not captured by the monitoring system of the NTD campaign, which indicators were limited to treatment and geographical coverage, they nevertheless suggest quality problems in campaign implementation processes.
Finally, several interviewees highlighted possible negative effects of free distribution of drugs on care seeking: as sick patients have to pay for drugs in routine circumstances: patients might, so they feared, wait for a next edition of the campaign rather than seek care.
While mass drug distribution itself relied on community volunteers, the NTD campaign nevertheless implied increased workload for both district and health centre staff to ensure drug supply, campaign monitoring, training and supervision.
Informants thought that the training component was a positive effect of the program. A training cascade was organized, starting with a “training of trainers” where national program coordinators trained district authorities. The cascade continued with district authorities training health centre nurses, who trained community volunteers. The training consisted in transmission of information on epidemiology, diagnosis and treatment of each of the targeted diseases; several participants said it was a repetition of previous training sessions.
Training and supervision activities entitled staff to receive allowances, which represented approximately an 80% increase of a district medical officer monthly salary (increase of 122 000 F CFA for an average salary of 150 000 F CFA), and a 45% increase of a health centre nurse salary (increase of 46 000 F CFA for an average salary of 100 000 F CFA). Some informants considered these incentives as contributing to the motivation and retention of health staff, but others reported that these allowances distracted staff's attention from their core activities, for which no extra allowances were provided.
Allowances were also given to community volunteers. Most informants considered this as necessary to attract volunteers. We witnessed a local dispute around the selection of volunteers, suggesting that becoming a volunteer was much sought after. Still, informants worried about the sustainability of allowances for volunteers (which were supposed to be paid by the communities themselves after the first round of the NTD control program), as well as about the inconsistency of allowance amounts between various donors, generating growing demands from the side of volunteers.
Preparatory to mass drug distribution, census data were collected which can be used for purposes other than NTD control. Improved information on NTD treatment coverage was also made available.
The NTD campaign however introduced a parallel monitoring and evaluation system. At district level, 12 new forms were introduced for drug supply management, 15 new forms for monitoring and supervision of campaign activities, and 1 new form for clinical complications of filariasis (Table 3). For each drug distributed, one report per village, one per health centre and one per district was required. Following donor instructions, a specific timetable for reporting was installed: village data were processed daily at health centre level, and transmitted to district level. Districts reported campaign results weekly to the regional level.
NTD campaign drugs were mostly donated by pharmaceutical companies, which increased their availability for preventive chemotherapy nationwide. But a parallel drug procurement system was established to ensure rapid distribution of drugs from national to regional and district level. As storage space at national and regional level was insufficient, trucks were especially rented for the occasion. Drug management forms and processes, distinct from the national ones, were created specifically for campaign implementation (see Table 3). Besides, some informants stressed imbalances in drug availability: while Azitromicine was distributed at no cost during the campaign, patients suffering from trachoma had no access to this drug in routine conditions, as it was substituted by payable tetracycline ointment. One local informant reported complaints from the community on this issue.
According to official documents [20], the provisional budget for mass drug administration – cost of drugs not included - was approximately US $ 12 million spread over 5 years, most of which was financed by USAID, with complements from other agencies. This budget covered drug distribution (29% of the budget), training of staff and volunteers (26%), supervision and evaluation (17%), drug procurement (9%), health education (7%), institutional strengthening (10%) and intersectoral collaboration (2%). Not included in the budget were State supported staff salaries and infrastructure used for the campaign. Institutional strengthening consisted mostly in intervention related equipment and staff, and several informants considered that the budget left no room for institutional strengthening beyond campaign related needs. Investments in general equipment were limited to motorbikes for districts; the increase of storage capacity at national level was not approved, nor was the acquisition (rather than rental) of trucks for drug procurement, which several informants regretted. Though acknowledging the relevance of the intervention, they expressed concerns about longer-term financing for sustainable campaign results.
Though the decision to implement the program was taken by Malian authorities, several informants considered that national authorities had little space for negotiation, as most decisions were taken at supranational level by donors and their grantees. Indeed USAID financing was not directly allocated to the country, but to sub-grantees which, in the case of Mali were the International Trachoma Initiative (ITI), replaced by Helen Keller International at the end of 2007.
A steering group with coordination and decision-making power was set up in parallel to the existing coordination structures in the Malian Ministry of Health. It included MoH officials and ITI staff. The strategic NTD control plan for 2007–2011 was designed at National Health Directorate level. Several informants considered the design of a single NTD control plan as a positive effect of the program, as it stimulated coordination between previously standalone program coordinators. However, the strategic plan had to adapt to donor and grantees requirements. Predefined strategies, earmarked funding and budgets kept tight for the sake of demonstrating efficiency left only limited margins of maneuver.
Several informants were also critical about distorting effects of the program on national priorities. They recognized the importance of NTDs and Mali's longstanding experience with mass treatment targeting distinct NTDs: onchocerchiasis since 1988, trachoma since 2004, schistosomiasis since 2005, and attempts to integrate mass treatment for onchocerchiasis, filariasis and helminthiasis started in 2005. Donor websites emphasize that the program met ongoing country efforts and contributed to scale them up at national level. But informants also reported that these ongoing efforts partly resulted from previous external financing opportunities as well. They considered that the accumulation of donor conditions was hampering resource allocation following national strategic orientations.
Finally, the top-down implementation process of the NTD campaign was felt by most informants to contradict local district leadership, central to Malian health policy (PRODESS II), and to interfere with planned activities. Indeed, as information concerning the campaign reached regional and district authorities with short notice, district authorities had to modify or adjourn their calendar of health centre supervisions with no space for negotiation.
This study is the first to address interactions between the integrated NTD control program and a country health system. It provides insights on positive and negative effects of the integrated NTD control program at point-of-delivery and on district systems. A previous study documented community resistance to free drug distribution for schistosomiasis and soil transmitted helminths in Uganda, but did not address health system effects [21].
A few limitations of this study should not remain unnoticed. Data collection was inevitably influenced by the presence of the researcher and relations between her and people in the field [17]. It is however plausible that investigator effects led authorities and staff to show the best of their performance and refrain from critical comments: biases are more likely to minimise rather than maximise problems. Furthermore this qualitative study was based on a limited number of interviews and contextualised observation units, and not intended to be generalisable as in quantitative research, though readers may assess their applicability to their own settings [17] Purposeful sampling [16]–[17] does not provide guarantees that the districts and health centres observed are typical of the country. As we selected “better performing” districts in terms of output indicators, we assume that campaign related problems are not more severe in these districts than elsewhere in the country, but this needs to be probed. Another reason for caution in transferring our findings to other settings or countries is that the Malian campaign was the very first NTD control program integrating five diseases and may have suffered from startup problems, avoided in further editions. As most USAID-supported NTD campaigns are based on similar principles and processes as in Mali, our study however provides plausible hypotheses to be tested in other contexts. The purpose of our study was exploratory, which implies that more research is needed, both qualitative and quantitative. We suggest that our findings are helpful in framing further research questions.
A large part of current understanding of interactions between GHI and health systems is based on HIV/AIDS related programs. Though the NTD control program appears as a “small” GHI compared to others such as Global Fund or PEPFAR [1], its analysis brings new insights to the ongoing debate.
The NTD control program, like other preventive programs, focuses on protection rather responsiveness to patient demand [22]. This distinguishes it from HIV/AIDS related GHIs scaling up access to ARV treatment for individual patients. We found that the integrated NTD program did not include curative care for NTDs, but also that extra workload and staff absences for program purposes disrupted access to general curative care at health centre level. A Bill and Melinda Gates Foundation study [23] reported extra workload for district staff in Angola and Tanzania resulting from donor requirements, but did not assess effects on service delivery. Evidence on effects of workload generated by GHI on access to health centre care is so far extremely limited [24].
A key finding of our study is the emerging differentiation between health centres within the same district in their capacity to cope with program's interferences. Only the most resilient services, characterized by a qualified, stable and motivated workforce, managed to maintain routine activities, and even to use the program as an opportunity for overall quality improvement. This finding is consistent with the notion that positive effects of GHIs are more likely to occur when the health system is robust [9], [25]. But it expands it from country health system to district and health service level, emphasizing the importance of human resources for robustness of services. Future research should further explore the relation between human resources characteristics and absorption capacity of programs at health service level. Indeed, this might bear consequences for the adaptation of programs to specific local health systems and services.
Other findings of our research are coherent with known effects of other GHIs on health systems and show a mixture of positive and negative effects. Like other programs, uptake in targeted services was scaled up, but duplications occurred, especially for drug procurement and monitoring and evaluation [8]–[9]; these parallel systems, meant to improve campaign efficiency, increased workload and total costs for the health system. Also like other GHIs, the NTD control program influenced priority setting [7]: pressure to absorb donated drugs and reach short term chemotherapy coverage results contributed to the distraction of energies away from other NTD control strategies such as treatment and sanitation, and more generally from overall health systems strengthening.
From a health systems perspective, the question is however not so much whether the balance of a specific GHI is positive or negative. The problem for Mali, as for other countries, is the cumulative effect of a large number of GHIs and other campaigns, each with implications at all levels. Besides NTD campaigns, Malian health services also run National Immunisation Days, Vitamin A and bed-net distribution as well as eradication or elimination campaigns of polio, tetanus or yellow fever. An estimation of time spent outside the health centre by head nurses of a Malian rural district in 2006 showed absences reaching 54% of working days; half of this time was dedicated to campaigns and trainings linked to vertical programs [24]. As the head nurse is usually the only staff qualified to provide first line curative care in Mali, disruptions in consultation schedules erode service responsiveness and community's confidence in their local health centre [26]. Another cumulative effect is the growing mobilisation of communities to meet top down defined targets, to the expense of an empowerment approach to community participation [21], [27].
The need for health system strengthening is increasingly acknowledged, also by promoters of integrated NTD control [28]. Most GHIs claim to contribute to health system strengthening with additional resources and capacity strengthening, but these interventions are mostly selective, targeting those system functions essential for implementation of their own program [22]. This was also the case for the Malian NTD control program. The prospect of adding vitamin A, bed nets and vaccines to the present campaign model [28] might contribute to the improvement of the protective function of health systems, but not to their responsiveness to population's demand for curative care, which may be even further undermined [22].
The control of NTDs in vulnerable communities is a necessity. But so is health systems strengthening, in order to respond adequately to other health problems and to ensure sustainable achievements, including of NTD control. A major challenge is how to engage in disease control – NTD and other diseases - without negatively impacting on existing health systems. Increased knowledge on interactions with the health system is needed to allow GHIs to plan for positive effects and alleviate potential negative effects. Presently, short term and quick win interventions are given priority, but more long term strategies are also needed. Health system strengthening should rely on country-specific development plans aligned with national policy, and requires a comprehensive approach across diseases and health problems and coordination among GHIs. For example, program specific in-service training should be organised in ways mitigating potential interruptions of service provision, but investments are also needed in pre-service education for qualified staff. The accumulation of program specific extra allowances, making targeted interventions more popular than routine activities [29], could gradually be replaced by comprehensive human resource management at national and district level. Parallel drug supply should be limited to exceptional emergencies, and investments redirected to reinforce national drug supply systems.
There are signs of GHIs learning from experience and gradually modifying some of their processes [8]. They also show increasing willingness to reduce fragmentation and to review processes [11]. Still the chaotic architecture for development assistance for health remains a major obstacle for health system strengthening. Progress towards effective and inclusive health systems will not result from the sum of selective GHI interventions.
|
10.1371/journal.pgen.1000360 | Reduced Neutrophil Count in People of African Descent Is Due To a Regulatory Variant in the Duffy Antigen Receptor for Chemokines Gene | Persistently low white blood cell count (WBC) and neutrophil count is a well-described phenomenon in persons of African ancestry, whose etiology remains unknown. We recently used admixture mapping to identify an approximately 1-megabase region on chromosome 1, where ancestry status (African or European) almost entirely accounted for the difference in WBC between African Americans and European Americans. To identify the specific genetic change responsible for this association, we analyzed genotype and phenotype data from 6,005 African Americans from the Jackson Heart Study (JHS), the Health, Aging and Body Composition (Health ABC) Study, and the Atherosclerosis Risk in Communities (ARIC) Study. We demonstrate that the causal variant must be at least 91% different in frequency between West Africans and European Americans. An excellent candidate is the Duffy Null polymorphism (SNP rs2814778 at chromosome 1q23.2), which is the only polymorphism in the region known to be so differentiated in frequency and is already known to protect against Plasmodium vivax malaria. We confirm that rs2814778 is predictive of WBC and neutrophil count in African Americans above beyond the previously described admixture association (P = 3.8×10−5), establishing a novel phenotype for this genetic variant.
| Many African Americans have white blood cell counts (WBC) that are persistently below the normal range for people of European descent, a condition called “benign ethnic neutropenia.” Because most African Americans have both African and European ancestors, selected genetic variants can be analyzed to assign probable African or European origin to each region of each such person's chromosomes. Previously, we found a region on chromosome 1 where increased local African ancestry completely accounted for differences in WBC between African and European Americans, suggesting the presence of an African-derived variant causing low WBC. Here, we show that low neutrophil count is predominantly responsible for low WBC; that a dominant, European-derived allele contributes to high neutrophil count; and that the frequency of this allele differs in Africans and Europeans by >91%. Across the chromosome 1 locus, only the well-characterized “Duffy” polymorphism was this differentiated. Neutrophil count was more strongly associated to the Duffy variant than to ancestry, suggesting that the variant itself causes benign ethnic neutropenia. The African, or “null,” form of this variant abolishes expression of the “Duffy Antigen Receptor for Chemokines” on red blood cells, perhaps altering the concentrations and distribution of chemokines that regulate neutrophil production or migration.
| A large proportion of healthy African Americans have been observed to have a white blood cell count (WBC) that is persistently lower than the normal range defined for individuals of European ancestry [1]–[5]. This condition, called “benign ethnic neutropenia”, can have important effects on medical decision-making, since WBC is a valuable indicator of immunocompetence, infection, and inflammation. To seek the genetic basis of benign ethnic neutropenia, we recently carried out an admixture mapping analysis in which we identified a locus on chromosome 1 where local inheritance of African or European ancestry is sufficient to account entirely for the epidemiological differences in WBC levels between African Americans and European Americans [6]. By genotyping samples from two epidemiological cohorts—the Health Aging and Body Composition Study (Health ABC) and the Jackson Heart Study (JHS)—at a panel of markers that were extremely differentiated in frequency between Africans and Europeans, we identified an approximately 900 kilobase locus on chromosome 1 (99% credible interval of 155.46–156.36 Mb) where individuals with low WBC had increased African ancestry compared with the average in the genome.
In the present study, we narrowed the region of association from 900 kb to a single base pair substitution that is likely to have a strong effect on variation in WBC. To achieve this, we increased our sample size from 1,550 in the initial study to 6,005, by pooling samples from the Jackson Heart Study (JHS), the Health ABC Study, and the Atherosclerosis Risk in Communities (ARIC) Study. We found that neutrophil count is responsible for the vast majority of the WBC association at the locus, and therefore focused on neutrophil count in the current analysis. We also showed that the genetic change that is probably responsible is the Duffy Null polymorphism (rs2814778, also called FY+/−), which is already known to protect individuals of African descent against Plasmodium vivax malaria infection [7],[8], and which has recently been associated with susceptibility to HIV infection and rate of progression to AIDS [9]. Our identification of this polymorphism as the probable cause of benign ethnic neutropenia should prompt further investigation of its effects on hematopoiesis and immunity.
We pooled 6,005 African American samples from three cohort studies: the Jackson Heart Study (JHS), the Atherosclerotic Risk in Communities (ARIC) Study, and the Health, Aging and Body Composition (Health ABC) Study. For each sample, we required a high quality genome-wide admixture scan (Materials and Methods), a genotype at SNP rs2814778, body mass index (BMI), age, gender, and a full differential white blood cell count (with the exception that for Health ABC samples we did not require a measurement of bands).
To explore correlations between the genetics and the phenotype, we first used the genotype at SNP rs2814778, which occurs at position 155,987,755 in Build 35 of the human genome reference sequence, within the 99% credible interval defined by our previous admixture mapping study [6]. This SNP is also known as the “FY+/−” or “Duffy” variant, and the FY− allele is very highly correlated to West African ancestry. For example, it is completely fixed in frequency in West African and European American samples from the International Haplotype Map [10] (although it is not completely fixed in larger sample sizes from these populations; see below). For Figure 1 and Tables 1 and 2, we used the genotype at rs2814778 as a surrogate for ancestry because the genotype can be conveniently read out as a discrete value (0, 1 or 2 copies) rather than as a continuous value, and is extraordinarily correlated to ancestry (r2>0.99). Later, we demonstrate that there is in fact a slightly stronger association to neutrophil count for rs2814778 than for ancestry, which is important in showing that the FY− allele at this polymorphism may actually be responsible for low neutrophil counts, and is not just in admixture linkage disequilibrium with the causal allele.
To test for heterogeneity in the strength of the genetic association to WBC among the different sample sets that comprised our study, we divided the samples into four groups. There were 658 samples from the Health Aging and Body Composition Study (“Health ABC”), 1,969 samples from the JHS cohort only (after randomly dropping samples until there was only one from each pedigree; “JHS only”), 2,476 samples from the ARIC cohort only (“ARIC only”), and 902 samples that overlapped between JHS and ARIC (“JHS-ARIC overlap”). For the JHS-ARIC overlap samples, we averaged all phenotype measurements, taken an average of 14 years apart at the time the participant entered each study (and having a correlation coefficient of r2 = 0.37), to provide a more precise estimate of the phenotype than would be available from either measurement alone.
Table 1 presents the characteristics of each of the groups of samples. We found that all sets of samples showed quantitatively similar associations to the chromosome 1 locus. In particular, for neutrophil count, individuals carrying at least one European-type (“FY+”) allele of rs2814778 had 1.58–1.65 times higher values, depending on the study, than individuals homozygous for African ancestry (FY−/−), a tight enough range that we decided to pool all four groups of samples for subsequent analyses. Despite the similar correlation of neutrophil count to local ancestry across studies, we observed that the correlation coefficient to “European carrier status” was significantly higher for JHS-ARIC overlap samples, ρ = 0.57, than for the samples for which only one measurement was made: ρ = 0.52 for JHS-only (P = 0.03 for a reduction) and ρ = 0.50 for ARIC-only (P = 0.002 for a reduction) (Table 1). This is likely to reflect a more accurate assessment of basal neutrophil count when it was measured twice and averaged over different environmental conditions (the baseline measurements in JHS and ARIC studies taken an average of 14 years apart) than when it was measured only once. In support of this hypothesis, the JHS-ARIC overlap samples contributed more per sample to the statistical signal than those measured in only one cohort: 28% more per sample on average, which we calculated by dividing the LOD score they contributed by the total number of samples.
Combining all samples (n = 6,005 in total) and working with normally transformed cell counts for each white blood cell lineage (Materials and Methods) we explored how counts of total WBC and each of the 6 differential counts were associated with ancestry at the locus, using the genotype at rs2814778 as a surrogate for ancestry (Table 2).
There was strong evidence that the allele at the locus that contributed to high white blood cell count had an almost purely dominant effect. As shown in Table 2, there was no significant difference in leukocyte counts between 247 African Americans with two copies of European ancestry at this locus (FY+/+) and 1,647 African Americans who were FY+/− (P>0.08 for WBC and all differential counts). By contrast, being FY+/+ or +/− (1,894 African Americans) vs. FY−/− (4,111 African Americans) was strongly associated to counts of all white blood cell types except bands (P<<10−4; Table 2).
The dominant effect of European ancestry on white blood cell count is also visually apparent in Figure 1, which shows the distribution of neutrophil count for individuals grouped according to genotype at rs2814778. Persons carrying at least one FY+ allele had a distribution of neutrophil counts that was shifted by 1.3 standard deviations above that of persons who were FY−/− (this was extraordinarily statistically significant: Z = 49.7). By contrast, there was no significant difference between individuals who carried either one or two FY+ alleles (Z = 0.6). For further analysis, we pooled individuals who were carriers of the FY+ allele at this locus.
The differential white blood cell count that was most significantly associated with ancestry was absolute neutrophil count (calculated as total WBC multiplied by the percentage of neutrophils). The correlation (ρ) of normally transformed absolute neutrophil count to carrier status for the FY+ allele was 0.519, which was higher than that of the general WBC phenotype originally mapped to the locus [6] (ρ = 0.458). In the 952 African Americans who had absolute neutrophil counts at least 1 s.d. below the mean (roughly <1,800 /mm3), the proportion of FY+ allele carriers was reduced by more than an order of magnitude compared with the genome wide average.
Neutrophil count was responsible for the vast majority of WBC association at the locus. After controlling for neutrophil count in a regression analysis, only monocyte count (ρ = 0.025, P = 0.05) and basophil count (ρ = −0.034, P = 0.009) remained nominally associated, and these associations were not significant after correcting for the 6 hypotheses tested (Table 2). The weak evidence of association to monocyte and basophil counts may reflect a real effect, or may be a false-positive due to multiple hypothesis testing. It is also possible that the result may be an experimental artifact related to the Coulter Counter technology used to measure differential WBC. In these measurements, the positions of monocytes and basophils were near those of neutrophils in the plots used for cell classification. Even a small amount of misclassification among neutrophils, monocytes, and basophils (a couple of percent) could cause their counts to be artifactually correlated, contributing to the signals we observe in the context of measurements in large sample sizes. Since neutrophil count appears to drive at least the great majority of association, we focused on this WBC phenotype for all further analysis
To assess whether the higher neutrophil count observed in European Americans compared with African Americans can be entirely accounted for by ancestry at the chromosome 1 locus, as is the case with total WBC [6], we examined samples from the Health ABC study (1,331 European Americans and 658 African Americans). Among African Americans who could be classified with confidence as carrying at least one chromosome of European ancestry at the locus, the absolute neutrophil count did not differ from that of European Americans (P = 0.99). Thus, genetic variation at the chromosome 1 locus was sufficient to account for the entire epidemiological difference across these populations.
The predictive effect of ancestry at the chromosome 1 locus was profound. Carrier status for the European-type (FY+) allele at the rs2814778 variant predicted 26.95% of the variance in normally transformed neutrophil count, which was far more than the 3.37% predicted by genome-wide European ancestry proportion. After controlling for rs2814778 genotype, there was no longer any association to genome-wide European ancestry. Similarly, after controlling for rs2814778 genotype, BMI and gender only predicted 0.79% and 0.14% of variance in neutrophil count respectively, while smoking (analyzed in JHS only) only predicted 0.8% of the variance. Age was not significantly associated to neutrophil count in our data (P = 0.25). We did not analyze other phenotypes like hypertension and coronary artery disease status for their correlation to neutrophil count. Because of the relatively weak contributions of all the non-genetic predictors we analyzed, we focused subsequent analyses on genotype at the chromosome 1 locus uncorrected for covariates.
We were able to place strong constraints on the frequency of the variant affecting neutrophil count by analyzing the distributions of neutrophil count for individuals with 0, 1 and 2 copies of European ancestry at the chromosome 1 locus, which in practice we marked by the genotype at rs2814778. The analysis in Figure 1A provides strong evidence of a dominant allele of European origin contributing to high neutrophil count. We modeled the frequency of the variant that causes high neutrophil count by defining 6 parameters. The frequency of this variant in Africans was specified as PA and its frequency in Europeans as PE. Individuals who were homozygous for the other allele were assumed to have a normal distribution of neutrophil count with mean μL and standard deviation σL, and carriers of the “high neutrophil” allele were assumed to have a normal distribution of neutrophil count with mean μH and standard deviation σH. Studying a grid of values of PA (Figure 1B), and another grid of values of PE (Figure 1C), we found the combination of the remaining variables that provided the best fit to the data, as assessed by a chi-square goodness-of-fit statistic. Given each set of 6 model parameters, we calculated a likelihood of the data for all 6,005 individuals. This resulted in a marginal likelihood surface for PA (Figure 1B) and PE (Figure 1C), which we used to place constraints on these parameters.
Fitting this 6-parameter model to the data, we inferred that the frequency of the allele contributing to high neutrophil counts was <4.9% in Africans and >95.2% in Europeans (Figure 1B, C), and that the difference in frequency between populations was >91.9%. Compared to 3.54 million autosomal SNPs in the November 2006 Phase2 HapMap data set [10], there were only 115 SNPs with a frequency differentiation at least this extreme, and only one in the region of admixture association: the SNP rs2814778 (at position 155.99 Mb), the same SNP we used as a marker of ancestry. This variant already has a known phenotype—susceptibility to Plasmodium vivax malaria—but it had not been hypothesized to be associated with low white blood cell count until it was found to lie within this locus [6]. While rs2814778 is a plausible candidate, the locus we described previously [6] spans 900 kb, and there could in principle be other variants within this span—unreported in the literature or in genome variation databases—that have a high enough frequency differentiation to explain the signal. In what follows, we present additional lines of evidence to rule out the great majority of sites other than rs2814778 as consistent with explaining the signal.
We used four strategies to increase the height of our admixture association peak and thereby to narrow the position of the allele affecting neutrophil count. First, we used the fact that with 6,005 samples in our admixture mapping analysis pooled across three studies (Table 3), we had a greater sample size than the 1,550 samples that were used initially [6]. Second, we designed an analysis (Materials and Methods) that used all of the samples instead of just the extremes of the distribution [6]. Third, we changed the phenotype from total WBC (ρ = 0.458 correlated to ancestry at the locus) to absolute neutrophil count (ρ = 0.519) to obtain a sharper statistical signal. Finally, we genotyped the JHS samples (including the JHS-ARIC overlap samples) at additional ancestry informative markers, spaced at a density of about 1 every 400 kb across the peak, to increase spatial resolution.
Using all 6,005 samples and the stronger phenotype of absolute neutrophil count, the LOD score (log base 10 of the Bayes score) rose to 363.1 (Table 3; Figure 2). The 99% credible interval was narrowed to ∼450 kb (155.957–156.407 kb), and still contained the Duffy null polymorphism at position 155,987,756 near the DARC gene (Entrez GeneID: 2532), as well as a handful of other genes listed in the lower panel of Figure 2.
We exploited the large sample size (6,005 individuals) to test whether the rs2814778 variant predicted low neutrophil count more than would be expected from the association to ancestry [6]. This is a difficult problem since the genotype at this SNP is highly correlated to ancestry. By using the ANCESTRYMAP software and the data from all 6,005 African Americans, we estimated that the frequency of FY+ allele at rs2814778 is 0.2±0.1% in Africans and 99.3±0.4% in Europeans (this frequency distribution is consistent with the allele frequencies inferred for the causal allele based on modeling of neutrophil counts in Figures 1B and 1C). Thus, if rs2814778 is the causal variant, there should be a small handful of individuals for whom the genotype at rs2814778 is discrepant with ancestry, who will be informative for our analyses.
To estimate the number of individuals who we expect to be informative for testing association of rs2814778 above and beyond ancestry, we used the fact that the cohort has 18.2% European ancestry on average (Table 2). Thus, we expected there to be about 13 individuals who are homozygous for the Duffy null allele at rs2814778 but heterozygous for European ancestry: 13 = (6005)×(2×18.2%×81.8%)×(0.7%). Similarly, we expected there to be about 8 individuals who are heterozygous at rs2814778 but homozygous for local African ancestry: 8 = (6005)×(81.8%×81.8%)×(0.2%).
To test for association to rs2814778 above and beyond ancestry, we first obtained estimates of European ancestry at the position of the SNP using the ANCESTRYMAP software [11]. We included rs2814778 in the ancestry estimation so that we could explicitly test whether the genotype at this SNP alone was more predictive of neutrophil count than this SNP plus flanking markers. This would be evidence that it was more associated than African ancestry itself. Our power to detect a signal was highest for JHS samples, which were genotyped at a high density at the chromosome 1 locus. Consistent with this observation, the 7 samples for which we could state with >50% confidence that the local ancestry was discrepant with the expectation from the rs2814778 genotype were all from JHS.
We performed three regression analyses (Table 4) to explore whether rs2814778 or ancestry status at the chromosome 1 locus was a better predictor of neutrophil count. (a) First, we obtained a χ2 statistic for association of carrier status for the rs2814778 FY+ allele to neutrophil count; (b) second, we obtained a χ2 statistic for association of carrier status for European ancestry to neutrophil count (using the rs2814778 genotype in the estimate); and (c) third, we obtained a χ2 statistic for association of both predictors together. We found that there was a significant difference between the strength of association of ancestry alone and ancestry and genotype together: (c)-(b) = 15.7 (P = 3.8×10−5). Testing for the reverse effect of ancestry above and beyond the genotype of rs2814778 produced no signal: (c)-(a) = 0.4 (P = 0.74). These results confirm that rs2814778 is predictive of neutrophil count, above and beyond the effect of ancestry.
To search for additional alleles in the admixture peak that might be associated to neutrophil count beyond the main effect, we genotyped a dense panel of 193 SNPs across the region in 148 individuals with low neutrophil count (<−0.7 standard deviations below the mean) and 74 individuals with high neutrophil count (1.3–2.8 standard deviations above the mean). We chose only individuals for whom we were >99% confident of all African ancestry at the locus, based on genotyping information at flanking markers excluding rs2814778, so that ancestry would not be a confounder of the analysis.
We genotyped these individuals for a set of SNPs chosen using Tagger [12] to capture the great majority of common variation across the admixture peak in both West Africans and Europeans (Materials and Methods). After the genotyping was complete, we had captured 94% of SNPs of >5% minor allele frequency in West Africans, and 96% of SNPs of >5% minor allele frequency in European Americans, both at a correlation of r2>0.8 (Figure 3B).
Case-control association analysis of these 193 SNPs identified only one, rs2814778, that was significantly associated (nominal P = 2.1×10−5; Figure 3A) after a Bonferroni correction for 193 multiple hypothesis tests. Thus, there was no evidence of any allele in the region that is associated to neutrophil count beyond the effect that is already captured by rs2814778.
We genotyped 10,062 self-identified European Americans in the ARIC study for rs2814778, searching for a decreased neutrophil count in association with the null allele. This analysis should have little power if the European American population is in Hardy-Weinberg equilibrium, since FY−/− homozygotes are expected to occur very rarely among Europeans: less than 1/10,000 based on the observed frequency of the null allele in this population (0.34 = 10,062×0.58%×0.58%). Interestingly, we observed 7 European Americans with FY−/− genotypes, a significant excess compared with expectation (P<4×10−9) suggesting that European Americans harbor population substructure with variable levels of African ancestry. Among the FY−/− homozygotes we found a non-significant reduction in WBC associated with the null allele: WBC was observed to be 5.9±2.6 for the 7 FY−/− homozygotes, 5.9±1.8 for the 103 FY+/− heterozygotes, and 6.3±1.9 for the 9,952 FY+/+ homozygotes (P = 0.06 with an additive model and P = 0.35 with a dominant model using 1-sided tests). Genotyping of rs2814778 in 1,339 self-identified European Americans from the Health ABC study identified 26 heterozygous individuals, and none homozygous for FY−/−.
These analyses strongly increase the likelihood that a single nucleotide change at the site of the rs2814778 polymorphism is responsible for low neutrophil counts, and provide no evidence of any other allele contributing a signal. However, these findings do not rule out the existence of undiscovered variants in the admixture peak that are differentiated enough to explain the signal. If such variants existed, then rs2814778 could simply be a marker in linkage disequilibrium with the causative variant rather than being causal itself.
We carried out an analysis in which we systematically ruled out the majority of other nucleotides in the region as potentially contributing to the signal. We examined genomic databases to identify DNA sequence fragments that are known to be of either African or European ancestry and that overlap the admixture peak, and considered nucleotides where all African chromosomes had one allele and all European chromosomes had the other. Based on our modeling in Figure 1, it is likely that the causative variant is sufficiently differentiated that it would be found by this discovery strategy.
We mined shotgun sequencing data from 6 individuals of West African ancestry and 5 individuals of European ancestry (Materials and Methods). We restricted analysis to nucleotides for which we had a high sequence quality score (Neighborhood Quality Score of ≥40) [13], randomly sampling one sequence to represent each individual. We supplemented the shotgun data with the human genome reference sequence, which is comprised of a mosaic of sequence from 5 BAC clones across the region. We determined that 3 of the clones (one from individual CIT978SK and two from individual RPCI-11) were of European ancestry, and 2 were of African ancestry (both from RPCI-11) (Figure 4). Interestingly, since RPCI-11 was heterozygous for African and European ancestry at this locus, this individual is probably African American. Because RPCI-11 is the source for most (∼74%) of the human genome reference sequence [14], we conclude that much of the public human genome reference sequence is that of an African American, and includes a substantial amount of sequence of African ancestral origin.
We found that 82.3% of the admixture peak was covered in at least one chromosome from each population (an average of 2.2× European and 2.1× African coverage). Of the 817 SNPs we identified, 594 could be ruled out as not completely differentiated in frequency between the European and African chromosomes used in SNP discovery, an additional 79 could be ruled out as not sufficiently differentiated across populations based on data from the International Haplotype Map database [10], and 49 could be ruled out by their allele frequencies in our own follow-up genotyping of HapMap samples. Thus, 88.5% ( = (594+79+49)/817) of SNPs discovered in the admixture peak could be ruled out. This allowed us to infer that 72.8% ( = 82.3%×88.5%) of nucleotides can be excluded as causal for the observed major effect on variation in neutrophil count.
These results provide yet another line of evidence that the rs2814778 single nucleotide change may itself be causing low neutrophil levels. There are now five reasons why we believe rs2814778 is likely to be the direct cause for low neutrophil count: (1) rs2814478 falls within the ∼450 kilobase admixture peak. (2) rs2814778 contributes a signal of association above and beyond the admixture signal, showing that the true underlying causal variant is in an even narrower region around rs2814478. (3) rs2814778 is already known to have functional consequences, based for example on past molecular work showing that it affects expression of an antigen on red blood cells and thus modulates resistance to P. vivax malaria. (4) rs2814778 is known to have a frequency differentiation across populations that makes it consistent with the underlying causal variant; a degree of differentiation that is extremely unusual, with only 0.003% of known SNPs in the genome having a differentiation this extreme. (5) Genome sequencing data directly rule out about three quarters of other nucleotides in the admixture peak as containing the variant. (We have not carried out a similar analysis of insertion/deletion polymorphisms, which are known to occur at about a tenth the rate of SNPs.)
We caution that an association study is always correlational, and can never by itself prove a functional effect of an allele. To prove causality, it is essential to follow up any association study with biological work. Nevertheless, the present study provides the best example of which we are aware of taking association analysis to its limit, and using association analysis to demonstrate a likely causal effect. Our study justifies further work to understand the biological mechanism by which a single nucleotide change at rs2814778 probably causes reduced neutrophil counts.
We have used admixture mapping to localize a variant affecting neutrophil levels to a region of about 450 kb centered on the Duffy null locus. We have further shown that the underlying variant must be >91% different in frequency between West Africans and European Americans, placing it among the top 115 HapMap SNPs in the genome in terms of allele frequency differentiation (top 0.003%). Since only one SNP in the admixture peak, the Duffy null polymorphism rs2814778, is known to be this differentiated in frequency across populations, we tested this SNP for evidence of association above and beyond the ancestry effect, and found a signal (P = 3.8×10−5). Finally, we ruled out the great majority of nucleotides across the region, apart from rs2814778, as sufficient to cause the signal.
Methodologically, these results provide a case-example of fine-mapping in a difficult context. We have moved from an initial association signal discovered via mapping by admixture linkage disequilibrium, to a more fine-grained association based on linkage disequilibrium inherited from the ancestral African and European populations, and finally to an analysis where we systematically excluded the great majority of nucleotides in the region as contributing to the association.
The study is also novel in demonstrating the value of mapping in multi-ethnic and admixed populations. The variant could not have been mapped in non-African Americans (either Africans or Europeans), since it is nearly fixed in both populations, showing how studying diverse populations is important in biology. For example, when we genotyped rs2814778 in more than 10,000 European Americans from the ARIC study, we could not obtain a replication despite the large sample size.
The mechanism of low neutrophil count in persons homozygous for the FY− allele is unknown. Interestingly, Yemenite Jews also have a high frequency of the FY− allele [15], which we hypothesize explains the occurrence of reduced neutrophil count in this group. The term “ethnic neutropenia” has been applied to persistently low neutrophil count in both Yemenite Jews and in populations of African ancestry, and the condition is clinically similar in these two populations. Persons with ethnic neutropenia have a reduced capacity to mobilize bone marrow neutrophil reserves in response to corticosteroids, despite normal cellularity and maturation of all cell lines in bone marrow aspirates [16]–[18]. Exercise-induced increments in neutrophil counts (demargination) are also smaller in persons with “ethnic neutropenia” than in healthy volunteers. Thus, the low neutrophil count is not the result of increased sequestration of neutrophils in the marginated granulocyte pools (cells adherent to the endothelium of post-capillary venules) [19].
The FY− allele of rs2814778 has a −46 T to C substitution (non-coding strand) in the Duffy Antigen Receptor for Chemokines (DARC) gene, which disrupts a binding site for the GATA1 erythroid transcription factor [20]. This substitution abolishes gene expression in erythrocytes but not other cell types, such as endothelial cells of the post-capillary venules [21]. The DARC gene product is a seven-transmembrane receptor that selectively binds “inflammatory” chemokines of both the CXC and CC families, including, for example, CXCL8 (interleukin 8) and CCL5 (RANTES), both of which are involved in neutrophil recruitment [22]–[25]. Unlike related chemokine receptors with signaling function, DARC lacks a G-protein binding motif. It is nevertheless capable of internalizing bound chemokine [21], and it has been hypothesized to affect leukocyte recruitment to sites of inflammation through its role in trancytosis of chemokines through endothelial cells [24]–[26]. DARC on red blood cells might affect the number of circulating neutrophils through any of several mechanisms, for example, by modulating the concentrations of chemokines in vascular beds in the bone marrow [25], or by acting as a chemokine “sink” to limit the stimulation and extravasation of circulating neutrophils, or through other mechanisms that are not yet understood. Interestingly, DARC-knockout mice, which lack DARC expression not only on red blood cells but also in other tissues, do not differ from wild-type mice in peripheral blood leukocyte levels [27], but have a different phenotype, of increased bone mineral density [28]. When we tested whether an association with bone mineral density existed in African Americans in the Health ABC cohort, however, rs2814778 genotype was not significantly associated with either total (n = 1,141, P = 0.57) or femoral neck (n = 1,141, P = 0.43) bone mineral density.
The expression of DARC on erythrocytes is known to modulate chemokine levels after endotoxin treatment [29]; thus the FY− allele could potentially have important effects in critically ill patients with sepsis. More subtle effects on innate immunity and inflammation could also exist through DARC modulation of chemokine concentrations in specific vascular and tissue microenviroments[25]. However, we were not able to directly demonstrate an effect of the FY− allele on health. When we carried out tests of association of FY− to a wide range of phenotypes in the Health ABC Study (M. Nalls and T. Harris unpublished data), we did not find association to any phenotype. It is not surprising that health impacts of this variant are subtle, since the allele has risen to nearly 100% frequency in some African populations without being retarded in its rise by natural selection. This forms a sharp contrast with the HBB allele, which confers resistance to Plasmodium falciparum in heterozygous individuals (analogous to FY− conferring resistance to Plasmodium vivax) but causes sickle cell disease in homozygous form, and as a result has never risen to more than about twenty percent frequency in any population.
Another recent study found that the FY− variant of DARC is associated to altered rates of HIV infection and disease progression, potentially suggesting a health effect of the low neutrophil count [9]. The authors found that there is a significantly slower rate of disease progression in FY−/− individuals infected by HIV-1 than in carriers of the FY+ allele. To explore this result, they performed functional studies showing that in vitro, HIV-1 attaches to erythrocytes via DARC and uses it as a means of transfer to target cells. They argued that such transfer might be impaired in FY−/− individuals, leading to slower disease progression. However, these authors also identified a second phenotype that is associated to the FY−/− genotype—a 40% higher rate of acquiring HIV-1 infection—that is difficult to attribute to the same mechanism, since the failure to express DARC on red blood cells might be expected to decrease access by HIV-1 to CD4-positive target cells. Potential explanations are that either low neutrophil counts or altered chemokine concentrations due the FY−/− genotype may have some role in modulating infection. These possibilities should be testable in the laboratory. It is also possible that differing frequencies of the FY− allele reflect stratification within the study population, and that differences in DARC expression are not actually involved in modulating susceptibility to infection or disease progression.
An immediate consequence of our finding is that genotyping of rs2814778 (or measuring Duffy antigen expression on red blood cells) might be used as a diagnostic guide in clinical situations in African Americans, helping to set a baseline expected neutrophil count for patients, and to guide treatment. Reduced neutrophil count has been cited as a potential cause of treatment delay and of less intensive therapy for early-stage breast cancer in African American women, perhaps contributing to ethnic disparities in breast cancer survival [30]. It also may alter the course of cytotoxic therapy for inflammatory diseases such as rheumatoid arthritis and systemic lupus erythematosus, perhaps needlessly. Finally, “ethnic neutropenia” may contribute to a diminished leukocytic response to infection [31]–[34], perhaps resulting in a lowered index of suspicion and delayed diagnosis of infection in selected patients. New studies are needed in each of these clinical settings to incorporate genotyping information from rs2814778 to help in the interpretation of neutrophil counts.
The 6,005 human subjects for this study were drawn from three observational cohorts in which large numbers of phenotypic measurements had been made: the Jackson Heart Study (JHS) [35], the Health, Aging and Body Composition (Health ABC) Study [36], and the Atherosclerosis Risk in Communities (ARIC) Study [37]. From each of these studies, we only included African Americans for whom a complete differential white blood cell count was available, including measurement of neutrophils, bands, lymphocytes, monocytes, eosinophils and basophils. An exception was the Health ABC study, which did not provide a measurement of bands. In addition, we restricted our analysis to individuals for whom we had an admixture scan that passed all our quality control filters; for whom we had a genotype at the Duffy polymorphism, rs2814778; and for whom we had information on gender, age, and body mass index (BMI). All WBC and phenotypic measures were from baseline data collected at the time of enrollment in each study.
The JHS cohort [35] consists of 5,302 self-identified African American men and women recruited between September 2000 and March 2004 from the three counties surrounding Jackson, MS. Unrelated persons aged 35–84 were enrolled from three sources: previous ARIC participants (31% of the total), random selectees from a commercial listing (17%), and members of an age- and sex-constrained volunteer sample (30%). The remaining participants, at least 21 years old, were members of a nested family cohort. A total of 3,945 JHS participants had the required phenotypic data and were successfully genotyped for a panel of admixture mapping markers. A subset of 2,871 was included in our analysis after randomly dropping samples of related individuals until there was only one individual included per family. Of these, 1,969 were “JHS unique” samples that were present only in JHS, and 902 were “JHS-ARIC overlap” samples, representing persons who had participated in both the JHS and ARIC studies. For the “JHS-ARIC overlap” samples, we averaged the baseline measurements for each individual at the time of their entry into each cohort, an average of 14 years apart.
The Health ABC cohort [36] consists of 3,075 men and women aged 70–79 who were enrolled between April 1997 and June 1998. All were Medicare beneficiaries living near Pittsburgh or Memphis and all reported having no difficulty performing basic physical activities. Of the 1,281 participants who identified themselves as African American, 658 had complete genotype and phenotype data and were included in the current study. Of the participants who identified themselves as European American, 1,331 were analyzed for the purpose of comparison with African Americans.
The ARIC cohort [37] consists of 15,792 randomly-selected participants aged 45–64 who were recruited between November 1986 and December 1989, in roughly equal numbers, from field centers in Jackson, MS, Minneapolis, MN, Forsyth County, NC, and Washington County, MD. The cohorts of the latter three field centers represent the ethnic mix of their communities. The Jackson-based cohort (n = 3,728) was limited to self-identified African Americans, and comprised 87.4% of all African Americans in ARIC (1,626 Jackson-based participants were later enrolled in JHS). A total of 3,378 African American participants were included in the present analysis after applying all data filters. Of these, there were 2,476 “ARIC unique” participants who were present only in ARIC. There were 902 “JHS-ARIC overlap” samples as described above. A total of 10,062 European American ARIC participants were also genotyped and analyzed with respect to a single variant, rs2814778.
For all three studies, cells in EDTA-anticoagulated venous blood were counted using a Beckman-Coulter Counter (Beckman Coulter, Inc., Fullerton, CA), which combines measures of electrical conductivity and light scatter to distinguish cell lineages in suspensions of unstained leukocytes, yielding an overall WBC and relative proportions for each of six leukocyte subgroups: neutrophils, monocytes, lymphocytes, basophils, eosinophils, and “band forms”, expressed as a percentage of total WBC. Absolute counts were obtained by multiplying the differential count (a percentage) by total WBC.
To create a phenotype for analysis, all counts were rank-ordered within one of the four groups of samples (Health ABC, JHS only, ARIC only and JHS-ARIC overlap), and then assigned a percentile. An inverse normal transformation was used to translate this percentile into a normally distributed phenotype.
Genotyping of African American samples on the admixture mapping panels was performed using the Illumina BeadLab platform [38], which can analyze a custom panel of 1,536 SNPs. We have developed three consecutive versions of a custom admixture mapping panel, each providing incrementally better coverage of the genome than the previous version, and all yielding excellent coverage. A total of 1,119 samples were genotyped in the “Phase 2” panel [39] (Health ABC samples and 16% of JHS samples) and 4,886 samples were genotyped in the “Phase 3” panel [6] (ARIC samples and 84% of the JHS samples). All panels include the rs2814778 polymorphism in the DARC gene. The genotyping of the Health ABC and JHS samples was carried out at the Broad Institute of Harvard and MIT in Cambridge as previously described [6]. Genotyping of the ARIC samples was carried out at the Center for Inherited Disease Research (CIDR) in Baltimore, MD.
Genotyping of the rs2814778 polymorphism in 10,062 ARIC European American samples and 1,339 Health ABC European American samples was done using the ABI TaqMan technology [40].
We used built-in data quality checks in the ANCESTRYMAP software [11],[39],[41],[42] to remove SNPs that were not appropriately intermediate in frequency in African Americans compared to the West African or European American ancestral populations, or that had evidence of being in linkage disequilibrium (LD) with each other in these ancestral populations [11]. After this filtering, the JHS samples had 1,265–1,532 SNPs available for analysis, the Health ABC study samples had 1,128–1,385 SNPs, and the ARIC study samples had 1,277–1,529 SNPs.
To refine the peak of admixture association, we used the Sequenom iPLEX platform [43] to genotype all JHS samples more densely in the region of highest interest on chromosome 1 (153.5–157 Mb in Build 35 of the reference sequence). After filtering to remove SNPs in LD in the ancestral populations or with poor genotyping performance, we had data from 9 markers across this region in JHS (rs2768744, rs2309879, rs7528684, rs1587043, rs857859, rs2814778, rs11265198, rs2494493 and rs11265352), compared with 2 in the other studies (rs2768744 and rs2814778).
The “C” allele of rs2814778 (also “FY−”) is known to be almost completely correlated to West African ancestry at the chromosome 1q23.2 locus. We therefore used this allele as a marker to study the epidemiological association of African ancestry to various WBC phenotypes (Figure 1 and Tables 1 and 2). In addition to using FY− as a surrogate for ancestry, in the final analyses of this study (Figure 3 and Table 4), we took advantage of the fact that the correlation between this allele and African ancestry, though >99%, is not perfect. Thus, we could test whether the allele is more predictive of neutrophil count than is African ancestry itself.
To test for association of the chromosome 1 locus to counts of leukocytes other than neutrophils, we carried out a regression analysis between absolute counts of neutrophils and absolute counts of each of the other white blood cell types. This generated a residual value for each cell type after correcting for the effect of neutrophil count. We then carried out 2-sided tests for association of each of these residuals to carrier status for European ancestry at the chromosome 1 locus (defined as having at least one FY+ allele at rs2814778), using Z-scores to indicate the difference, in standard deviations, in the population means between groups of samples. These Z scores can be approximately translated into a P-value by referring to the corresponding percentile in the cumulative normal distribution function.
The ANCESTRYMAP software [11] was used to better define the region of chromosome 1 associated to neutrophil count. Since the software is optimized for dichotomous traits, we divided the 6,005 samples into 12 strata (with 399–593 samples each) based on their normally transformed neutrophil counts (Table 3). For each stratum, we identified a risk model (increased probability of observing a sample with 1 or 2 copies of European ancestry compared with the expectation from the genome-wide average) that optimized the peak LOD score at the locus. This ranged from a 25-fold decrease in the relative probability of European ancestry for individuals with neutrophil counts <−1.5 standard deviations below the mean, to a 6.5-fold increase for those with neutrophil counts >1.5 standard deviations above the mean. For each stratum, the risk model and LOD score at the chromosome 1 locus are given in Table 3.
To use these strata to define an admixture peak, we carried out an admixture scan for each group separately, and then summed the LOD scores at loci interpolated every tenth of a centimorgan. The peak LOD score was 363.1 as shown in Table 3. The 99% credible interval of 155.957–156.408 Mb in Build 35 of the human genome reference sequence was determined by the region where the score was within log10(e6.63/2) = 1.44 of its maximum (Figure 2). This is calculated from a likelihood ratio test, using the fact that a χ2 statistic with 1 degree of freedom of 6.63 corresponds to P = 0.01.
To assess whether genotype at rs2814778 is more predictive of neutrophil count than ancestry, we calculated two numbers for each DNA sample. First, we recorded whether the individual was a carrier of the FY+ “functional” allele ( = 1) at rs2814778, or was homozygous for the FY− “null” allele ( = 0) (homozygosity for the “null” allele abolishes expression of the Duffy antigen on red blood cells but not on other cell types). Second, we used the ANCESTRYMAP software [11] to estimate the probability that in the region spanning this SNP, at least one of the individual's chromosomes was of European ancestral origin. Importantly, we included the genotype of rs2814778 in the ancestry estimation. Thus, if neutrophil count were better correlated to rs2814778 genotype than to an ancestry estimate that included information from both rs2814778 and closely neighboring SNPs, it would indicate that the neighboring SNPs did not add relevant information, and that rs2814778 is either the causal variant or is in strong LD with the causal variant.
To determine whether genotype at rs2814778 or ancestry at the chromosome 1 locus was the better predictor of neutrophil status, we carried out three regressions:
To test for evidence of an association of the genotype at rs2814778 above and beyond ancestry, we subtracted the χ2 statistics of (c)-(b). To test for evidence of association to ancestry above and beyond SNP genotype, we subtracted the χ2 statistics of (c)-(a).
To test whether there were SNPs apart from rs2814778 that contributed evidence for association at the chromosome 1 locus, we densely genotyped a subset of especially informative samples. To select cases and controls for fine-mapping, we identified individuals from JHS for whom we were >99% confident of African ancestry on both chromosomes at the admixture peak. By limiting cases and controls to persons with more confident estimates of entirely African local ancestry, it was easier to detect whether any signal of association was significant above and beyond the admixture association. For this analysis, the estimate of ancestry at the admixture peak using ANCESTRYMAP [11] excluded SNPs within the peak.
Among individuals who had >99% confidence of entirely African ancestry at the locus, we identified 696 individuals who had a “low” neutrophil count, defined based on visual inspection of the distribution as an absolute count of <2,100/mm3 (and corresponding to <0.7 s.d. below the population mean for the entire JHS sample). We also identified 77 individuals who had a “high” neutrophil count, defined as an absolute count of 5,100/mm3–9,100/mm3 (1.3–2.8 s.d. above the population mean). For genotyping, we selected a random subset of individuals with “low” neutrophil count, and all of the individuals with “high” neutrophil count. We successfully genotyped 148 subjects with low and 74 subjects with high neutrophil count that we could use in this analysis.
To identify additional SNPs across the admixture peak that might be associated with neutrophil levels, we used the Tagger software [12] to choose a panel of SNPs from the International Haplotype Map database [10] that captured all SNPs of >5% minor allele frequency in West African samples with a correlation of r2>0.8. Forcing these SNPs into the analysis, we chose additional SNPs across the region until we had similarly captured all SNPs in European Americans with >5% minor allele frequency. All SNPs identified in this way were selected for genotyping on the Sequenom iPLEX platform [43]. Cases and controls were successfully genotyped at a densely spaced panel of 193 tag SNPs across the ∼450 kilobase admixture peak. Each SNP was tested using a χ2 statistic assuming an additive effect on neutrophil count for each additional copy of the allele. We did not test a dominant or recessive model because none of the genotyped SNPs (apart from rs2814778) had a frequency differentiation across populations consistent with that SNP explaining the admixture signal (and thus being the main effect SNP; see above). We chose an additive model to search for variants that might modulate the neutrophil count beyond the main effect because it is known that for substantial minor allele frequencies this provides reasonable power to detect an association, whether the true underlying effect is dominant, additive, or recessive [44].
While we found that rs2814778 was more predictive of neutrophil count than ancestry, we were concerned that the variant might not itself be causal, but instead only in LD with the causal variant. To search for additional candidate SNPs in the region that are highly different in frequency between Africans and Europeans, we examined shotgun genome sequence data derived from public databases, from 6 individuals who were known to have West African ancestry across the region (NA18507, NA18517, NA19129, NA19240, NA17109 and NA17119), and 5 individuals who were known to have European ancestry across the region (NA12156, NA12878, NA07340, HuAA and HuBB). These sequences include 4 West Africans and 2 Europeans examined as part of a fosmid end-sequencing project [45], 2 European Americans sequenced as part of the Celera Genomics human genome sequencing project [46],[47], 1 European American sequenced for the purpose of SNP discovery [48],[49], and 2 African Americans also sequenced for SNP discovery [48],[49], who we determined had entirely African ancestry at the locus by using the ANCESTRYMAP software [11] and by unpublished methods (Simon Myers, Alkes Price and Alon Keinan, personal communication). For each individual, we only analyzed nucleotides for which we had a high quality base call at the locus (Neighborhood Quality Score of ≥40) [13]. At sites where we had more than 1× coverage, we randomly sampled one sequence to represent the individual.
To determine the ancestry of the human genome reference sequence across the admixture peak, we first observed that it was spanned by a mosaic of 5 fully sequenced Bacterial Artificial Chromosomes (BACs), each representing a clone of 86,000–196,000 base pairs (http://genome.ucsc.edu). The problem of determining ancestry of the human reference sequence across the region thus amounts to determining the ancestry of each of the clones separately. To do this, we obtained the allele of the human genome reference sequence at each of 284 HapMap SNPs across the region that had been genotyped in both West Africans and European Americans. For each window of 8 consecutive SNPs in HapMap, we calculated the likelihood that the human reference sequence was of African or European ancestry. To determine this likelihood empirically, we compared the findings to 120 phased European chromosomes and 120 phased African chromosomes from the HapMap database, counting the number of matches to the human reference sequence in each population over that window. We conservatively added 1 to the counts of the tested haplotype for each population to preclude an estimate of zero probability for either population. Under the assumption that only European and African ancestry were possible, the results showed with confidence that 3 of the clones were of European ancestry and 2 were of African ancestry (Figure 4).
|
10.1371/journal.ppat.1000364 | Distinct Roles for FOXP3+ and FOXP3− CD4+ T Cells in Regulating Cellular Immunity to Uncomplicated and Severe Plasmodium falciparum Malaria | Failure to establish an appropriate balance between pro- and anti-inflammatory immune responses is believed to contribute to pathogenesis of severe malaria. To determine whether this balance is maintained by classical regulatory T cells (CD4+ FOXP3+ CD127−/low; Tregs) we compared cellular responses between Gambian children (n = 124) with severe Plasmodium falciparum malaria or uncomplicated malaria infections. Although no significant differences in Treg numbers or function were observed between the groups, Treg activity during acute disease was inversely correlated with malaria-specific memory responses detectable 28 days later. Thus, while Tregs may not regulate acute malarial inflammation, they may limit memory responses to levels that subsequently facilitate parasite clearance without causing immunopathology. Importantly, we identified a population of FOXP3−, CD45RO+ CD4+ T cells which coproduce IL-10 and IFN-γ. These cells are more prevalent in children with uncomplicated malaria than in those with severe disease, suggesting that they may be the regulators of acute malarial inflammation.
| While Tregs have been implicated in regulation of the immune response to chronic infections, their potential in determining disease outcome in acute infections is unclear. In this study we have found that Tregs are unable to control the florid inflammation during acute, severe P. falciparum malaria infections, suggesting that this component of the immunoregulatory arsenal may be rapidly overwhelmed by virulent infections. Further, we identified, for the first time in an acute human infection, a population of IL-10-producing Th-1 effector cells and found that IL-10-producing Th-1 cells were associated with development of uncomplicated as opposed to severe malaria, leading us to suggest that such “self-regulating” Th-1 cells may contribute to clearing malaria infections without inducing immune-mediated pathology. In addition, we found evidence that malaria-induced Tregs may limit the magnitude of malaria-specific memory responses detectable 28 days later, which may reduce the risk of immune-mediated pathology upon reinfection and may explain how immunity to severe disease can be gained after as little as one or two infections. We conclude that vaccines designed to induce cell-mediated responses should be assessed for their ability to induce IL-10 producing Th-1 cells and Tregs.
| The clinical spectrum of P. falciparum infection ranges from asymptomatic parasite carriage to a febrile disease that may develop into a severe, life-threatening illness. The factors that determine disease severity are not completely understood but are likely to include both parasite and host components [1]–[3]. Ultimately, the interplay between the parasite and the immune response likely determines the outcome of the infection [4]. Although sterile immunity - completely preventing re-infection - is hardly ever seen and protection against clinical symptoms of uncomplicated disease is only acquired after repeated infections [5], immunity to severe disease and death may be acquired after as few as one or two infections [6] suggesting that different immune mechanisms underlie these different levels of immunity.
While there is a growing consensus that killing of malaria parasites or malaria-infected red blood cells requires the synergistic action of antibodies and cell-mediated immune responses [5],[7], the mechanisms conferring protection against severe disease are less clear. Given that pathology of severe disease has repeatedly been linked to sustained and/or excessive inflammatory responses [4], acquiring the ability to regulate these responses adequately may be a key determinant of immunity that protects against severe disease [8]. Thus, while an early inflammatory response is needed to control parasite replication in human P. falciparum malaria [9]–[11], excessive levels of pro-inflammatory cytokines such as TNF-α [12]–[15], IFN-γ, [16],[17], IL-1β and IL-6 [14],[18],[19] are associated with severe pathology. Conversely, low levels of regulatory cytokines such as TGF-ß have been associated with acute [20] and severe malaria [21],[22], a relative deficiency in IL-10 was seen in those who succumbed to severe malaria [23], significantly lower ratios of IL-10 to TNF-α were found in patients with severe malarial anaemia [24],[25], and high ratios of IFN-γ, TNF-α and IL-12 to TGF-β or IL-10 were associated with decreased risk of malaria but increased risk of clinical disease in those who became infected [26]. In summary, therefore, immunity against severe malaria may depend upon the host's ability to regulate the magnitude and timing of the cellular immune response, allowing the sequential induction of appropriate levels of inflammatory- and anti-inflammatory cytokines at key stages of the infection.
Given these associations between severe disease and exacerbated immune pathology, a number of studies have explored the role of CD4+CD25hiFOXP3+CD127−/lo regulatory T cells (Tregs) in determining the outcome of malaria infection. Induced and/or activated in response to malaria infection [27], Tregs may be beneficial to the host in the later part of the infection - when parasitaemia is being cleared - by down-regulating the inflammatory response and thereby preventing immune-mediated pathology. On the other hand, if Tregs mediate their suppressive effects too early, this could hamper the responses required for initial control of parasitaemia, permitting unbridled parasite growth which may also lead to severe disease. Malaria-specific induction of Tregs has been observed in a variety of experimental malaria infections in mice [28]–[31], but their role in preventing severe malarial pathology is unclear. Thus, in BALB/c mice infected with a lethal strain of P. yoelii, ablation of Treg activity by depletion of CD25+ cells either allowed mice to control parasitaemia and survive [32] or had no impact on the course of disease [33]. Depleting CD25+ cells of BALB/c mice infected with either P. berghei NK65 [34] or P berghei ANKA [35] reduced neither parasitaemia nor mortality, but increased the severity of symptoms in the diseased mice, suggesting at least some benefit from Tregs in this model. Rather oddly, infection of CD25+ T cell-depleted BALB/c mice with P. chabaudi adami DS led to increased parasitaemia and more severe anaemia [30]. Finally, CD25+ T cell depletion around the time of parasite inoculation reduced the incidence of experimental cerebral malaria in C57BL/6 mice infected with P. berghei ANKA in two independent studies [29],[31], but not when CD25+ T cells were depleted 30 days prior to infection [31]. Whilst various explanations have been offered for these discrepant results, including differences in the various strains of mice and parasites employed, the microbial microenvironment in which the mice are kept, and the precise CD25 depletion protocols employed, these studies are currently not very helpful when trying to understand the role of Tregs in human malaria infections.
Malaria naïve individuals undergoing experimental P. falciparum sporozoite infection showed an increase in FOXP3 mRNA expression and expansion of Tregs 10 days after infection; Treg induction correlated with high circulating levels of TGF-β, low levels of pro-inflammatory cytokines and rapid parasite growth [27] suggesting - but not proving - that Treg activation early in infection may inhibit the development of effective cellular immunity. More recently, we have observed that Treg populations appear to be transiently expanded and activated during the malaria transmission season in individuals from a malaria endemic community [36], again suggesting that naturally acquired malaria infection can drive the expansion and activation of Tregs. However, although Tregs have been implicated in IL-10-mediated down-regulation of Th1-like responses in the placenta of malaria-infected women [37] and reduced Treg frequencies and function have been linked to enhanced anti-parasite immunity in certain ethnic groups in West Africa [38] the potential for Tregs to influence the clinical outcome of malaria infections is still unclear.
To investigate the role of Tregs during clinical malaria infection, we have compared cellular immune responses of children with either severe or uncomplicated malaria. Interestingly, although we did not observe any significant differences in Treg numbers or function between severe and uncomplicated malaria cases, our data do indicate that malaria-induced Tregs may limit the magnitude of malaria-specific Th1 memory responses and thus moderate pro-inflammatory responses to subsequent infections, providing a possible explanation for the very rapid acquisition of immunity to severe malaria. Moreover, we have identified a population of FOXP3−, CD45RO+, CD4+ T cells which co-produce IL-10 and IFN-γ and which are more prevalent in children with uncomplicated malaria than in those with severe disease. We suggest that these IL-10 producing effector T cells may contribute to clearing malaria infection without-inducing immune-mediated pathology.
Immune responses of 59 Gambian children with severe P. falciparum malaria were compared with those of 65 children with uncomplicated clinical malaria and with 20 healthy (control) children of similar age and recruited from the same study area at the same time (Table 1). On admission, only 12 (9.4%) patients had a white blood cell count (WBC) above the age-specific norm and there was no significant difference in median WBC count between uncomplicated and severe cases, suggesting that few if any children had a concomitant systemic bacterial infection. No difference was observed in the differential WBC between the two groups. As expected, [39]–[41] numbers of lymphocytes, CD3+ and CD4+ T cells were significantly lower during the acute disease than during convalescence in both the severe and the uncomplicated groups (Table 1A). Parasite density on admission was two-fold higher in patients with severe malaria than in those with uncomplicated malaria; severely ill children also had significantly lower hemoglobin levels and were on average 2.4 years younger than children with uncomplicated disease. The number of P. falciparum clones per clinical isolate ranged from 1 to 4 (with an overall mean of 2 (CI95%: 1.8–2.1)), and – as has been observed previously [42] - did not differ significantly between the three groups (p = 0.3, Table 1B). Other factors potentially confounding immune responses, such as the degree of malnutrition or intestinal helminth infections were of similarly low prevalence in both severe and uncomplicated malaria cases and were not associated with severity of disease (Table 1B).
For statistical analysis, patients were classified as uncomplicated or severe cases, with the latter being further subdivided into those patients suffering from cerebral malaria (CM), severe anaemia (SA) or severe respiratory distress (SRD) (grouped together as SA) and those suffering only from severe prostration (grouped as SB). Data were analysed using linear regression, with a random effect to allow for the within subject measurements over time, adjusting for age, sex, duration of prior symptoms and numbers of clones causing the infection. Due to the multiplicity of comparisons that were made within the model, resulting from multiple responses and multiple comparisons within response, hypotheses rejected with a probability of less than 0.012 have a false discovery rate of 5% [43].
We hypothesized that children with severe malaria would have fewer circulating Tregs than children with uncomplicated malaria, or that Tregs of severely ill children would be less active than those of children with uncomplicated disease. However, the proportion of cells expressing a Treg phenotype (defined by flow cytometry as CD3+CD4+ lymphocytes being FOXP3+and CD127−/low; Figure 1A, 1B, and 1C) was similar in the acute (D0) and the convalescent phase (D28) for both uncomplicated and severe cases (SA+SB) and in healthy control children; on average, 2–3% of CD4+ T cells expressed the regulatory phenotype (Figure 1D). However, when the number of cells expressing a Treg phenotype was calculated using lymphocyte and monocyte counts from the differential WBC, we found that the absolute numbers of Tregs (per litre of blood) were significantly and similarly elevated in both severe and uncomplicated malaria cases during convalescence when compared to the acute phase (p<0.001) or when compared to the control group of healthy children (p = 0.037) (Figure 1E). A similar kinetic was observed for FOXP3 mRNA levels (Figure 1F).
Although not supportive of our original hypothesis, these observations are consistent with the notion that acute malaria infection drives expansion of Treg populations which then persist for some weeks to maintain immune homeostasis during the contraction phase of the effector response [36]. In accordance with this notion, and in agreement with our previous observation that increased levels of Tregs were associated with faster parasite growth during the early stages of blood stage infection [27], we observe here in - children with either severe or uncomplicated malaria infections - a significant positive correlation between parasite density and the frequency of Tregs within the CD4+ T cell population (p = 0.002, Figure 2).
Since fully differentiated Tregs predominantly express an activated/memory phenotype [44] T cells from children with severe and uncomplicated malaria were analysed for expression of CD45RO (Figure 3A and 3B). In both uncomplicated and severe cases, the proportion of all T cells expressing CD45RO was significantly higher (p<0.001) during acute infection than during convalescence (data not shown). Irrespective of disease severity, more than 90% of Tregs expressed CD45RO during the acute phase of infection but expression of this marker decreased significantly (to approx 70%) during convalescence (Figure 3C). Likewise, the median fluorescence intensity (MFI) of CD45RO staining was 1.5 fold higher (p = 0.0025) during acute disease than during convalescence (Figure 3D). Taken together, these data indicate that in both uncomplicated and severe cases of malaria Tregs are predominantly of a memory phenotype and are activated during acute malaria infection.
Three different indicators were used to assess the regulatory potential of Tregs during acute malaria infection. Firstly, using a classical anti-CD25 depletion assay, we assessed the ability of Tregs to suppress P. falciparum shizont extract (PfSE)-driven lymphocyte proliferation. Anti CD25 treatment removed approximately half (geometric mean 48.8%; CI95%: 41–58%) of the CD4+ T cells that were FOXP3+CD127−/low and this was associated with a 1.76 fold and 1.57-fold (geometric means) increase in PfSE-induced lymphoproliferation in severe and uncomplicated cases, respectively, with no significant difference between the groups (p = 0.343, Figure 4A).
Next, since reduced expression of SOCS-2, a member of the suppressors of cytokine signaling family confined to Tregs [45], has been linked to impaired Treg function in Africans [38] we compared SOCS-2 mRNA levels among severe and uncomplicated cases. While SOCS-2 levels were found to be significantly reduced during acute disease compared to convalescence (p = 0.0076), no difference was observed between those with severe and those with uncomplicated malaria (data not shown).
Finally, since high concentrations of TNF-α have been reported to impair Treg activity (by upregulating and then signaling via TNFR2, leading to decreased FOXP3 mRNA and protein expression [46]), and the functional impairment of Tregs observed in rheumatoid arthritis patients can be reversed by anti-TNF-α antibodies [47], we considered the hypothesis that the high levels of TNF-α seen in severe malaria patients [13],[15], might upregulate TNFR2 and impair Treg function. TNFR2 expression on Tregs was assessed by flow cytometry (Figure 4B and 4C). However, although a significantly higher proportion of Tregs expressed TNFR2 (Figure 4D) - with higher MFI (data not shown, p = 0.028) - during acute disease compared to convalescence, no difference was seen in TNFR2 expression on Tregs between severe and uncomplicated cases (Figure 4D). Moreover, there was no correlation between plasma levels of TNF-α and TNFR2 expression on Tregs, neither among severe or uncomplicated cases nor among all cases combined; neither did we observe any inverse correlation between TNF-α concentration and FOXP3 expression. Rather, the MFI of TNFR2 on Tregs was positively correlated with the MFI of FOXP3 in Tregs (r: 0.476; p<0.0001). Thus, our data seem to be more in line with data from mice suggesting that the interaction of TNF-α with TNFR2 on Tregs promotes their expansion and upregulation of FOXP3 [48] than with the data from studies of human rheumatoid arthritis.
Since the balance of T-effector to Treg responses is likely to be as important, or more important, than the absolute levels of either [26], we compared the ratio of the levels of mRNA for the Th1 transcription factor T-BET with those for FOXP3, currently considered the best marker for Tregs, and the ratio of T-effector cells (defined as CD3+CD4+CD25+FOXP3−, T-effector) over Tregs among the various groups. As shown in Figure 5A, in all groups the T-BET/FOXP3 ratio was significantly higher during acute disease than during convalescence and a similar, albeit not significant, trend was observed for the ratio of T-effector/Tregs (data not shown). Moreover, since the absolute number of circulating T-effector cells was significantly higher in severe cases than in uncomplicated cases (p = 0.01, data not shown), the T-effector/Treg ratio tended to be higher among severe cases than uncomplicated cases on day 0 (p = 0.039) and a similar trend was seen for the T-BET/FOXP3 ratio (p = 0.058). The ratio of FOXP3 to GATA-3 (Th2 lineage factor) mRNA was similar for both time points in all groups (data not shown) but the Th1/Th2 ratio (T-BET/GATA-3 mRNA) was significantly higher during acute disease in children with CM, SA or SRD compared to those suffering from severe prostration (p = 0.0075), indicating that the expansion of the T-effector population is biased towards Th1 responses.
These data confirm previous studies indicating a shift towards a more inflammatory response during acute and severe malaria but, significantly, our data extend the previous observations by revealing that this inflammation is not balanced by a commensurate increase in Treg function. Indeed, our data strongly suggest that the potent inflammation induced during an acute malaria infection overwhelms the normal homeostatic capacity of the immune system and, in particular, that the Treg response in children with severe malaria is insufficient to balance a much stronger Th1 effector response.
To investigate further the dynamics of pro-inflammatory/regulatory responses during clinical malaria, plasma concentrations and mRNA transcripts of inflammatory (IFN-γ, TNF-α) and regulatory cytokines (IL-10) were assayed. In accordance with previous observations, plasma concentrations of IFN-γ, TNF-α and IL-10 were all significantly higher during acute disease than during convalescence, with significantly higher levels in severely ill children compared to uncomplicated cases (Figure 5B, 5C, and 5D). Levels of mRNA transcripts for IL-10 and IFN-γ were also significantly elevated in all groups during acute disease, but there was no significant difference between severity groups (Figure S1A and S1B). For both severe and uncomplicated cases, levels of IFN-γ mRNA were highly correlated with levels of IL-10 mRNA during the acute phase (severe: r = 0.833 p<0.001, uncomplicated: r = 0.693 p<0.001), suggesting that IFN-γ production is being balanced by IL-10 production. Interestingly, IFN-γ mRNA levels on day 0 correlated with FOXP3 mRNA on day 0 for both severe (r: 0.39 p = 0.003) and uncomplicated cases (r: 0.44 p = 0.0001), suggesting that IFN-γ may also be driving FOXP3 expression.
The balance of pro-and anti inflammatory cytokine responses clearly changed with time, but somewhat surprisingly, there were no marked differences in cytokines ratios between children with differing levels of disease severity. Thus, the ratios of TNF-α or IFN-γ to IL-10 on day 0 were similar in all disease severity groups (Figure 5E and 5F) and ratios of IFN-γ to IL-10 mRNA were similar in all disease severity groups both on day 0 and day 28 (Figure S1F). However, IFN-γ mRNA levels were on average only 3.2-fold higher on day 0 than day 28 but IL-10 mRNA levels were 29-fold higher on day 0, resulting in a significantly lower IFN-γ/IL-10 mRNA ratio on day 0 than day 28 (Figure S1F).
IL-10 is a crucial immunoregulatory cytokine in both human [23] and murine [33],[49] malaria; we have recently identified CD4+ T effector cells as a major source of IL-10 [33], but the source of IL-10 in human malaria infection is unknown. In other protozoal infections of mice CD4+ effector T cells that co-produce IFN-γ and IL-10 have been identified [50]–[52]. We therefore cultured freshly isolated PBMCs from 30 children with acute malaria (17 severe, 13 uncomplicated) and 20 healthy control children, with or without PMA and Ionomycin (PI), for 5 hours and analyzed them for the presence of intracellular IL-10 and IFN-γ by flow cytometry (Figure 6A). No cytokine production was observed in unstimulated cells (data not shown), and PBMC from healthy children failed to produce any IL-10 in response to PI (data not shown), indicating that stimulation with PI predominantly induces cytokine production from recently activated cells.
By contrast, distinct populations of IL-10+ and IFN-γ+ cells were seen among the PI-stimulated cells from children with acute severe or uncomplicated malaria, with a small but easily distinguishable population of cells (approx 1% of all PBMC) producing both cytokines simultaneously (Figure 6A, right plot). In both severe and uncomplicated cases, IL-10 producing cells were predominantly CD45RO+ CD4+ T cells (Figure 6B and 6C) and were almost exclusively FOXP3− and CD25− (Figure 6D). Moreover, although a transient increase of FOXP3 in activated human T-effector cells has been reported [53], in our hands less than 1% (median 0.97%, CI95%: 0.67–1.27%) of IFN-γ producing cells were FOXP3+ (Figure 6E).
Overall, among children with acute malaria, approx 4% of PI-stimulated CD4+ T cells produced IL-10 and approx 8% produced IFN-γ and neither the proportions of cells producing one or the other cytokine (Figure 6F) nor the ratio of IFN-γ/IL-10 producing cells (data not shown) differed significantly between severe and uncomplicated cases. However, intriguingly, the proportion of CD4+ T cells simultaneously producing IL-10 and IFN-γ was three fold higher in uncomplicated cases than severe cases (geometric mean 5.2% vs 1.6%, p = 0.041, Figure 6F). Moreover, the proportion of IFN-γ+ CD4+ T cells that also produce IL-10 was almost twice as high among uncomplicated cases as among severe cases (p = 0.045, Figure 6G).
Taken together, these data indicate that during acute, uncomplicated or severe, malaria infections IL-10 producing cells are overwhelmingly T effector cells and that Th1 effector cells that also produce IL-10 are more prevalent in children with uncomplicated malaria than in children with severe malaria.
It has been reported that Tregs present at the time of infection [54] or vaccination [55] may restrict the development of subsequent Th1 memory responses. To determine whether Tregs present during acute malaria infection might similarly affect the induction of immunological memory, we compared FOXP3 mRNA levels on day 0 with malaria specific IFN-γ memory responses (as assessed by PfSE-specific cultured ELISPOT) among PBMC collected from 34 of our convalescent malaria patients (19 severe and 15 uncomplicated) on Day 28.
As shown in Figure 7A, cells from uncomplicated and severe cases mounted similarly strong IFN-γ memory responses following culture with PfSE. When plotted against FOXP3 mRNA levels measured on day 0, a linear by linear hyperbolic fit revealed that higher levels of FOXP3 mRNA on day 0 were highly significantly (p = 0.009) associated with lower malaria-specific IFN-γ memory responses on Day 28, suggesting that Tregs induced during the acute infection may limit the magnitude of subsequent Th1 responses (Figure 7B). For neither group could a significant effect of parasitaemia on the memory response be observed (r = −0.12 p = 0.962 for severe and r = 0.015, p = 0.957 for uncomplicated cases).
We hypothesized that the balance of inflammatory to regulatory immune responses would be biased towards a more inflammatory response in children with severe malaria than in children with uncomplicated malaria, that this balance would be restored during convalescence and – crucially – that this would be associated with differences in the proportion, absolute number or function of circulating classical (CD4+ CD127−/lo FOXP3+) regulatory T cells. In partial support of these hypotheses, the number of cells expressing a Treg phenotype and FOXP3-mRNA levels were both significantly higher during convalescence than during the acute clinical episode and the ratio of the Th1 transcription factor T-BET to the Treg transcription factor FOXP3 was significantly higher during acute disease than during convalescence in both severe and uncomplicated cases, compatible with the notion that Tregs fail to sufficiently regulate pro-inflammatory responses which might contribute to the onset of symptomatic malaria infection.
Given our previous observation of Treg expansion during the pre-patent phase of malaria infection [27], we suggest that Tregs are induced/activated shortly after parasite emergence from the liver, that their numbers in peripheral blood then decline as a result of sequestration of CD4+ T cells during acute disease [40],[41],[56] and then, as has been described for other T cell subsets [39],[57], Tregs regain access to the circulation after malaria is cured. The significant positive correlation of Treg numbers with parasitaemia, as well as the correlation between FOXP3 mRNA and IFN-γ mRNA levels in acute samples, further supports the notion that the initial infection induces a proportional increase in Tregs, attempting to balance the effector T cell response, and is in line with the recently proposed concept that antigenic challenge will give rise to an antigen specific Treg response, proportional in size to the inflammatory response [58]. Moreover, the Tregs circulating during acute malaria infections almost exclusively expressed an activated memory phenotype suggesting that they have expanded from a pre-existing pool of memory T-cells. This interpretation would be in line with recent elegant work in humans demonstrating that Tregs are derived by rapid turnover of memory populations in vivo [59], and with data from murine studies where, after CD25-depletion, malaria infection very rapidly drives differentiation of Tregs from circulating mature CD4+ T cells [60]. Obviously, it would be of interest to study the relationship between Tregs and effector T cell kinetics and parasite biomass, which is not readily measurable. Future studies may explore the usefulness of P. falciparum Histidine Rich Protein 2 in this context, which has recently been suggested as a surrogate marker for parasite biomass [61].
However, despite clear evidence of Treg induction and reallocation during acute malaria infection, we could not find any robust differences in Treg parameters between children with severe and uncomplicated disease. Thus, neither Treg numbers nor FOXP3 mRNA levels differed significantly between children with uncomplicated malaria and those with severe malaria, and three different indicators of Treg function - their capacity to suppress lymphoproliferation, their expression of SOCS-2 [45] and TNFR2 [46],[48] were all similar in severely ill children and children with uncomplicated disease. Furthermore, the similar distortion in the T-BET/FOXP3 mRNA ratio during acute disease and the lack of any marked differences between the two groups in ratios of inflammatory to anti-inflammatory cytokines, as well as the close correlation between IFN-γ and IL-10 in both groups which is in line with previous observations in experimental human malaria infections [62], suggests that the systemic shift towards a pro-inflammatory immune response is similar in children with either severe or uncomplicated disease.
At first glance, these data do not appear to support the hypothesis that deficiencies in Treg function underlie the tendency of some children to develop severe, life threatening malaria.
However, we did observe significantly higher Th1 effector responses (more T-effector cells, higher concentrations of IFN-γ and TNF-α) in severely ill children than in children with uncomplicated disease, suggesting that the classical FOXP3+ Treg response that develops during acute malaria infection may be insufficient to balance the florid effector T cell response that develops particularly in children with severe disease. This would be in line with evidence showing that as the strength of the inflammatory stimulus increases, the suppressive capacity of human Tregs declines and the resistance of T-effector cells to regulation increases [63]. The situation observed during an acute, clinical malaria infection is thus in clear contrast to the situation in healthy, malaria-exposed individuals where Treg numbers closely track numbers of T-effectors, precisely maintaining an apparently optimal T-effector∶Treg ratio [36].
IL-10 is well-established as a vital homeostatic regulator of malaria-induced inflammation that prevents immune-pathology in mice [49],[64], promotes the necessary switch from early Th1 to subsequent Th2 responses [65],[66], and has been linked to protection from severe malaria anaemia [24],[25], and death [23] in humans. However, the cellular source of IL-10 in human malaria cases was, until now, ill defined. Contrary to our expectations, but in striking agreement with observations in P. yoelii-infected mice [33], CD45RO+ CD4+ T cells (that are CD25− and FOXP3−) and not classical Tregs are the only substantial source of IL-10 during acute malaria infection. This observation is reminiscent of that made by Nylen [67] in patients with acute visceral leishmaniasis. Moreover, in our patients, a significant proportion of IL-10 producing CD4+ T cells were simultaneously producing IFN-γ, identifying them as Th1 cells. Although IL-10 secreting Th1 cells have been described recently in two murine models of toxoplasmosis [50], and cutaneous leishmaniasis [51], as far as we are aware, this is the first demonstration of IL-10 producing Th1 cells during human infections. Intriguingly, the proportion of these cells within the total CD4+ T cell population was significantly higher in children with uncomplicated malaria than in children with severe malaria suggesting that in human P. falciparum infection, as in murine T.gondii infections [50], IL-10 producing Th1 cells, activated by a strong inflammatory stimulus, may act as anti-parasitic effector cells with a “built in” control mechanism to prevent the onset of immune pathology. If so, then the ability of these self-regulating effector cells to localize to sites of parasite sequestration in tissues, where they mediate parasite killing whilst simultaneously blocking tissue damage, may be key to clinical immunity to malaria. Thus, our data strongly suggest that the percentage of IL-10-producing Th1 effector cells, rather than the cocktail of circulating cytokines, may be the most relevant biomarker of effective immunity to severe malaria.
Although Tregs may not seem to determine the outcome of current P. falciparum infections we did find evidence that they affect the magnitude of the malaria specific memory response induced by the current infection. A similar observation has been made in P. berghei ANKA-infected mice; animals that were depleted of CD25+ cells prior to infection and drug-cured on day 5 developed significantly stronger IFN-γ memory responses on day 14 than did intact infected/cured mice, and these mice also developed much more severe, and frequently fatal, clinical symptoms upon reinfection, despite more efficient parasite clearance [35]. Thus, malaria specific Tregs acquired during a primary infection may limit the magnitude of Th1 effector responses to subsequent infections to a level that allows parasite clearance without causing immunopathology. Future studies should be designed to test the hypothesis that Tregs may contribute to the very rapid development of resistance to severe malaria.
In summary, our data indicate that classical FOXP3+ Tregs are unable to control the florid inflammation that accompanies acute malaria infections and this component of the immunoregulatory arsenal is rapidly overwhelmed in children with either mild or severe malaria. Importantly however we have identified, for the first time in an acute human infection, a population of IL-10 producing Th1 effector cells which appear to be a major source of this key anti-inflammatory cytokine during acute malaria infection, and which are associated with development of uncomplicated as opposed to severe malaria. We propose that IL-10-producing Th1 cells may be the essential regulators of acute infection-induced inflammation and that such “self-regulating” Th1 cells may be essential for the infection to be cleared without inducing immune-mediated pathology. Moreover, we have found evidence in support of the hypothesis that Tregs limit the magnitude of the Th1 memory response raising the intriguing possibility that they may play an important role in the rapid evolution of clinical immunity to severe malaria.
A case-control study was conducted in Gambian children with severe or uncomplicated malaria, resident in a peri-urban area within a 40 km radius south of the capital, Banjul, with low levels of malaria transmission [68],[69]. Patients were enrolled at Brikama Health Centre, the MRC Fajara Gate Clinic or the Jammeh Foundation for Peace Hospital in Serekunda between September 2007 and January 2008, after written informed consent was obtained from the parents or guardians. Uncomplicated disease was defined as an episode of fever (temperature >37.5°C) within the last 48 hours with more than 5000 parasites/µl detected by slide microscopy. Severe disease was defined using modified WHO criteria [70]: SA, defined as Hb<6 g/dl; SRD defined as serum lactate >7 mmol/L; CM defined as a Blantyre coma score ≤2 in the absence of hypoglycaemia, with the coma lasting at least for 2 hours. To avoid the confounding effects of other pathogens in children with concomitant systemic bacterial infections [71], children with clinical and/or laboratory evidence of infections other than malaria were not enrolled into the study. For some experiments, healthy children of the same age and recruited from the same area at the same time of the year were enrolled as controls. In total, 59 severe, 65 uncomplicated and 20 control cases were enrolled.
On admission (D0) and after 4 weeks (D28±3 days) one ml of blood was collected in RNA stabilizing agent (PAXgene™ Blood RNA system, Pre-AnalytiX) and a maximum of 4 mls of blood (mean: 3.2 mls CI 95%: 3.1–3.3 mls) were collected into heparinized vacutainers® (BD). All patients received standard care according to the Gambian Government Treatment Guidelines, provided by the health centre staff. The children's health was reviewed 7 days after admission. The study was reviewed and approved by the Joint Gambian Government/MRC Ethics Committee and the Ethics Committee of the London School of Hygiene & Tropical Medicine (London, UK).
P. falciparum parasites were identified by slide microscopy of 50 high power fields of a thick film. Full differential blood counts were obtained on days 0 and 28 using a Medonic™ instrument (Clinical Diagnostics Solutions, Inc); the presence of intestinal helminths was assessed by microscopy from stool samples collected into BioSepar ParasiTrap® diagnosis system, following the manufacturers' instructions. Sickle cell status was determined by metabisulfite test and confirmed on cellulose acetate electrophoresis [72].
Blood samples were processed within 2 hours of collection. Plasma was removed, stored at −80°C and replaced by an equal volume of RPMI 1640 (Sigma-Aldrich). PBMC were isolated after density centrifugation over a 1.077 Nycoprep (Nycomed, Sweden) gradient (800 g, 30 min) and washed twice in RPM 1640. Cells were either stained for flow cytometry directly ex-vivo, or cultured in RPMI 1640 containing 10% human AB+ serum, 100 µg/ml streptomycin, 100 U/ml penicillin (all Sigma-Aldrich), and 2 mM L-glutamine (Invitrogen Life Technologies), referred to as complete growth medium (GM).
Fresh PBMC were stained using the following fluorochrome labeled mouse or rat anti-human antibodies: FITC anti-TNF-receptor II (R&D), PE anti-FOXP3 (clone PCH101), Pacific Blue anti CD3, APC-Alexa Fluor 750 anti-CD127 (all Ebioscience), APC anti-CD25, PerCP anti CD4 (BD systems), ECD anti-CD45R0 (Beckman-Coulter), and appropriate isotype controls.
IL-10 and IFN-γ production by PBMC from 30 children with acute P. falciparum (D0) was assessed after 5 hours stimulation in GM containing PMA (50 ng/ml) and Ionomycin (1000 ng/ml) or GM alone.
Cells were stained with FITC anti-IFN-γ, PE anti-FOXP3, PE-Cy7 anti-CD25, APC-AF750 anti-CD8, Pacific Blue anti-CD3 (all Ebioscience), PerCP anti-CD4, APC anti-IL-10 (both BD), and ECD anti-CD45R0 (Beckman-Coulter). To ascertain specificity of the intracellular cytokine staining, aliquots of some samples were incubated with saturating amounts of purified non-labelled antibody of the same clone prior to staining with the fluorochrome labeled ICS antibody. The FOXP3 staining buffer set (Ebioscience) was used following the manufacturer's protocol. Samples were acquired on a 3 laser/9 channel CyAn™ ADP flowcytometer using Summit 4.3 software (Dako). Analysis was performed using FlowJo (Tree Star Inc.). All flowcytometric analysis was performed at the MRC laboratories, The Gambia on freshly isolated cells.
Plasma concentrations of IFN-γ, TNF-α and IL-10 were determined for each subject and time point on the Bio-Plex® 200 system, using X-Plex™ assays (both Bio-Rad Laboratories), according to the manufacturer's instructions. Data were analysed using the Bio-Plex® Manager software. The detection limit was defined as the concentration corresponding to a fluorescence value above the mean background fluorescence in control wells plus 3 SD, being 8.76 pg/ml for IFN-γ, 5 pg/ml for TNF-α and 0.57 pg/ml for IL-10. Values below this threshold were set to these levels.
P. falciparum parasites (3D7 strain) were cultured in vitro as described [73] and were routinely shown to be mycoplasma free by PCR (Bio Whittaker). Schizont-infected erythrocytes were harvested from synchronized cultures by centrifugation through a Percoll gradient (Sigma-Aldrich). PfSE was prepared by two rapid freeze-thaw cycles in liquid nitrogen and a 37°C water bath. Extracts of uninfected erythrocytes (uRBC) were prepared in the same way.
PBMC from 10 severe and 10 uncomplicated malaria cases collected on day 0 were depleted of CD25hi cells or mock depleted using magnetic beads (Dynal Biotech, UK), at a bead to PBMC ratio of 7∶1, and cell proliferation was determined by [3H]-thymidine (Amersham, UK) incorporation after 6 days in culture with PfSE, uRBC (RBC∶PBMC ratio equivalent to 2∶1), GM, or 2 days culture with PMA (10 ng/ml)+ Ionomycin (100 ng/ml), as described [27].
Cultured ELISPOTs were performed to assess malaria specific IFN-γ memory responses, adapting an established method [74]. Up to 1 million PBMCs collected on day 28 were cultured in 24 well plates for 6 days in 1 ml GM and stimulated with either PfSE, uRBC (RBC∶PBMC ratio equivalent to 2∶1), or GM respectively. At day 3, half the medium was exchanged and rIL-2 (final concentration 20 IU/well) was added. On day 6 cells were harvested, washed three times, and 1.5×105 cells seeded into duplicate wells onto Millipore MAIP S45 plates and restimulated overnight with PfSE, uRBC (concentrations as above), GM or PHA-L (5 µg/ml). IFN-γ ELISpot was performed using MabTech antibodies according to the manufacturer's instructions. Spot forming cell numbers were counted using an ELISPOT plate reader (AutoImmuneDiagnostica, Vers. 3.2). Results are expressed as spot forming units (SFU) per million PBMC after subtraction of individual background values (GM for PHA-L, uRBC for PfSE) being deducted. Assays were discounted if the positive control (PHA-L) was <50 SFU, or the negative control was >30 SFU.
For quantitative reverse transcription-polymerase chain reaction (RT-PCR), total RNA was extracted from PAX tubes following the manufacturer's instructions and reverse transcribed into cDNA using TaqMan® reagents for reverse transcription (Applied Biosystems), following the manufacturer's protocol. Gene expression profiles for FOXP3, IL-10, SOCS-2 and IFN-γ were measured by RT-PCR on a DNA Engine Opticon® (MJ Research) with QuantiTect SYBR Green PCR kits (Qiagen Ltd) using primers (all Sigma Genosys) previously described: IFN-γ, IL-10; FOXP3 designed by [75], and SOCS-2 designed by [76].
T-BET and GATA-3 gene expression was determined using the TaqMan® Probe kit using the primers (all Metabion) designed by [77]. 18S rRNA, amplified using a commercially available kit (rRNA primers and VIC labeled probe, Applied Biosystems), was used as an internal control. Data were analysed using Opticon Monitor 3™ analysis software (BioRad) and are expressed as the ratio of the transcript number of the gene of interest over the endogenous control, 18S rRNA.
Genomic DNA from each parasite isolate was genotyped by sequencing the highly polymorphic block 2 region of the msp1 gene to assess the number of clones infecting each patient [78].
Analysis was performed using linear regression, with a random effect to allow for the within subject measurements over time, where the response variables were log transformed to improve the normality and constant variance assumptions. Significance (measured at the 5% level) tests for the effects of malaria group (uncomplicated, SA or SB), time (day 0 and day 28) and their interaction were adjusted for the possible confounding effects of age, sex, duration of prior symptoms and numbers of clones causing the infection. Where there was no significant malaria group and time interaction, p-values for the overall comparison of day 0 vs. day 28 are given. Within day 0, comparisons of severe vs. uncomplicated and the two groups of severely ill patients (SA vs SB) were adjusted for any malaria group and time interactions. To allow for the multiplicity of tests resulting from multiple responses and multiple comparisons within a response performed in the model, a false discovery rate (FDR) of 5% was assumed. Using the Benjamini and Hochberg approach [43] only tests with a p-value below 0.012 have an FDR of ≤5%. Due to the large number of tests family-wise error rate correction methods were too conservative. Analyses were performed using Stata version 9 and Matlab version R2008a.
|
10.1371/journal.pntd.0000269 | Intestinal Transcriptomes of Nematodes: Comparison of the Parasites Ascaris suum and Haemonchus contortus with the Free-living Caenorhabditis elegans | The nematode intestine is a major organ responsible for nutrient digestion and absorption; it is also involved in many other processes, such as reproduction, innate immunity, stress responses, and aging. The importance of the intestine as a target for the control of parasitic nematodes has been demonstrated. However, the lack of detailed knowledge on the molecular and cellular functions of the intestine and the level of its conservation across nematodes has impeded breakthroughs in this application.
As part of an extensive effort to investigate various transcribed genomes from Ascaris suum and Haemonchus contortus, we generated a large collection of intestinal sequences from parasitic nematodes by identifying 3,121 A. suum and 1,755 H. contortus genes expressed in the adult intestine through the generation of expressed sequence tags. Cross-species comparisons to the intestine of the free-living C. elegans revealed substantial diversification in the adult intestinal transcriptomes among these species, suggesting lineage- or species-specific adaptations during nematode evolution. In contrast, significant conservation of the intestinal gene repertories was also evident, despite the evolutionary distance of ∼350 million years separating them. A group of 241 intestinal protein families (IntFam-241), each containing members from all three species, was identified based on sequence similarities. These conserved proteins accounted for ∼20% of the sampled intestinal transcriptomes from the three nematodes and are proposed to represent conserved core functions in the nematode intestine. Functional characterizations of the IntFam-241 suggested important roles in molecular functions such as protein kinases and proteases, and biological pathways of carbohydrate metabolism, energy metabolism, and translation. Conservation in the core protein families was further explored by extrapolating observable RNA interference phenotypes in C. elegans to their parasitic counterparts.
Our study has provided novel insights into the nematode intestine and lays foundations for further comparative studies on biology, parasitism, and evolution within the phylum Nematoda.
| Biological properties of the nematode intestine warrant in-depth investigation, the results of which can be utilized in the control of parasitic nematodes that infect humans, livestock, and plants. Both the importance of intestinal antigens from Haemonchus contortus in immunity and the damage to H. contortus intestine by anthelmintic fenbendazole have highlighted the versatility of the intestine as an emerging target. However, biological information regarding fundamental intestinal cell functions and mechanisms is currently limited. Conserved intestinal genes across nematode pathogens could offer molecular targets for broad parasite control. Furthermore, qualitative and quantitative comparisons on intestinal gene expression among species and lineages can identify basic adaptations relative to a critical selective force, the nutrient acquisition. This study begins to identify intestinal cell characteristics that are conserved across representatives of two clades of nematodes (V and III) and further clarifies diversities that likely reflect species- or lineage-specific adaptations. Results consistent with functional data on digestive enzymes from H. contortus and RNAi in Caenorhabditis elegans, as examples, support the potential for the comparative genomics approach to produce practical applications. This study provides a platform on which extensive investigation of intestinal genes and a more comprehensive understanding of the Nematoda can be gained.
| The intestine is one of the major organs in nematodes, creating a key surface at the intestinal apical membrane that interacts with the environment. While specific cellular characteristics of the intestine can be diverse among nematode species, they typically conform to polarized epithelial cells with the apical membrane composed of microvilli lining the digestive tube. In apparent contrast to other surfaces of nematodes, digestive and assimilative functions, as well as various metabolic pathways and cellular trafficking, are expected to be extremely active at the intestinal surface. For example, an adult Caenorhabditis elegans is capable of producing oocytes with about the same total biomass as its own body per day [1], but the average intestinal residence time for foods was estimated to be less than two minutes in C. elegans [2], suggesting that the microvillous membrane must have an enormous capacity for nutrient digestion and absorption. In addition, the intestine has to offer innate immunity against invasive pathogens, and adaptations at the apical intestinal membrane may be required to protect parasitic nematodes against host immune systems. Furthermore, the nematode intestine has been suggested to be involved in other biological processes such as stress responses, body size control, and aging [1].
Three lines of evidence indicate that the intestine is an important target for the control of parasitic nematodes. First, intestinal antigens enriched for apical membrane-associated proteins have been successfully used to immunize against Haemonchus contortus, a hematophagous nematode of small ruminants [3]–[7]. Surface-bound nematode proteases are a dominant, but not exclusive, group of proteins that have been implicated in inducing this protection. A prospective mechanism of the immunity involves perturbing nutrient digestion and acquisition at the intestinal surface by the ingested host-derived antibodies capable of neutralizing parasite digestive proteases [7]. Further investigations conducted with hematophagous hookworms also produced similar effects [8]. Second, adult H. contortus intestinal cells are hypersensitive to benzimidazole anthelmintics, apparently through the target protein beta-tubulin isotype 1 [9],[10]. It was suggested that the drug inhibited vesicle transport in the apical secretory pathway, causing the intracellular release of the digestive enzymes destined for secretion and subsequent cytotoxic effects [9]. Third, parasite control has been demonstrated by inhibition of an intestinal enzyme, cathepsin L cysteine protease, by either RNA interference or a chemical inhibitor in the plant parasitic nematode Meloidogyne incognita [11]. These observations generate great interests to uncover the basic characteristics of the intestinal cells that might be further exploited for the broad control of parasitic nematodes. However, the dearth of relevant experimental systems and molecular information such as gene repertoires for many parasitic species has impeded rapid progress.
Five major clades (I–V) are currently recognized to comprise the phylum Nematoda [12],[13]. So far, almost all studies of the intestine at the gene level have focused on the clade V nematodes. A small-scale sampling of expressed sequence tags (ESTs) from the dissected intestine from adult H. contortus females identified 51 intestinal genes including cysteine proteases [14], this list was later expanded via a proteomic approach to include a number of apical intestinal membrane proteases from H. contortus and hookworms [15]. Intestinal EST libraries generated from laser-dissected materials from Necator americanus and Ancylostoma caninum allowed the identification of 544 intestine-expressed genes [16]. Although a more comprehensive dataset with >5,000 intestinal genes is available in C. elegans [17]–[19], it is unclear, given the evolutionary diversity within Nematoda, to what extent the molecular and cellular functions of the intestine can be extrapolated across nematode species.
In this study, we sampled the transcribed genomes from several tissues and developmental stages from two parasitic nematodes: the clade III nematode Ascaris suum, which presumably feeds on the semi-digested contents in the host intestine, and the clade V blood-feeding parasite H. contortus. Nearly 10,000 and 5,000 genes were identified from the two nematodes, respectively. More importantly, given the attention to the intestine, we produced the largest collection of intestinal genes in parasitic nematodes by dissecting adult intestine from each species, a procedure that is not practical for many other nematodes because of their small sizes and the lack of laboratory culturing systems. Extensive cross-species comparisons were made among the adult intestinal genes from the parasites and those expressed in the adult intestine of the free-living bacterivore C. elegans. Both diversification and conservation of intestinal gene repertories were evident among the species investigated. The diversities of intestinal transcriptomes by clade and species may reflect the substantial life style differences among these nematodes. A group of 241 protein families were found conserved in the intestine of all three nematodes, accounting for ∼20% of the intestinal gene repertoires from the three species. These genes may include core intestinal functions that are indispensable among many nematodes. Functional annotations were generated for the intestinal genes. Molecular characteristics of the intestinal genes were further explored to highlight various physiological aspects of the nematode intestine.
Dissection of the adult intestine was carefully performed under microscopy as described previously [14],[20],[21]. The samples used in this study had also passed another round of visual inspection microscopically to ensure they did not contain other tissues such as muscle, esophagus, or hypodermis. Detailed information on genetic materials and cDNA library construction are available at www.nematode.net. ESTs were processed and clustered as described before [22]–[25]. EST contig sequences were translated individually by Prot4EST, a 6-tier translation pipeline combining both similarity-based methods and de novo predictions [26], for downstream analysis.
Databases used for sequence comparisons were: i) Caenorhabditis spp., all amino acid sequences in the complete genomes of C. elegans (Wormbase Release v150), C. briggsae (June, 2006), and C. remanei (June, 2006), ii) Other Nematoda, all non-Caenorhabditis nematode nucleic acid sequences in GenBank excluding those from A. suum (when analyzing A. suum sequences) or H. contortus (when querying H. contortus sequences) (October 18, 2006), and iii) Non-Nematoda, all amino acid sequences in the non-redundant protein database NR excluding those from nematode species (September 20, 2006). WU-BLASTP (wordmask = seg postwe B = 1000 topcomboN = 1) was used to query the translated sequences against protein databases, and WU-TBLASTN (wordmask = seg lcmask B = 1000 topcomboN = 1) for searching against nucleotide databases [27]. The E-value cutoff of 1.0e−5 was used to accept sequence similarities in all BLAST searches.
Each intestinal EST cluster was assigned two counts according to the numbers of times it was sampled from either the intestinal or non-intestinal cDNA libraries, respectively. Similarly, each C. elegans intestinal gene was assigned two counts for the numbers of times it was sampled by SAGE tags from either the glp-4 dissected gut or the glp-4 adult whole worm, respectively. The mutants lack the gonad when raised at 25°C, therefore contamination by other tissues is less likely [17]. The SAGE data was downloaded with sequence quality filter = 0.99, no normalization, duplicate ditags and ambiguous or antisense tags removed (April 19, 2006; mapped to Wormbase Release v150) [17]. A Poisson-based enrichment test, considering both the total sampling sizes and random variations [28], was implemented to compute an P-value to represent the likelihood of intestinal enrichment for each EST cluster or C. elegans gene using these two counts. The P-value cutoff of 0.001 was chosen to define the putative intestine-enriched genes from the three nematodes.
A hidden Markov modeling-based algorithm, Phobius [29], was used with default setting. Each query sequence was further annotated as TM-only, TM with SP, SP-only, or intracellular based on raw Phobius outputs. For each EST cluster, Phobius annotation was predicted for each contig and summarized at the EST cluster level. A modified Wormbase Release v150 containing only the longest splicing isoform at each gene loci was used as the complete gene set of the C. elegans genome.
For tissue-level comparisons made between intestine and gonad, InParanoid [30] was used at default settings to identify a total of 1,764 putative orthologous groups between all the A. suum EST clusters and the complete gene set of C. elegans (the modified Wormbase Release v150 containing only the longest splicing isoform at each gene loci). InParanoid-generated main orthologous pairs, which are essentially the mutual-best matches between all the available genes from the two species, were further screened against the 447,546 A. suum Genome Survey Sequences (GSSs) that were generated recently (Mitreva, unpublished), resulting in the final group of 1,652 putative main orthologous pairs in which the C. elegans members do not have better matches in GSSs than the A. suum EST partners assigned by InParanoid. C. elegans gonad-expressed genes were extracted from SAGE data generated from dissected gonad (March 12, 2007) [17].
An all-against-all WU-BLASTP was performed on all the 9,918 translated intestinal genes from the three species (including sequences for EST contigs from the two parasites and 5,056 C. elegans genes). Raw BLAST results were fed to a C-language implementation of Markov Cluster (MCL) Algorithm (www.micans.org/mcl), a fast and scalable unsupervised cluster algorithm based on simulation of flow in graphs [31]. An Inflation Fact of 1.6 was chosen for the MCL clustering. The MCL output was then summarized at the EST cluster level, during which we applied an additional filtering step to remove an EST cluster from a MCL protein family if less than 10% of its total contigs were clustered into that family. These parameters were based on manual inspection of the results on a test set consisting of the putative intestine-enriched genes with 210 parasite EST clusters and 247 C. elegans genes (false positive rate of 3%; data not shown).
Default parameters for InterProScan v13.1 [32] were used to search against the InterPro database [33]. Raw InterProScan results for the translated EST contigs were summarized at the EST cluster level. Gene ontology (GO) terms were further assigned and displayed graphically by the AmiGO browser with default parameters and the ontology data released on March 15, 2007[34]. Complete GO mappings for the three intestinal transcriptomes are available at www.nematode.net. For each GO term, its enrichment in an IntFam group (such as the IntFam-241 group) was measured over the complete set of 9,918 translated intestinal genes using a hypergeometric test, the p-value cutoff of 1.0e−5 was chosen for enrichment. The less informative ontologies, including those at level 4 or higher for Biological Process or Molecular Function, and those at level 2 or higher for Cellular Component, were removed from the enrichment list. Also removed were redundant ontologies by keeping only the lower level more informative ontology if the same group of genes was mapped to more than one GO term.
An empirical mixed approach was used for mapping the novel genes to canonical pathways. The E-value cut-off of 1.0e−10 reported by WU-BLASTP against the Genes Database Release 39.0 from Kyoto Encyclopedia of Genes and Genomes (KEGG) was first used for finding homologous matches. Then the top match and all the matches within a range of 30% of the top BLAST score, if meeting the cut-off, were accepted for valid KEGG associations [35]–[37]. A hypergeometric test, measuring the relative coverage of the KEGG-annotated orthologous groups assigned to a pathway, was implemented to identify the enriched pathways for each intestine [38].
Nucleotide sequences data reported in this paper are available in the GenBank, EMBL and DDBJ databases. The accession numbers for ESTs from A. suum are: BI781215-BI784439, BM032617-BM034650, BM280443-BM285290, BM318846-BM319958, BM515079-BM518821, BM566483-BM567588, BM568416-BM569529, BM732977-BM734435, BM964439-BM965448, BQ094886-BQ096565, BQ380669-BQ383404, BQ835081-BQ835723, BU965907-BU966430, CA849193-CA850481, CA953713-CA955182, CB100077-CB102042, DV018957-DV019894, EB186562-EB187079. The accession numbers for ESTs from H. contortus are: CA033335-CA034379, CA868595-CA870175, CA956361-CA959150, CB018493-CB022024, CB063882-CB065260, CB099467-CB100076, CB190871-CB192419, CB331948-CB333475.
We constructed 18 A. suum and 6 H. contortus stage- or tissue-specific cDNA libraries, and sequenced 31,416 and 14,014 5-prime ESTs from the two species, respectively. These ESTs totaled to 13.6 and 6.3 million bases for A. suum and H. contortus, accounting for 77% and 63% of the total nucleotides from the two species currently available in public databases (Table S1). Supplemented by 9,354 A. suum and 8,146 H. contortus ESTs previously deposited in GenBank (retrieved in January, 2006), all available ESTs were grouped into 17,989 A. suum and 9,842 H. contortus EST contigs, each containing ESTs derived from nearly identical transcripts according to overlapping sequences to reduce sequence redundancy [23],[24]. The contigs were further assembled into 9,947 A. suum and 5,058 H. contortus EST clusters based on sequence similarities identified among contigs as well as in previously identified genes (Table S1). Each EST cluster likely represents transcripts derived from a single genomic locus and therefore is approximated as one gene [22]–[24]. Given that C. elegans and C. briggsae each contains ∼19,000 protein-coding loci, and between 14,500 and 17,800 genes were inferred from the Brugia malayi draft genome [39], we have consequently identified a substantial portion of the complete gene sets from the two parasites. These data will vastly facilitate the genome assembly and annotation in the related nematode genome sequencing projects currently underway. Initial investigation of the identities of these novel genes was performed by comparing the translated sequences with known proteins from other organisms (Text S1; Figure S1).
To study the intestinal transcriptomes, four cDNA libraries (out of the 18) from A. suum and three (out of the 6) from H. contortus were constructed from dissected adult intestine with methods based on either Poly-A [40] or spliced leader sequences [24]. Among all the ESTs we generated, a total of 9,586 A. suum and 7,068 H. contortus ESTs were derived from these intestinal libraries. These ESTs occurred in 3,121 A. suum and 1,755 H. contortus EST clusters, accounting for about 30% of the total genes sampled in each nematode. Since these EST clusters contained ESTs sampled from the adult intestine, they were considered to represent adult intestinal genes, making this the largest tissue-level gene discovery in parasitic nematodes thus far (Table 1).
In contrast to the two gastrointestinal parasites, the free-living model nematode C. elegans is a bacterivore obtaining nutrients primarily or exclusively from the consumption of bacteria. Two previous studies reported identification of genes expressed in the adult C. elegans intestine: i) sequence tags generated by serial analysis of gene expression (SAGE) from the dissected adult intestine were mapped to over 4,000 C. elegans genes [17],[18]; ii) a study using mRNA tagging and microarray gene expression profiling identified ∼1,900 intestine-expressed genes [19]. Consolidating the two efforts provided us with a non-redundant set of 5,065 intestinal genes from adult C. elegans, covering over 25% of all coding loci in its entire genome (Table 1).
The phylum Nematoda is ancient and diverse. Even though the evolutionary distance between clade III A. suum and clade V C. elegans was estimated to be ∼350 million years [41], the nematode intestine has maintained high similarity in both tissue morphology and presumably physiology (i.e. involvement in feeding). However, it is unknown how much the intestine is conserved, or diversified, at the molecular level across species. The tissue-level gene sampling in this study offered an opportunity to investigate this question.
Differences in the intestinal gene repertoires were obvious among the three nematodes. In total, 39% of A. suum and 19% of H. contortus intestinal genes were found to be novel compared to all known proteins in the public databases (Figure 1). Such novel intestine-expressed parasite genes contained no match in the complete genome of the free-living C. elegans, thus not in the C. elegans intestine, making them unique by comparison to C. elegans. In addition, for the sampled intestinal genes from both parasites, the non-Caenorhabditis nematodes offered the largest numbers of homologous matches than either the Caenorhabditis species or the non-nematode organisms (Figure 1). Such differences may suggest the existence of lineage- or species-specific diversification in the nematode intestine.
Furthermore, we observed higher levels of diversification in the putative intestine-enriched genes from the three nematodes. Taking into consideration sample size and random sampling fluctuation [28], we identified 150 A. suum, 60 H. contortus, and 247 C. elegans putative intestine-enriched genes based on the “digital” expression levels revealed in EST and SAGE data (at the Poisson distribution-based P-value cutoff of 0.001) (Table S2; Table S3; Table S4). Many of these predicted enrichments suggested unique intestinal functions for the individual species. For example, the group of 60 genes from the blood-feeding H. contortus includes 2 fibrinogen-related proteins that may function as thrombin inhibitors to prevent clotting of ingested blood. Also included are putative enzymes that may be involved in the digestion of hemoglobin, one of the major food sources of blood-feeding parasites, including a serine-type protease, a metallopeptidase, and 13 different cysteine-type proteases that were reported previously [42] (Table S3). Interestingly, a significantly higher percentages of these genes (e.g. 15%-31% higher than all the sampled intestinal genes) encode proteins predicted as secreted or trans-membrane [29] (Figure 2), suggesting that they interact with the extracellular environment. However, 64%, 54%, and 69% of them, from the three species respectively, were distinct from members of the protein families conserved in the intestine of all three nematodes (IntFam-241; see below), indicating that a large portion of these putative intestine-enriched genes are specific to the intestine of individual nematode lineages or species. This further underlines the diversification of intestinal transcriptomes in accommodating the different life styles and feeding patterns among nematodes.
To evaluate common characteristics of the nematode intestine, we first sought evidence for the molecular conservation of the tissue in the context of phylogeny. We made comparisons among genes expressed in the intestine of A. suum and C. elegans and those expressed in another tissue, namely the gonad. These two species have the largest numbers of sequences available, and they also represent the most distant relationship among the three nematodes investigated. The gonad was chosen because the next largest group of genes was sampled from this tissue in A. suum after the intestine. H. contortus was excluded from this analysis because a gonad-expressed gene set was not available from this nematode. Genes expressed in the intestine and gonad were divided into four putative tissue-specific groups: i) 2,453 A. suum and ii) 2,557 C. elegans genes expressed in the intestine but not in the gonad (the two intestine groups), and iii) 2,690 A. suum and iv) 2,589 C. elegans genes that were found in the gonad but not in the intestine (the two gonad groups). The use of the similar numbers of genes in each group is expected to reduce false results caused by over-representation from any single category.
Molecular conservation was first evaluated by comparing the numbers of putative homologous pairs identified among the intestine and gonad gene groups. The number of the putative homologs between the two intestine groups was significantly larger than that between the intestine and gonad groups (p-value = 2.5e−04 at the bit-score cutoff of 100 in a permutation two-tailed Z-test; p-value = 4.2e−08 at the bit-score cutoff of 50; Figure S2), suggesting that for genes expressed in the intestine of one nematode, their homologous matches in another species are significantly more likely to be expressed in the intestine than in the gonad of the second nematode. These results provide evidence for the molecular conservation of the intestine across these distantly related nematodes. In contrast, the number of putative homologs between the two gonad groups was not statistically different from that between the gonad and intestine (Figure S2), indicating that the gonad genes appeared to be less conserved than those expressed in the intestine in this two-tissue comparison.
To increase the confidence of analysis, we next focused on the putative orthologous pairs predicted among the intestine and gonad gene groups, which was a smaller data set than the homologous pairs used above but with higher stringency. Among the total of 1,652 putative orthologous pairs predicted from A. suum and C. elegans (see Materials and Methods), 289 were paired among genes from the intestine and gonad groups. They were used in a Chi Square statistical test, with random distribution of orthologous pairs as the null hypothesis. Compared to the expected numbers, there was a 31% enrichment of orthologous pairs observed between the A. suum and C. elegans intestine groups (Figure 3), whereas the enrichment between the two gonad groups was only marginal (5%), and the observed numbers of orthologous pairs between the gonad and intestine groups were less than expected (Figure 3). Overall, a significant χ2 value of 11.9 rejects the null hypothesis at a confidence level higher than 99% (p-value <0.01) [43], and selective pressure is evident on molecular conservation of the intestinal gene repertories.
Although the use of the incomplete transcriptomes and a bias towards relatively abundant transcripts in EST sampling can affect results, analyses of either homologous or orthologous pairs both provide direct support for the molecular conservation of the nematode intestine. With the obvious pattern of diversification in the nematode intestine (discussed earlier), our results indicate that a subset of the intestinal gene repertoires, which likely contribute to the intestinal characteristics conserved across diverse nematode species, remain conserved during the evolution of Nematoda.
Interestingly, genes expressed in the gonad appear to be less well conserved based on both analyses. However, these results do not suggest the lack of evidence for the conservation of the gonad. Instead, the two-tissue comparisons indicate that the levels of conservation are lower in the gonad than in the intestine, suggesting that the levels of molecular conservation may differ in different nematode tissues. In fact, the conserved characteristics of the gonad may become more evident with larger sample sizes and/or by comparisons with another tissue with a lower level of conservation than the intestine, when new sequence data becomes available. Similarly, differences at the levels of molecular conservation were observed in different tissues between human and mouse, which diverged only about 25 million years ago [44]. Future comparisons with more complete expression data across multiple tissues in different nematode species should offer additional insights into this aspect of nematode evolution.
To compare the intestinal transcriptomes of A. suum, H. contortus, and C. elegans in a single analysis, we built protein families from the complete set of 9,918 translated intestinal genes combined from the three nematodes. A total of 5,587 intestinal protein families (IntFam) were identified conservatively based on sequence similarities by MCL clustering [31] (Figure 4). Proteins assigned into the same protein family contain putative homologous or orthologous matches among the three species. Both diversification and conservation of the intestinal transcriptomes was obvious at the protein family level in this 3-species comparison. A total of 59% of all the sampled intestinal genes were members of the protein families containing proteins from only one nematode (Figure 4). Although the assignments for many of these single-species families are likely to change when more complete intestinal gene repertories become available, this group includes the genes contributing to the unique intestinal features in each species. The remaining 41% of the intestinal genes formed 910 multi-species protein families; they are conserved in the intestine of at least two nematodes. Among these multi-species families, 241 had members from all three species, accounting for ∼20% of all the intestinal genes under investigation (Figure 4). Given the differences in life styles and feeding patterns among the three nematodes, we propose that these 241 intestinal protein families represent an ancestral intestinal transcriptome involved in core cellular and physiological intestinal functions common to the investigated species or even across the Nematoda. Therefore, we referred to them as the “core” IntFam-241 group.
The 9,918 translated intestinal genes sampled from the three nematodes were annotated and classified using Gene Ontology [34],[45]. Ontologies were assigned at a higher ratio (58%) to the C. elegans intestinal genes than to those from A. suum (31%) or H. contortus (35%; Table 1). In addition, genes in the multi-species IntFam groups, which contained members from at least two nematodes, were annotated at higher ratios (47%–74%), whereas only 8% of the genes were annotated from the two single-species IntFam groups containing members only from A. suum or H. contortus (data not shown). These data may indicate that novel intestinal genes have independently evolved in relation to the different lineages of parasitism. Complete GO mappings for the three intestinal transcriptomes are presented in the searchable AmiGO browser at www.nematode.net [46]. Furthermore, A hypergeometric test was implemented to identify ontologies that are statistically enriched, thus indicating enriched features, in the core IntFam-241 (Table 2) as well as other IntFam groups (Text S1; Table S6).
Five of the 17 enriched Molecular Function ontologies in IntFam-241 are related to protein kinases (Table 2; Table 3). Protein kinases are one of the largest and most influential protein families, accounting for about 2% of genes in a variety of eukaryotic genomes including C. elegans and B. malayi. They regulate almost every aspect of cellular activities and may phosphorylate up to 30% of entire proteomes [39],[47]. Based on GO annotations, protein kinases were enriched by ∼3.5 fold in IntFam-241 over the complete set of intestinal genes (5.3% vs. 1.5% of the total genes for each group). Both serine/threonine- and tyrosine-types of protein kinases were found to be enriched. Novel protein kinases from the parasites were further classified based on their C. elegans homologs (Table S5). Interestingly, molecular functions such as adenyl nucleotide binding, ATP binding, and GTP binding were also enriched. The involvement of these functions in protein kinase activities further suggested key roles of cellular signaling in the nematode intestine (Table 2).
The other major Molecular Function terms enriched in IntFam-241 were the proteases (Table 2; Table 3). All but one of the six subtypes of proteases (glutamic acid-type proteases as the exception) had been identified in IntFam-241 (Table 3), suggesting conservation of essential protease functions, such as nutrient digestion and acquisition, among the three species or even across many species of Nematoda. Because each species feeds on distinct food sources, it is possible that related digestive proteases have evolved within each species to adapt for digestion of the different food types. Given the success of parasite control achieved by immunization with H. contortus and hookworm intestinal protease-type antigens [3],[8], these proteases may warrant further investigation in A. suum and other parasites.
Analysis of the IntFam groups other than IntFam-241 was also conducted. However, in absence of deeper sampling of the intestinal transcriptomes, it is difficult to interpret the results in relation to broadly conserved or lineage- and species-specific characteristics (Text S1; Table S6).
To identify the biological pathways that are active in the nematode intestine, we mapped the 9,918 intestinal translated sequences, and for comparison, the complete C. elegans genes (Wormbase Release v150), to the reference canonical pathways in Kyoto Encyclopedia of Genes and Genomes [35]–[37] (Table 4). Complete listing of all KEGG mappings including graphical representation is available for navigation at www.nematode.net [46].
The enrichment of specific major KEGG pathways was evident for each intestine by comparisons to the complete KEGG mappings for all C. elegans genes (Table 4) [38]. Carbohydrate metabolism, energy metabolism, and translation were identified as the statistically enriched pathways in all three intestinal transcriptomes (at the p-value cutoff of 0.05). Interestingly, immune system was an enriched KEGG cellular process in the A. suum intestine; this pathway barely missed the cutoff for enrichment in H. contortus (with a p-value of 9.9e−02), but no enrichment was indicated for the C. elegance intestine (Table 4). The KEGG immune system was built based on studies in mammalian systems. Many of those from the two parasites were mapped to intracellular proteins of immune cells involved in, for example, intracellular signaling or antigen processing (Table S7; Table S8). Therefore, the potential for their involvement in interactions with the host are not a primary suggestion here, but it cannot be completely excluded either.
RNA interference (RNAi) has been developed and successfully applied to genome-wide gene silencing to inhibit gene functions in C. elegans [48]–[51]. C. elegans RNAi information can be further extrapolated in understanding functions of orthologous genes in other nematodes, especially in parasitic nematodes where high-throughput screening is not yet practical [52]. For the 3,455 IntFam protein families containing C. elegans genes, observed RNAi phenotypes for their C. elegans members (Wormbase Release v150) were extracted and extrapolated to a total of 45% of these IntFams (Table 5). Protein families from the IntFam-241 were assigned at a higher ratio (73%) than those from other IntFam groups with C. elegans members (Table 5). Among the IntFams-241 families with RNAi phenotypes assigned, 74% (131/176) had severe phenotypes including embryonic, larval, or adult lethal, sterile, sterile progeny, and larval or adult growth arrest (data not shown). Since the IntFam-241 families represent proteins conserved in all the three species, these results further support our hypothesis that the core IntFam-241 protein families likely play key roles in the nematode intestine across many species.
We have performed large-scale sampling of the transcribed genomes in A. suum and H. contortus from various tissues or developmental stages, accounting for 77% and 63% of total available bases for the two nematodes, respectively. The identification of 9,947 A. suum and 5,058 H. contortus genes in this study will vastly facilitate the related genome sequencing projects currently underway. The research has produced the largest samplings of the adult intestinal transcriptomes thus far in parasitic nematodes by identifying 3,121 A. suum and 1,755 H. contortus intestinal genes, making possible the extensive comparative studies with the adult intestinal transcriptome of the free-living C. elegans. We found that, even with the evolutionary distance of an estimated 350 million years separating clades III and V nematodes [41], both significant conservation and diversification of gene repertories were evident for the intestine. A group of 241 intestinal protein families, each containing members from all three intestines, were further identified. The IntFam-241 group, containing ∼20% of all intestinal genes sampled from the three species, was proposed to represent an ancient intestinal transcriptome responsible for core cellular and physiological intestinal functions that are conserved in the investigated species or many other nematodes. In addition, various aspects of nematode intestinal physiology were revealed by GO and KEGG classifications of the intestinal transcriptomes, and the examination and extrapolation of available RNAi phenotypes from C. elegans. Overall, this study has contributed to a better understanding of nematode biology, providing central information for the development of novel and more effective parasite control strategies. Finally, the use of the C. elegans model to dissect basic parasite biology has been slow to evolve. Results presented here identified numerous specific areas of research where C. elegans might contribute in this way.
|
10.1371/journal.pgen.1000265 | A Motor Function for the DEAD-Box RNA Helicase, Gemin3, in Drosophila | The survival motor neuron (SMN) protein, the determining factor for spinal muscular atrophy (SMA), is complexed with a group of proteins in human cells. Gemin3 is the only RNA helicase in the SMN complex. Here, we report the identification of Drosophila melanogaster Gemin3 and investigate its function in vivo. Like in vertebrates, Gemin3 physically interacts with SMN in Drosophila. Loss of function of gemin3 results in lethality at larval and/or prepupal stages. Before they die, gemin3 mutant larvae exhibit declined mobility and expanded neuromuscular junctions. Expression of a dominant-negative transgene and knockdown of Gemin3 in mesoderm cause lethality. A less severe Gemin3 disruption in developing muscles leads to flightless adults and flight muscle degeneration. Our findings suggest that Drosophila Gemin3 is required for larval development and motor function.
| The childhood disease spinal muscular atrophy (SMA) has a drastic impact on motor neurons and muscles. The cause has been linked to a deficiency in the survival motor neuron (SMN) protein. SMN interacts with various proteins termed Gemins to form the SMN complex, among which Gemin3 is the only one with an RNA unwinding activity. Here, we study the function of D. melanogaster Gemin3 in the context of development. The association of Gemin3 with SMN, which had been reported previously in humans, is conserved in flies. Loss of Gemin3 resulted in death at larval stages. Before they die, gemin3 mutant flies become sluggish and develop large synapses, which are the contacts between motor neurons and muscles. Disruption of Gemin3 in mesodermal tissues, especially muscles, causes development defects, degeneration of flight muscles, and flies that are unable to fly. This study demonstrates that Gemin3 plays a critical role in fruit fly development, especially in motor function, which raises the question of whether disruption of Gemin3 contributes to SMA.
| Spinal muscular atrophy (SMA) is an autosomal recessive disorder characterised by degeneration of spinal cord motor neurons, as well as progressive muscular weakness, dysphagia, dyspnoea, and in severe cases, death [1],[2]. The majority of SMA patients harbour deletions or mutations in the survival motor neuron (SMN1) gene, which encodes an RNA-binding protein, SMN. In mammalian cells, the SMN protein is stably complexed with a group of proteins including Gemin2 [3], Gemin3 [4],[5], Gemin4 [6], Gemin5 [7], Gemin6 [8], Gemin7 [9], and Gemin8 [10]. Biochemical studies in vertebrate systems suggested that the SMN complex plays an essential role in small nuclear ribonucleoprotein (snRNP) assembly. The SMN complex binds directly to small nuclear RNAs (snRNAs) and ensures that a set of seven Sm or Sm-like (Lsm) proteins are assembled onto snRNAs [11].
Gemin3, the only RNA helicase in the SMN complex, contains nine conserved motifs including the Asp-Glu-Ala-Asp motif (or DEAD box in one-letter code). The RNA helicase activity of Gemin3 is ATP-dependent with a 5′ to 3′ direction [12]. RNAi-mediated knockdown studies indicated a role for Gemin3 in the assembly of snRNP complexes as an integral component of the macromolecular SMN complex [13],[14]. Furthermore, a recent study demonstrated that intracellular Gemin3 proteolysis by a poliovirus-encoded proteinase led to reduced Sm core assembly activity in poliovirus-infected cells [14].
In addition to snRNP biogenesis, Gemin3 was also implicated in transcriptional and microRNA regulation. Gemin3 was originally isolated as a cellular factor that associates with the Epstein-Barr virus nuclear proteins EBNA2 and EBNA3C, which play a role in the transcriptional regulation of both latent viral and cellular genes [15]. The non-conserved C-terminal domain of Gemin3 has been shown to interact with and modulate a variety of cellular transcription factors including steroidogenic factor 1 [12],[16], early growth response protein 2 [17], forkhead transcription factor FOXL2 [18], and mitogenic Ets repressor METS [19]. Although the majority of Gemin3 and its associated protein, Gemin4, are found in the SMN complex, a less abundant Gemin3-Gemin4 complex has been isolated from HeLa and neuronal cells. The Gemin3-Gemin4 complex contains Argonaute 2 and numerous microRNAs, co-sedimenting with polyribosomes [20]–[22].
Despite the detailed studies in vertebrate systems and a recent study in Drosophila culture cells [23], the function of Gemin3 in Drosophila development remains elusive. Here we identify the orthologue of Gemin3 in Drosophila melanogaster and demonstrate that Drosophila Gemin3, like its vertebrate counterpart, associates with SMN. Loss-of-function gemin3 mutants are lethal as third instar larvae and/or prepupae. Before they perish, gemin3 mutants exhibit dramatic loss of mobility and neuromuscular junction (NMJ) defects. Tissue-specific expression of a dominant-negative transgenic construct and RNAi studies suggest that the function of Gemin3 in mesoderm, particularly in muscles, is essential for animal survival. Furthermore, disruption of Gemin3 in muscles causes flight muscle degeneration and loss of flight. Thus our study demonstrates that Drosophila Gemin3 plays a critical role in development and motor function.
We carried BLAST searches of the Drosophila melanogaster genome using human and mouse Gemin3 sequences, and found that the DEAD/DEAH RNA helicase 1 (Dhh1) or CG6539 is the putative Drosophila Gemin3 orthologue. This gene, renamed for the present studies as gemin3, is located on the third chromosome in region 67E3, and is composed of 2 exons separated by a short intron. The Drosophila melanogaster Gemin3 protein is composed of 1028 amino acids and shows 33% identity and 55% similarity (BLASTP) to the respective human orthologue (Figure 1A, B). This level of conservation is quite similar to that observed between the Drosophila and human SMN, which have an overall identity and similarity of 31% and 49%, respectively. The N-termini of Gemin3, in which all nine DEAD-box helicase motifs reside, are more conserved than the C-termini. A region in the middle (451–573aa) of Drosophila melanogaster Gemin3 corresponds to the SMN-binding domain identified in higher eukaryotes [5].
Aiming to test whether the physical interaction between SMN and Gemin3 reported in higher eukaryotes [24] is conserved in Drosophila, a co-immunoprecipitation approach was pursued. We have generated a transgenic line expressing CFP::Gemin3. The CFP::Gemin3 gene is functional as it can rescue gemin3 mutants, which we describe later. In extracts derived from CFP::Gemin3 transgenic larvae, anti-SMN antibodies co-immunoprecipitate CFP::Gemin3 (Figure 2).
Two recessive lethal gemin3 alleles were identified: PBac{RB}e03688 (gemin3W) and P{PZ}Dhh1rL562 (gemin3R). We used PCR to confirm that the transposon insertion site of the gemin3W allele is located at 92 nt upstream of the transcription start site (Figure 3A; Figure S1). Part of the 5′ and 3′ piggyBac ends in the gemin3W allele were found to have been lost during the insertion. In the gemin3R allele, the P element inserted at 108 nt downstream of the transcription start site (Figure 3A; Figure S1). Since the P{PZ}-element insert sequence generates several premature stop codons, gemin3R is hypothesised to be an amorph.
Several studies were pursued to demonstrate that the recessive lethality of both transposon insertions is specific to gemin3 disruption, thereby confirming that gemin3 is an essential gene. First, complementation crosses revealed that both gemin3 alleles retain their recessive lethality in trans to each other and to a chromosomal deficiency that completely eliminates the gemin3 gene (Df[3L]ED4457). Second, a re-mobilisation screen of the P-element in the gemin3R allele, which is the only transposon that could be excised, recovered homozygous viable precise excision alleles or revertants. Third, both low ubiquitous gemin3 and CFP::gemin3 transgenic expression driven by 1032-GAL4 [25] rescued the lethality of gemin3R homozygotes and gemin3R/gemin3W transheterozygotes. However, neither of the above gemin3 transgenes can rescue the lethality of homozygous gemin3W, suggesting that a non-specific mutation may be causing the lethality associated with the gemin3W allele. Since the lethality observed in gemin3 heteroallelic mutants was specific to the loss of gemin3, further analysis concentrated on this genotype.
Expression of the CFP::gemin3 transgene under the control of tissue-specific drivers such as G7-GAL4 (muscle), elav-GAL4 (neuron), or the combination of both could not rescue the lethality of gemin3R homozygotes and gemin3R/gemin3W transheterozygotes, suggesting that animal survival also depends on the basal level of Gemin3 in tissues not covered by the expression of G7-GAL4 or elav-GAL4 drivers. Homozygous gemin3R mutants survive to the third instar larval stage, while the transheterozygotic gemin3R/gemin3W animals survive to the prepupal stage after both genotypes experience a prolonged wandering third instar larval stage. The expression of gemin3 at different developmental stages was compared by two-step RT-PCR. Essentially gemin3 mRNA was expressed at all developmental stages (Figure 3B). Supporting the amorphic allele hypothesis, we observed that expression of gemin3 mRNA was dramatically reduced in transheterozygous animals throughout their entire larval life, whereas the housekeeping control Tat-binding protein-1 (Tbp-1) transcripts remained detectable (Figure 3B). Heterozygous gemin3R adults have approximately half of the gemin3 mRNA transcript as that in wild-type animals (Figure 3B).
Although showing no dramatic mobility changes throughout the first and second larval stages, the gemin3R/gemin3W transheterozygotes exhibit a significantly decreased contraction rate at the third instar larval stage (Figure 4A and Video S1). The puparium formed by gemin3 heteroallelic mutants exhibited failed eversion of the spiracles and a large axial ratio (Figure 4B, C), the latter of which is most probably the result of a failure in body wall muscle contraction. Ubiquitous expression of the CFP::gemin3 transgene within this mutant background rescues the defects in mobility, spiracle eversion and abnormal axial ratio, confirming that the CFP::gemin3 transgene is functional and the above phenotypes exhibited by gemin3R/gemin3W transheterozygotes are specifically due to the disruption of Gemin3 function (Figure 4A–C). Mobility failure is probably not secondary to compromised muscle structure since gemin3 mutant larval fillets have an ordered pattern of muscle fibres without obvious muscle losses. In addition, there are no gross defects in the sarcomeric organisation in the gemin3 mutants (Figure 4D).
The obvious larval contraction defects of the gemin3 transheterozygotic mutants directed the research focus on the larval neuromuscular junction (NMJ). The present studies focus on the highly characterised type I NMJ innervating ventral longitudinal muscles 6 and 7, and aim at unveiling the presence of any morphological abnormalities in a gemin3 mutant background. To this end, larval muscle fillets were dissected and double-labelled with anti-HRP antibodies, which allow visualisation of the neuronal membrane, and an antibody against Discs-large (Dlg), a primarily postsynaptic scaffold protein localised to the subsynaptic reticulum that surrounds each bouton. Although no obvious motor neuron denervation was detected, gemin3 heteroallelic mutants exhibit an appreciative synaptic overgrowth before pupariation (Figure 5A) and a significantly increased synaptic area even when normalised to muscle size (Figure 5B). Expression of a gemin3 transgene in a mutant or wild-type background resulted in an increase in both NMJ and muscle area (data not shown). When normalized to muscle area, the NMJ area and branches in rescued gemin3 mutants restore to the wild-type range, whereas normalized NMJ area and branch numbers within single NMJs are significantly decreased when gemin3 was overexpressed (Figure 5B, C).
A truncated gemin3 transgene (gemin3ΔN), which lacks 424 amino acid residues from the N-terminus of Drosophila melanogaster Gemin3 and hence lacks the helicase core (Figure 3A), causes lethality on ubiquitous expression. Whilst highlighting the importance of the helicase domain to the function of Gemin3, the N-terminal truncated Gemin3 isoform is hypothesized to be a dominant-negative mutant. We used various drivers to investigate the effect on animal survival when gemin3ΔN is expressed in various temporal and spatial expression patterns (Table 1). No dramatic effect is observed when gemin3ΔN is expressed at 25°C under the control of elav-GAL4, nrv2-GAL4, D42-GAL4, OK6-GAL4, mef2-GAL4, or C57-GAL4 drivers (Figure 6A). However, expression of gemin3ΔN at 25°C by Act5C-GAL4, how-GAL4 or C179-GAL4 driver results in lethality, and that by the G7-GAL4 driver leads to a significant decrease in viability (Figure 6A). When the temperature shifted to 29°C to allow for maximal GAL4 activity, expression of gemin3ΔN by Act5C-GAL4, C179-GAL4, how-GAL4, or G7-GAL4 driver causes lethality, while that by mef2-GAL4 and C57-GAL4 drivers results in decreased viability (Figure 6B). Co-expression of an extra full-length gemin3 transgene but not a control gene such as GFP with the gemin3ΔN transgene significantly alleviates the driver-associated lethality (Figure 6 and data not shown). These experiments indicate that the lethality or low viability associated with the expression of gemin3ΔN in the mesoderm and larval muscles is specifically due to the disruption of Gemin3 function.
To confirm the driver-specific lethality pattern induced by the gemin3ΔN transgene, several gemin3 RNAi transgenic flies were isolated and tested to establish whether lethality can be induced when gemin3 knockdown occurs ubiquitously throughout the entire organism. Two RNAi transgenes, gemin3dwejra and gemin3munxar, fit this criterion. Reducing gemin3 gene activity using elav-GAL4, nrv2-GAL4, or D42-GAL4 has no effect on fly viability (Figure 7). In contrast, Gemin3 knockdown at both 25°C and 29°C via C179-GAL4 resulted in lethality. The how-GAL4 driver gave a similar effect when the gemin3dwejra and gemin3munxar RNAi transgene was expressed at both temperatures or at a temperature of 29°C, respectively (Figure 7). The lethality induced by gemin3munxar could be rescued by co-expressing a functional gemin3 transgene, thus excluding the possibility that lethality is the result of ‘off-target’ effects (Figure 7A, B).
Knockdown of gemin3 in the mesoderm and larval somatic musculature results in lethality at the late pupal stage, that is, pharate adults enclosed in pupae fail to eclose. Animals expressing gemin3ΔN under the control of the how-GAL4 driver often lead to pupariation and puparia have increased axial ratios, similar to the defects exhibited by the gemin3R/gemin3W transheterozygotes. In addition, how-GAL4≫gemin3ΔN pupae exhibited several morphological abnormalities, including head eversion defects, short legs, and short wings, although segmentation of the abdomen and mature eye pigments appear normal (Figure 8).
While they can walk and jump normally, eclosed flies with an mef2-GAL4-driven gemin3ΔN expression have a reduced ability to fly. In a flight assay, those flies show defective flight ability, similar to wild-type flies with clipped wings, which are flightless (Figure 9A and Video S2). The indirect flight muscles (IFMs) in mef2-GAL4≫gemin3ΔN flies are shrunken, resulting in increased spacing, and breakages are obvious between the muscle fibers. Frequently, large tears within the indirect flight muscles are observed in mef2-GAL4≫gemin3ΔN flies but not in wild-type flies (Figure 9B).
We have shown that CG6539, the Drosophila orthologue of vertebrate Gemin3, plays critical roles in larval and pupal development, especially in motor function.
Gemin3 or DP103 was first identified in mammalian culture cells through biochemical approaches [5],[15]. The Gemin3 protein has three critical features. First, the N-terminus of Gemin3 contains multiple helicase motifs including a DEAD-box. Second, Gemin3 interacts with SMN in vitro and in vivo [24]. Third, the Gemin3 and SMN proteins have a similar subcellular localization pattern [5],[26].
In Drosophila there are 29 DEAD-box RNA helicases [27]. Using human and mouse Gemin3 to BLAST the Drosophila melanogaster genome, CG6539, previously identified as DEAD/DEAH RNA helicase 1 (Dhh1), is the top hit. In the N-terminus, CG6539 contains 9 conserved RNA helicase motifs including a DEAD-box. A segment in the middle of CG6539, which corresponds to the SMN-binding domain in human Gemin3, is less conserved. Moreover, co-immunoprecipitation experiments using Drosophila larval muscle extracts show that Gemin3 binds to SMN in vivo. We have also carried localization assays, which demonstrate that Gemin3 co-localizes with SMN in the cytoplasm and nucleus [28] (RJC, KED, and JLL, unpublished data). Taken together, we feel confident that we have identified the Drosophila orthologue of vertebrate Gemin3.
Recently, an independent study by Fischer and colleagues also identified CG6539 as Drosophila Gemin3 through bioinformatic and biochemical approaches using Drosophila culture cells [23]. Both their study in Drosophila culture cells and this study in Drosophila tissues have shown that Gemin3 interacts with SMN, suggesting that Gemin3 is a bona fide component of the SMN complex in fruit flies, similar to that in vertebrate systems.
In this study, we have multiple lines of evidence demonstrating that Drosophila Gemin3 is essential for animal development and survival. Firstly, homozygous loss of gemin3 through a specific transposon insert (gemin3R) or a transheterozygous combination of two transposon inserts which do not complement each other (gemin3R/gemin3W) results in lethality at the larval and/or prepupal stage. Secondly, a functional gemin3 transgene specifically rescues the lethality and developmental defects caused by loss of gemin3. Thirdly, expression of a dominant-negative allele of gemin3 (gemin3ΔN) or Gemin3 knockdown by RNAi ubiquitously or even in a tissue-specific pattern results in lethality or reduced viability.
Gemin3-null mutants have recently been described in the mouse [29]. Heterozygous gemin3 mutant mice are healthy and fertile, with minor defects in the female reproductive system, whereas homozygous gemin3 knockout in mice leads to death at the 2-cell embryonic stage [29]. Thus, the lethality caused by loss of Gemin3 in Drosophila is consistent with the findings in Gemin3-null mice. However, while Gemin3-null mice died at an early embryonic stage, gemin3 mutant flies exhibit delayed lethality, which probably results from maternal contribution of the gemin3 transcript. In a separate study in female ovaries, we observed severe defects in nurse cells and oocytes when gemin3 is disrupted in germline cells (RJC, KED, and JLL, unpublished data).
The earliest clues pointing towards a motor function were a progressive loss of mobility and consequent long and thin puparia when Gemin3 function is lost. Similar phenotypes have previously been observed in mutants with disrupted Mlp84B, a muscle sarcomeric protein [30], or Tiggrin, an extracellular matrix ligand for the position-specific 2 integrins [31]. We also observe that gemin3 mutants have an overgrown NMJ though these could be a secondary response to the progressive loss of muscle power. The size ratio of NMJs to muscles is reduced when gemin3 is overexpressed raising the possibility that Gemin3 might also play a role in synaptic growth.
The requirement of Gemin3 in mesoderm and larval muscles for adult viability suggests a function of Gemin3 at the post-synaptic side. Based on the tissue-specific phenotypes uncovered, such a function is critical for pupal metamorphic changes and flight muscles. However, another possible explanation is that an earlier and wider disruption of Gemin3 by mesodermal-related drivers is responsible for the lethality, while late and local disruption of Gemin3 by neuroectodermal-related drivers causes milder phenotypes. More studies on the expression details of Gemin3 in pre- and post-synaptic tissues would help to distinguish those views.
Studies in vertebrate systems, in vitro and in vivo, have shown that Gemin3 directly binds to SMN [24]. A recent study in Drosophila culture cells [23] and this study in fly tissues confirm that the interaction between Gemin3 and SMN is conserved from fly to human.
This study raises the possibility of a functional interaction between Gemin3 and SMN. Loss of gemin3 phenocopies the larval mobility phenotypes observed in smn mutants [32]. Strong Gemin3 disruption in mesoderm and muscles led to striking developmental defects during metamorphosis, similar to those reported on disruption of SMN in a similar expression pattern [33]. A less severe gemin3 disruption in the developing musculature results in viable but flightless adult flies, which have flight muscle degeneration, similar to the phenotype in a hypomorphic smn mutant [34].
We observed that gemin3 mutants exhibit an overgrown NMJ before puparation and overexpression of gemin3 leads to a significant decrease in NMJ area and branches relative to muscle size. Interestingly, two studies describe a range of NMJ phenotypes for smn mutants [32],[35]. It is still not clear whether smn and gemin3 mutants have similar morphologic defects at the NMJ as the parameters and the segments used for NMJ analysis vary in different studies. Comparison of smn and gemin3 mutant NMJs with the same standard, as well as analysing the NMJ phenotype in smn and gemin3 double mutants would help to address this question.
The motor defects unravelled on disruption of Gemin3 function in Drosophila are very intriguing in view of its association with SMN, and the possible link to SMA. More studies are necessary to clarify the roles of SMN-Gemin3 interaction in development, which may help us to understand the molecular mechanisms of the devastating neurodegenerative disorder SMA.
The y w stock was used as the wild-type control. Transposon insertion alleles gemin3R (P{PZ}Dhh1rL562) and gemin3W (PBac{RB}e03688) were obtained from the Bloomington Drosophila Stock Centre (BDSC) at Indiana University and the Exelixis collection at Harvard Medical School, respectively. Complementation tests, transposon remobilisation and rescue studies were carried out according to standard genetic crossing schemes. The RNAi transgenic constructs UAS-gemin3dwejra (49505) and UAS-gemin3munxar (49506) were obtained from the Vienna Drosophila RNAi Center and their generation was described in Dietzl et al. [36]. GAL4 lines used in this study included 1032-GAL4, Act5C-GAL4 (BDSC), elav-GAL4 (BDSC), nrv2-GAL4 (gift from Paul Salvaterra, City of Hope National Medical Center, Duarte, California, USA), D42-GAL4 (BDSC), OK6-GAL4 (gift from Cahir O'Kane, University of Cambridge, Cambridge, UK), C179-GAL4 (BDSC), how-GAL4 (BDSC), mef2-GAL4 (gift from Barry Dickson, Research Institute of Molecular Pathology, Vienna, Austria), G7-GAL4 (gift from Aaron DiAntonio, Washington University, St. Louis, Missouri, USA) and C57-GAL4 (gift from Vivian Budnik, University of Massachusetts, Worcester, Massachusetts, USA); the spatial and temporal expression patterns are described in the Results. All stocks were cultured on standard molasses/maizemeal and agar medium in plastic vials or bottles at 25°C.
For the generation of the P{CFP::gemin3} transgenic construct, the PCR-amplified full-length coding sequence of gemin3 was ligated into the KpnI and XbaI restriction sites of the pUAST vector. The NotI and KpnI restriction sites of the resulting recombinant vector were then used to insert the cyan fluorescent protein (CFP) coding portion of the pECFP-C1 vector (BD Biosciences Clontech, Palo Alto, California, USA) upstream of the gemin3 sequence. The P{UAS-gemin3} construct was produced by ligating the gemin3 cDNA (Drosophila Genomics Resource Centre, Indiana University) in the pUAST vector using the KpnI and NotI restriction sites. The generation of the P{UAS-gemin3ΔN} involved PCR-amplification of the C-terminus of gemin3 followed by ligation into the KpnI and XbaI restriction sites of the pUAST vector. In both cases, the ligation products were used to transform E. coli competent cells using standard protocols. Correct transformants were further propagated and their harbouring plasmids were purified (Qiagen HiSpeed Plasmid Midi Kit, Qiagen Ltd., West Sussex, UK) prior to microinjection in y w embryos (BestGene Inc., Chino Hills, California, USA).
RNA was first extracted using the RNeasy kit (Qiagen Ltd.) and then reverse transcribed into cDNA using the QuantiTect Reverse Transcription Kit (Qiagen Ltd.) following manufacturer's instructions. PCR amplification of mRNA transcripts was performed using primers specific to gemin3 (forward: 5′-CACTGGCCAAAATGGATCTAA-3′ and reverse: 5′-GGCATTGCCTCAATGAGTTT-3′) and Tbp-1 (forward: 5′-CACCGAAAAGATCAAGGTCAA-3′ and reverse: 5′-CTTTGTTGACTCCGACCAGA-3′) mRNAs. RT-PCR products were resolved by electrophoresis on a 1.7% agarose gel containing ethidium bromide and bands were visualized by ultraviolet light.
Measurement of larval mobility involved placing age-matched larvae individually at the centre of a 0.7% agar plate and measuring the forward body wall contractions exhibited by each larva for 1 minute. Puparial axial ratios were calculated by dividing the length by the width of the puparia, both of which were measured from still images.
Adult viability assays were conducted by crossing GAL4 driver stocks to lines harbouring knockdown or truncated gemin3 transgenes. A week following eclosion, adult flies were screened and counted. Adult viability was calculated as the percentage of the number of adult flies with the appropriate genotype divided by the expected number for the cross.
The flight assay was done according to a modified protocol originally designed by Benzer [37]. In brief, a 1000 ml-graduated cylinder divided into 5 sectors was coated internally with mineral oil. Flies were introduced into the top of the cylinder through a funnel and the flies stuck in each sector were counted. The height flies stick in the cylinder is indicative of their flight capabilities.
Protein A beads washed and suspended in protein lysis buffer (2× protein lysis buffer [50 mM Tris pH8, 150 mM NaCl, 1 mM EDTA, and 1% v/v NP-40]+21× protease inhibitor cocktail [complete, Mini; Roche Diagnostics Ltd.]) were incubated with preimmune serum or an antigen-specific antibody, including rabbit anti-GFP (Abcam plc., Cambridge, UK) and rabbit anti-SMN (gift from Marcel van den Heuvel, University of Oxford). Sample lysates were prepared by dissecting body wall larval muscle fillets (∼30/IP) into cold 1× PBS followed by grinding into cold 2× protein lysis buffer. Following pre-clearing, lysates were incubated with beads coated with the appropriate target antigen-specific antibody. The beads were then washed in lysis buffer, and mixed with 4× NuPAGE LDS Sample Buffer (Invitrogen Ltd., Paisley, UK), 10× NuPAGE Reducing Agent (Invitrogen Ltd.) and deionised water. The mixture was then heated at 70°C in order to dissociate the immunoprecipitated antigen and any other macromolecules bound to it, followed by a brief spin. The bead-free supernatant was loaded onto a 4–12% NuPAGE Novex Bis-Tris pre-cast gel (Invitrogen Ltd.), resolved and probed for GFP according to standard Western blotting procedures.
Larvae were dissected in 1× PBS, fixed in 4% paraformaldehyde in PBS and then washed in 1× PBS+0.1% (v/v) Triton X-100 (PBT). The tissues were next subjected to overnight staining at 4°C by mouse anti-Discs large antibodies (1∶100; Developmental Studies Hybridoma Bank, University of Iowa, Iowa, USA). The next day, tissues were washed in PBT and stained for ∼2 hours at room temperature with anti-rabbit Alexa Fluor 488-conjugated secondary goat antibodies (1∶50), and anti-HRP goat antibodies conjugated to TRITC (1∶50; Jackson ImmunoResearch Laboratories Inc, West Grove, Pennsylvania, USA). Samples were then counterstained with nuclear-staining Hoechst 33342 (1∶500) and Cy5-conjugated actin-binding phallodin (1∶200) and mounted in Vectashield medium (Vector Laboratories Ltd., Peterborough, UK) prior to viewing with a Zeiss LSM 510 META confocal microscope.
ImageJ software (NIH) was used to quantify branch number, NMJ area, and muscle area from z-projections of confocal image stacks capturing ventral longitudinal muscles 6 and 7 (Segment A1). NMJ area constituted the presynaptic region stained by the anti-HRP antibody whereas branch number calculates the number of arborisations containing at least two boutons within a single NMJ. Both NMJ area and branch numbers were normalised through dividing each by the total muscle area of ventral longitudinal muscles 6 and 7.
Adult flies were fixed overnight in 4% (v/v) paraformaldehyde+2.5% (v/v) glutaraldehyde+0.1 M phosphate buffer pH7.2. The flies were then washed in 0.1 M phosphate buffer pH7.2 and post-fixed with 2% (w/v) osmium tetroxide for 2 hours at room temperature. Following a wash in water, the samples were subjected to a series of progressive dehydration steps in ethanol : water mixtures prior to embedding in Spurr's resin. Ultrathin sections were then made with a diamond knife, stained with Toluidine Blue and viewed under a light microscope.
|
10.1371/journal.pgen.0030148 | Gene Duplication and Adaptive Evolution of Digestive Proteases in Drosophila arizonae Female Reproductive Tracts | It frequently has been postulated that intersexual coevolution between the male ejaculate and the female reproductive tract is a driving force in the rapid evolution of reproductive proteins. The dearth of research on female tracts, however, presents a major obstacle to empirical tests of this hypothesis. Here, we employ a comparative EST approach to identify 241 candidate female reproductive proteins in Drosophila arizonae, a repleta group species in which physiological ejaculate–female coevolution has been documented. Thirty-one of these proteins exhibit elevated amino acid substitution rates, making them candidates for molecular coevolution with the male ejaculate. Strikingly, we also discovered 12 unique digestive proteases whose expression is specific to the D. arizonae lower female reproductive tract. These enzymes belong to classes most commonly found in the gastrointestinal tracts of a diverse array of organisms. We show that these proteases are associated with recent, lineage-specific gene duplications in the Drosophila repleta species group, and exhibit strong signatures of positive selection. Observation of adaptive evolution in several female reproductive tract proteins indicates they are active players in the evolution of reproductive tract interactions. Additionally, pervasive gene duplication, adaptive evolution, and rapid acquisition of a novel digestive function by the female reproductive tract points to a novel coevolutionary mechanism of ejaculate–female interaction.
| In a broad range of organisms, including humans, molecular interactions between the male ejaculate and the female reproductive tract play integral roles in sexual reproduction. Although these interactions are essential, the biochemical composition of the male ejaculate can change rapidly over short evolutionary time periods. It is often hypothesized that this rapid evolution reflects a coevolutionary relationship with the female reproductive tract. The paucity of research on females, however, presents a formidable challenge to empirical tests of this hypothesis. In this study, we sought to identify proteins in the female reproductive tracts of D. arizonae that may be interacting or coevolving with the male ejaculate. Unexpectedly, we discovered that D. arizonae females produce an array of “digestive” enzymes in their reproductive tracts. These classes of enzymes are normally found in the gut, where they degrade ingested food for nutritional uptake. In D. arizonae, these enzymes have resulted from recent gene duplications, and natural selection has caused rapid and radical changes in their amino acid sequences. We propose that this pattern of duplication and diversification reflects the “female side” of a coevolutionary relationship with the male ejaculate. Exploring the “male side” of this relationship is an important avenue for future research.
| Extensive research across a broad range of taxa has revealed that the proteins involved in sexual reproduction often evolve rapidly due to positive selection (reviewed in [1–3]). Although the selective forces that underlie this pattern remain unclear, it frequently has been postulated that adaptive evolution of reproductive proteins may result from intersexual coevolution [1–3]. Indeed, this has been demonstrated in the fertilization proteins of the free-spawning marine gastropod abalone, in which the male protein lysin and its female receptor, vitelline envelope receptor for lysin (VERL), both exhibit signatures of adaptive evolution [4–7]. In internally fertilizing organisms, however, such as mammals or insects, the biochemical interactions between male and female reproductive proteins may be vastly more complex. Reproductive outcomes depend not only on interactions between male and female gamete proteins, but additionally on interactions between male seminal proteins and proteins in the lumen of a female's reproductive tract [8–11].
Fruit flies of the genus Drosophila provide an important model system for exploring the function and evolution of reproductive tract interactions (reviewed in 9–12]). In Drosophila melanogaster, the male ejaculate comprises just under 100 proteins, several of which are known to stimulate important processes in mated females such as ovulation, oogenesis, and sperm storage (reviewed in [9–11]). Several male proteins either undergo proteolytic cleavage in mated females [13–15], or localize to specific portions of the female reproductive tract [16–18], indicating that ejaculate–female interactions are mediated biochemically by females. Between species, rapid changes in ejaculate composition frequently have resulted in lineage-specific seminal proteins [19–21], many of which may be novel coding sequences [22]. Additionally, molecular evolutionary studies indicate that a significant portion of this ejaculate is subject to positive selection in the melanogaster [23–25], obscura [26], and repleta species groups [27].
By comparison, the female side of reproductive tract interactions has received little attention. Female reproductive tract proteins have been identified transcriptionally only in D. simulans [28], and their functions remain entirely unknown. Furthermore, although several female reproductive tract proteins [28–30] and egg membrane proteins [31] show evidence of positive selection, these analyses largely have been confined to the melanogaster species group. It is unclear, therefore, how diversity in female reproductive physiology and mating system across the genus [reviewed in 12,32] is reflected in their reproductive proteins. This overall paucity of research on females presents a major obstacle to understanding the evolution of ejaculate–female interactions and the role of intersexual dynamics in the divergence of reproductive proteins.
Here we use a comparative expressed sequence tag (EST) approach to characterize candidate female reproductive tract proteins in D. arizonae. D. arizonae is a repleta group species that exhibits important differences from the melanogaster group in mating system and female physiology. D. arizonae females remate daily, while D. simulans females wait several days before remating [12]. Female promiscuity may affect the evolution of reproductive proteins by increasing the number of competing male ejaculates [33]. Females of D. arizonae additionally exhibit two remarkable post-mating physiological processes not seen in the melanogaster group. First, they incorporate peptide components of the male ejaculate into somatic tissues and oocytes [34], an adaptation which may help defray the cost of egg production during periods of resource limitation [35]. Second, they exhibit an insemination reaction, an opaque white mass of unknown biochemical composition that forms in the female uterus after copulation [36].
By comparing post-mating outcomes in inter- and intrapopulation crosses, several studies have presented evidence for ejaculate–female coevolution in natural populations of D. arizonae and its sister species D. mojavensis (most recent common ancestor, ∼1.5 million years ago [MYA]) [37–41]. Intrapopulation crosses of both species produce larger eggs than interpopulation crosses [38], a process known to be stimulated by several components of the male ejaculate in D. melanogaster (reviewed in [9–11]). Additionally, the insemination reaction exhibits a larger size and duration in interpopulation crosses relative to intrapopulation crosses, suggesting this trait is subject to sexually antagonistic coevolution [39]. Finally, desiccation resistance is higher in mated than unmated females [40], and the magnitude of this effect differs between inter- and intrapopulation crosses [41]. Such extensive evidence for physiological coevolution indicates this will be an exciting system to explore the molecular basis of reproductive tract interactions.
Our study identifies 241 candidate female reproductive proteins in D. arizonae, of which 31 show elevated rates of amino acid substitution suggestive of adaptive evolution. Unexpectedly, we also discovered three lineage-specific gene families of digestive proteases whose expression is specific to the lower female reproductive tract. These proteins exhibit strong signatures of adaptive evolution, and selected sites cluster near functionally important amino acids. The implications of these findings for ejaculate–female interactions and intersexual coevolution are discussed.
We sequenced a total of 2,304 ESTs derived from the D. arizonae lower female reproductive tract (parovaria, oviduct, spermathecae, seminal receptacle, and uterus) representing 649 unique proteins (for a complete list see Table S1). Of particular interest are proteins found on cell surfaces or in the lumen of this tissue, which interact directly with the male ejaculate and likely play an integral role in reproductive tract interactions [28]. We therefore designate candidate female reproductive proteins as those that exhibit secreted signal sequences, or transmembrane domains. The gross functional composition of the 241 candidate female reproductive proteins identified in this study (Figure 1) are similar to those of D. simulans [28], and include transport, signal transduction, and proteolysis.
To explore the evolutionary histories our candidate female reproductive proteins, we calculated the ratio of replacement to silent substitutions (dN/dS) between our D. arizonae ESTs and their orthologs in the D. mojavensis genome. Candidate female reproductive proteins exhibit significantly larger dN/dS values than intracellular proteins in our dataset (median test, p > 0.0001), suggesting that these proteins evolve more rapidly than their intracellular counterparts. This elevated rate of amino acid substitution is predicted if adaptive evolution of secreted and transmembrane proteins is a frequent consequence of molecular coevolution with components of the male ejaculate.
Under strict neutrality, only dN/dS ≫ 1 can be considered robust evidence of adaptive evolution. While several of our candidate genes show dN/dS > 1, none of these tests is statistically significant (Table 1). A literature survey has shown, however, that 95% genes that exhibit a pairwise dN/dS > 0.5 contain a class of sites with dN/dS ≫ 1 [28]. Of 227 pairwise comparisons, 31 (14%) were identified with dN/dS > 0.5, indicating they are likely experiencing positive selection (Table 1). This result is largely independent of gene duplication, as the estimated frequency of adaptive evolution it is still 13% when recent duplicates are excluded from the dataset.
On a functional level, several protein classes that commonly occur in seminal and fertilization proteins, including lipases, lectins, glycoproteins and proteases, are found in our candidates for adaptive evolution (Table 1). Roughly half of these 31 candidates, however, have no known function, and several others belong to functional classes that are not commonly represented among reproductive proteins. Proteins with unusual or unknown functions make excellent candidates for discovering genes which have acquired novel functions in a biochemical network which likely evolves rapidly. Future studies of these 31 candidates will yield significant insight into the function and evolution of reproductive tract interactions in the repleta species group.
Gene duplication plays an integral role in the evolution of D. arizonae female reproductive tract proteins. Specifically, 47% (16) of all secreted proteases in D. arizonae female reproductive tracts have at least one closely related paralog that also is expressed in these same tissues. Duplication events have been extremely recent; as multiple, tandemly-duplicated paralogs in the D. mojavensis genome correspond to only a single gene in D. virilis, the most closely related fully sequenced outgroup (most recent common ancestor, ∼23 MYA; reviewed in [42]). We therefore estimate that the duplication rate of secreted proteases expressed in D. arizonae tracts is 0.0298 (duplications per gene per million years, see Materials and Methods), which is 21-fold higher than the genome wide estimate for D. melanogaster (0.0014, [43]). Although the selective forces involved are yet obscure, such recent and pervasive gene duplication has not been seen in any class of reproductive protein yet studied, including D. simulans female reproductive proteins [28].
Four (of 16) duplicated proteases have resulted from two single gene duplication events. The remaining 12 duplicated proteases, however, are associated with small lineage-specific gene families. Each family contains four to six tandemly duplicated paralogs in the genome of D. mojavensis that are syntenic to a single ortholog in the genome of D. virilis (Figure 2). For brevity, we hereafter refer to these three families of tandem duplicates as protease gene family 1, 2, and 3. Phylogenetic analysis of D. arizonae ESTs, and coding sequences from the genomes of D. mojavensis, D. virilis, and D. grimshawi (http://rana.lbl.gov/drosophila), reveals the majority of these tandem duplicates in the D. mojavensis genome have a D. arizonae ortholog that is expressed in the lower female reproductive tract (Figure 3). This strongly suggests that the gene duplication events relate in some way to the reproductive function of these proteases. Indeed, reverse transcriptase PCR (RT-PCR) of all three gene families reveals that in adult D. arizonae these genes are exclusively expressed in the lower female reproductive tract (Figure 4). Gene copies present in the D. mojavensis genome that do not correspond to D. arizonae ESTs are likely not highly expressed.
While the function of these duplicated proteins in D. arizonae female reproductive tracts is unknown, they are often similar or identical in their key amino acid residues to several families of digestive proteases found almost exclusively in gastrointestinal tracts (Table 2). Specifically, protease gene families 1 and 2 share appreciable homology with trypsin, chymotrypsin, and elastase, serine endopeptidases commonly found in digestive tracts of both insects and mammals [reviewed in 44]. While, serine endopeptidases can also function in immune signaling cascades across a broad array of organisms, such proteases generally have secondary protein–protein interaction domains that allow for localized regulation of physiological responses [45]. No such domains are seen in either protease gene family 1 or 2, suggesting these proteases exhibit a primarily digestive function. Similar to the two families of serine endopeptidases, protease gene family 3 contains zinc metalloendoproteases very similar to astacin, a prominent digestive enzyme in the crayfish midgut [reviewed in 46]. The reproductive tract-specific expression of these proteases, coupled with recent, lineage-specific gene duplications, suggest that D. arizonae female reproductive tracts recently have acquired a novel digestive function. Digestive enzymes in female reproductive tracts likely have important implications for male reproductive success, and therefore, the evolution of the male ejaculate.
There is compelling evidence that directional selection has played an important role in the evolution of reproductive tract-specific secreted digestive proteases in D. arizonae females. All three families of digestive proteases exhibit a class of sites whose ratio of nonsynonymous to synonymous substitutions (dN/dS) is significantly greater than the neutral expectation of 1 (Table 2). dN/dS values for these selected sites range from 2 to 11.96, indicating certain amino acids in these proteins have experienced strong positive selection. Notably, the two single gene duplication events show no evidence of adaptive evolution (Table 2), indicating that directional selection has been exclusive to the lineage-specific families of digestive proteases.
In order to interpret selection in terms of both duplication and speciation events, we used the PAML free ratios model [47] to estimate dN/dS along every branch in each of the three phylogenies (Figure 3). Positive selection associated with three different speciation events suggests that ongoing changes in the biochemical environment of the female reproductive tract, including possible male contributions to this environment, have resulted in adaptive evolution in some of these proteins. A total of five gene duplication events are also immediately followed by a period of positive selection in one of the paralogous branches (dN/dS > 1), indicating neofunctionalization of a duplicate gene copy. The other seven duplication events however, are followed by elevated amino acid substitution rates (dN/dS = 0.2–1) but no evidence of adaptive evolution. This suggests that relaxed constraint created by functional redundancy between paralogs has also played an important role in the evolution of these gene families.
Evidence for adaptive amino acid evolution in duplicated genes implies that selection has acted to diversify the paralogs functionally. Indeed, in all three of the protease gene families, polar, nonpolar, and charged amino acids are seen to inhabit the same selected site in different paralogs. This indicates that directional selection has resulted in recurrent and radical amino acid substitutions, likely affecting the structure and function of the encoded proteins. By mapping selected sites onto predicted molecular structures, it is possible to make more specific inferences about how the biochemical function of these enzymes has been impacted by adaptive evolution. In the two families of serine endopeptidases (protease gene families 1 and 2), positive selection clusters near the catalytic triad: the three amino acids essential for proteolytic function (reviewed in [44]) (Figure 5). Furthermore, in protease gene family 1, positive selection is found adjacent to, and in one case synonymous with, three amino acid sites known to effect substrate specificity (reviewed in [48]). Collectively, these data indicate that directional selection has acted to diversify the catalytic activity of both families of serine endoproteases, and that protease gene family 1 has concomitantly undergone adaptive evolution for increased breadth in substrate specificity. Future functional studies of these enzymes, particularly in terms of how they interact with the male ejaculate, will yield significant insight into the selective pressures that underlie diversification of these extraordinary gene families.
Our most striking result was the observation of three lineage-specific radiations of secreted digestive proteases in D. arizonae female reproductive tracts. Although the biological significance of these gene duplications is yet unclear, they may relate to two unusual physiologies exhibited by both D. arizonae and D. mojavensis females. First, the insemination reaction must be degraded by females prior to oviposition or remating [36], a process that could require specialized digestive machinery. Second, female incorporation of ejaculate-derived protein, as observed in D. arizonae and D. mojavensis, could be facilitated by degrading seminal proteins and/or sperm into smaller fragments that are more easily absorbed.
Regardless of their physiological function, lower female reproductive–tract specific expression of digestive enzymes points to a novel form of ejaculate–female interaction, in which females may actively degrade, rather than process or activate [13–15], protein components of the male ejaculate. Digestion of seminal proteins or sperm would undoubtedly have important implications for male reproductive success, predicting an evolutionary response from males. Indeed, the association of these proteases with recent gene duplications and strong signatures of adaptive evolution suggests they are involved in an intersexual arms race. Exploring the male side of this interaction, therefore, is an important avenue of future research.
The 31 candidates for adaptive evolution also have important implications for reproductive tract interactions and intersexual coevolution. Roughly half of these proteins have no known function or conserved domain, suggesting they are enriched for novel biochemical functions. Additionally, the candidates include several classes of proteins that have not been implicated previously in reproductive tract interactions. Particularly intriguing are three transmembrane proteins with the conserved transporter domain MFS_1, for inorganic solutes (Table 1). Although the biochemical composition of the Drosophila ejaculate is largely unknown outside of its protein constituents, females of several species incorporate ejaculate-derived phosphorus into somatic tissues and oocytes [49]. It is unclear if these transporters underlie such a process in D. arizonae. Their presence and evolutionary history point, however, to nonpeptide biochemical interactions in female reproductive tracts which also may evolve rapidly.
If divergence of reproductive proteins is driven by intersexual dynamics, particularly sexually antagonistic coevolution [50–52], species with more promiscuous mating systems are predicted to exhibit comparatively more adaptive evolution in their reproductive proteins. D. arizonae is significantly more promiscuous than its previously examined congener D. simulans [28], and, consistent with the prediction, we find evidence that this difference in mating system may be reflected in the evolution of their female reproductive proteins. Specifically, we observed that candidate female reproductive proteins in our dataset exhibit higher dN/dS values than intracellular proteins, while this effect was not seen in similar comparisons between D. simulans and D. melanogaster [28]. Additionally, the estimated frequency of adaptive evolution in D. arizonae female reproductive tract proteins (14%) is significantly higher (Fisher's Exact Test p = 0.003) than that of D. simulans (5%) [28]. Although the experimental approach for these two studies was quite similar, differences in divergence times between D. arizonae and D. mojavensis (∼1.5 MYA, [37]), and D. simulans and D. melanogaster (∼3 MYA, [53]), could result in more stochastic influence on our measures of dN/dS. Firm conclusions about the effect of mating system on the evolution of female reproductive proteins therefore requires further empirical testing across a broader array of taxa.
Although the function and evolution of male seminal proteins have been researched extensively in both insects and mammals, our understanding of the female reproductive tract proteins with which they interact remains sparse. Our data, as well as previous research in the melanogaster group [28–30], indicate that rapid evolution is common among female reproductive tract proteins. We furthermore present compelling evidence that differences in female physiology and possibly mating system between Drosophila species are reflected in their reproductive tract proteins. Our research indicates that female reproductive proteins are active players in reproductive tract interactions, and that rapid evolution of seminal proteins must be considered in terms of their relationship with female counterparts.
D. arizonae used in this study were collected in December 2005 in Tucson, Arizona by E. S. K. A total of 873 lower reproductive tracts (parovaria, oviduct, spermathecae, seminal receptacle, and uterus) were dissected from mature adult females 9 d or older. In order to maximize transcriptional diversity obtained, dissected females were sampled from a diverse array of mating states. Of the females, 662 were from population bottles, while approximately 40 females were dissected from each of the following treatments: virgin, homospecifically mated 4–8 h postcopulation, homospecifically mated 24 h postcopulation, heterospecifically (to D. mojavensis) mated 4–8 h postcopulation, and heterospecifically mated 24 h postcopulation.
The harvested tracts were pooled into four separate aliquots of TRIZOL reagent (Invitrogen, http://www.invitrogen.com) and total RNA was extracted according to manufacturer instructions. Quality of these samples was verified with an Agilent 2100 bioanalyzer (http://www.home.agilent.com/), at which point they were pooled. mRNA enrichment was achieved by binding poly-A tails on Oligotex (Qiagen, http://www.qiagen.com/) spin columns. Quality of enriched mRNA was verified with an Agilent 2100 bioanalyzer, and the total yield (1.5 μg) was used for library construction with the Cloneminer cDNA library construction kit (Invitrogen). Approximately 300,000 colony-forming units were obtained with an estimated insert size of 1kb. Of these clones, 10,000 were picked with a QBOT (Genetix, http://www.genetix.com/) operated by the Arizona Genomics Institute (http://www.genome.arizona.edu/). Of these clones, 1,920 were sequenced bidirectionally, and an additional 384 were sequenced exclusively from their 5′ ends. All sequencing was done on at the Arizona Genomics Institute on an ABI 3700 DNA analyzer (https://products.appliedbiosystems.com/) with big-dye terminator chemistry.
Base calling and assembly were implemented in Phred and Phrap [54]. All bases with a Phred quality score below 20 (99% accurate) were excluded from further analysis. The estimated frequency of sequencing errors in included bases was 0.04%. BLASTN [55] (e-value = 0.01) against the GLEANR coding sequence annotations (from CAF1 assembly http://rana.lbl.gov/drosophila/) of the D. mojavensis genome was used to identify orthologs of D. arizonae ESTs. For ESTs with no good BLASTN hit to annotated coding sequence, BLASTN (e-value = 0.01) was implemented against the complete CAF1 assembly of the D. mojavensis genome. ESTs with BLAST hits in the D. mojavensis genome that contained long open reading frames were used to annotate additional genes in D. mojavensis by eye. No examples of ESTs with long open reading frames but no good BLASTN hit in the D. mojavensis genome were identified.
Translations of these coding sequences were used to identify secreted proteins and cell surface receptors using SignalP [56], and transmembrane proteins using TMHMM [57]. Conserved protein family (Pfam) domains were identified with hmmpfam [58]. Gene Ontology (GO) terms [59] were obtained from FlyBase (http://flybase.bio.indiana.edu/) for D. melanogaster homologs, or based on conserved Pfam domains if no D. melanogaster homolog was found. For explicit definitions of GO terms see http://www.geneontology.org/.
In total, the D. arizonae ESTs corresponded to 649 unique proteins in the D. mojavensis genome. The orthologous genes were aligned using CLUSTALW [60] and alignment accuracy was verified by eye. Maximum-likelihood estimates of nonsynonymous substitutions rate (dN), synonymous substitution rate (dS), and the ratio of nonsynonymous substitutions per nonsynonymous site to synonymous substitutions per synonymous site (dN / dS), were obtained from PAML [47]. For duplicated genes, only reciprocally monophyletic homologs were compared in pairwise analyses.
Sequence data for D. arizonae was obtained from the EST library, while sequences from D. mojavensis, D. virilis, and D. grimshawi were obtained from their unpublished, publicly available genomes (http://rana.lbl.gov/drosophila/). GENECONV was used to test for gene conversion between paralogs, using the method of Sawyer [61]. Phylogenetic reconstruction of multigene families was implemented in Mr. Bayes v3.0b4. Nested maximum-likelihood models of codon evolution were implemented in the codeml program of PAML [47] and compared using likelihood ratio tests. Two tests of positive selection were performed. In the first test, the neutral model (M1) is compared with the selection model, in which a class of sites is permitted to exhibit dN/dS (ω) > 1 (M2). In the second test, a beta distribution of site classes in which the most rapidly evolving is fixed to ω = 1 (M8a) is compared to a similar model in which the most rapidly evolving site class is permitted to exhibit ω > 1 (M8) [62]. Multiple initial values of ω were used to ensure convergence on the likelihood optima. For the second test, critical values of the test statistic are determined from Wong et al [63]. Lineage-specific selection patterns of dN/dS were determined by implementing branch-specific models [64].
A total of 34 secreted proteases were identified in D. arizonae female reproductive tracts. Using BLASTN homology and maximum-likelihood phylogenetic reconstruction implemented in PAUP*, we determined these 34 proteins correspond to 37 orthologs in the genome of D. mojavensis, and 23 orthologs in the genome of D. virilis (http://rana.lbl.gov/drosophila/). Assuming no gene conversion or gene loss, the total copy number of these genes was 23 at the divergence of the D. mojavensis and D. virilis lineages. Duplication rate can therefore be estimated by the following exponential growth equation:
Where CM is copy number of D. mojavensis (37), CA is the ancestral copy number (23), t is the divergence time between D. mojavensis and D. virilis (t = 23 MYA [42]), and r is the estimated rate of duplication per gene per million years.
D. arizonae RNA was extracted from 20 whole males, 70 reproductively mature females from population bottles lacking their lower reproductive tracts, and 70 lower reproductive tracts preserved in TRIZOL (Invitrogen) according to manufacturer instructions. Purified RNA was treated with DNAseI (Gibco, http://www.invitrogen.com/), and reverse transcribed with the iScript cDNA synthesis kit (Bio-Rad, http://www.bio-rad.com/). Resultant cDNA was diluted to 10 ng/μl, and used as a template for standard PCR using universal primers, with D. arizonae genomic DNA as a positive control. Primer sequences are as follows: Dmoj\GLEANR_8528-F, 5′-AAGAAGCGCACCAAGCACTTCATC-3′; Dmoj\GLEANR_8528-R 5′-TCTGTTGTCGATACCCTTGGGCTT-3′; protease gene family 1 -F1 5′-ATGTGGAATCTAAGCCCAGCCAA-3′; protease gene family 1 -F2 5′-RTAGATGGCAGTTGCTYCTYGTG-3′; protease gene family 1 -R1 5′-GATGYGATACCAATCACRGTGCT-3′; protease gene family 1 -R2 5′-ACGATRCCAATCACRGTGCYAGA-3′; protease gene family 2 -F1 5′-CTCAAACCGCARTAGYTRTCCT-3′; protease gene family 2 -F2 5′-CTTCAAGCCGCMGTWGCTGTCCT-3′; protease gene family 2 -R1 5′-CACCRCTGTGYTYCCTRATCCATTC-3′; protease gene family 2 -R2 5′-CACCGCWGTGCTCYYTGATCCATT-3′; protease gene family 3 -F1 5′-TGAAACCGATCCCAGACTTATAGC-3′; protease gene family 3 -F2 5′-ATGAAACCGATCCCGAGTTGATAG-3′; protease gene family 3 -R1 5′-ATCAGCCATGCTCAATTCTTGTCG-3′; and protease gene family 3 -R2 5′-ATCAGCCCAGCTTAATTCTAGTCG-3′.
Three dimensional structure was predicted by SWISS-MODEL [65], and visualized by Deep View. Selected sites were determined from Bayes Emperical Bayes calculation [66] implemented under M8 in PAML [47].
All sequences for this study are available from the National Institute for Biotechnology Information (NCBI) GenBank (http://www.ncbi.nlm.nih.gov/Entrez/index.html) accession numbers EV41299147751410–EV41383447752253 |
10.1371/journal.pgen.1004633 | Identification of a Regulatory Variant That Binds FOXA1 and FOXA2 at the CDC123/CAMK1D Type 2 Diabetes GWAS Locus | Many of the type 2 diabetes loci identified through genome-wide association studies localize to non-protein-coding intronic and intergenic regions and likely contain variants that regulate gene transcription. The CDC123/CAMK1D type 2 diabetes association signal on chromosome 10 spans an intergenic region between CDC123 and CAMK1D and also overlaps the CDC123 3′UTR. To gain insight into the molecular mechanisms underlying the association signal, we used open chromatin, histone modifications and transcription factor ChIP-seq data sets from type 2 diabetes-relevant cell types to identify SNPs overlapping predicted regulatory regions. Two regions containing type 2 diabetes-associated variants were tested for enhancer activity using luciferase reporter assays. One SNP, rs11257655, displayed allelic differences in transcriptional enhancer activity in 832/13 and MIN6 insulinoma cells as well as in human HepG2 hepatocellular carcinoma cells. The rs11257655 risk allele T showed greater transcriptional activity than the non-risk allele C in all cell types tested. Using electromobility shift and supershift assays we demonstrated that the rs11257655 risk allele showed allele-specific binding to FOXA1 and FOXA2. We validated FOXA1 and FOXA2 enrichment at the rs11257655 risk allele using allele-specific ChIP in human islets. These results suggest that rs11257655 affects transcriptional activity through altered binding of a protein complex that includes FOXA1 and FOXA2, providing a potential molecular mechanism at this GWAS locus.
| GWAS have identified more than 1200 loci contributing to risk of disease, including more than 70 loci associated with type 2 diabetes. With a majority of associated variants localized to non-coding regions of the genome, focus has moved to identifying the functional variants explaining the association signals. One mechanism by which variants may act is to affect activity of enhancer elements regulating target gene expression. In this study, we take advantage of recent advances in genome-wide annotation of human regulatory elements to prioritize candidate functional variants at the CDC123/CAMK1D locus. We identify two T2D-associated variants that overlap predicted regulatory enhancer elements. We demonstrate that one variant, rs11257655, shows allele-specific transcriptional enhancer activity in mammalian cell lines relevant to type 2 diabetes. We also show differential protein-DNA binding suggesting that the rs11257655 type 2 diabetes- risk allele increased transcriptional activity through binding a protein complex that includes FOXA1 and FOXA2. This study demonstrates that genome-wide maps of regulatory elements are a useful resource to guide identification of variants differentially affecting transcriptional activity and provides insight into molecular mechanisms underlying a T2D susceptibility locus.
| Type 2 diabetes is a complex metabolic disease with a substantial heritable component [1]. Over the past seven years, genome-wide association studies (GWAS) have successfully identified over 70 common risk variants associated with type 2 diabetes [2]–[5]. Association signals at many of these loci localize to non-protein-coding intronic and intergenic regions and likely harbor regulatory variants altering gene transcription. In recent years great advances have facilitated identification of regulatory elements genome-wide using techniques including DNase-seq and FAIRE-seq (formaldehyde-assisted isolation of regulatory elements), which identify regions of nucleosome depleted open chromatin, and ChIP-seq (chromatin immunoprecipitation), which identify histone modifications to nucleosomes and transcription factor binding sites. Several studies have successfully integrated trait-associated variants at GWAS loci with publicly available regulatory element datasets in disease-relevant cell types to guide identification of regulatory variants underlying disease susceptibility [6]–[10].
The CDC123 (cell division cycle protein 123)/CAMK1D (calcium/calmodulin-dependent protein kinase ID) locus on chromosome 10 contains common variants (MAF>.05) strongly associated with type 2 diabetes in Europeans (rs12779790, P = 1.2×10−10) [3], East Asians (rs10906115, P = 1.5×10−8) [4], and South Asians (rs11257622, P = 5.8×10−6) [5]. Fine-mapping using the Metabochip identified rs11257655 as the lead SNP [2]. The index variant and proxies (r2>.7) span an intergenic region of at least 45 kb between CDC123 and CAMK1D and overlap the 3′ end of CDC123 [3]. None of the type 2 diabetes-associated variants at this locus are located in exons. Analysis of the beta cell function measurements HOMA-B and insulinogenic index, derived from paired glucose and insulin measures at fasting or 30 minutes after a glucose challenge, demonstrated association of the risk allele at the CDC123/CAMK1D locus with reduced beta cell function, suggesting the beta cell as a candidate affected tissue [2], [11]. Another intronic variant (rs7068966, r2 = 0.18 EUR, 1000G Phase 1) located 50 kb away from rs12779790 is associated with lung function [12].
The transcript(s) targeted by risk variant activity at this locus remain unknown. CDC123 is regulated by nutrient availability in yeast and is essential to the onset of mRNA translation and protein synthesis through assembly of the eukaryotic initiation factor 2 complex [13], [14]. Evidence from previous GWA studies suggest cell cycle dysregulation as a common mechanism in type 2 diabetes; for example, type 2 diabetes association signals are found close to the cell cycle regulator genes, CDKN2A/CDKN2B and CDKAL1 [15]. CAMK1D is a member of the Ca2+/calmodulin-dependent protein kinase family which transduces intracellular calcium signals to affect diverse cellular processes. Upon calcium influx in granulocyte cells and hippocampal neurons, CAMK1D activates CREB-dependent gene transcription [16], [17]. Given the roles of cytosolic calcium in regulation of beta cell exocytotic machinery and of CREB in beta cell survival, CAMK1D may have a role in beta cell insulin secretion. In cis-eQTL analyses, the rs11257655 type 2 diabetes risk allele was more strongly and directly associated with increased expression of CAMK1D than CDC123 in both blood and lung [18], [19].
In this study we aimed to identify the variant(s) underlying the association signal at the CDC123/CAMK1D locus using genome-wide maps of open chromatin, chromatin state and transcription factor binding in pancreatic islets, hepatocytes, adipocytes and skeletal muscle myotubes. We measured transcriptional activity of variants in putative regulatory elements using luciferase reporter assays, and identified a candidate cis-acting SNP driving allele-specific enhancer activity in two mammalian beta cell-lines as well as hepatocellular carcinoma cells. We then evaluated DNA-protein binding in sequence surrounding this variant and identified allele-specific binding to key islet and hepatic transcription factors. Thus, our study provides strong evidence of a functional variant underlying the type 2 diabetes association signal at the CDC123/CAMK1D locus acting through altered regulation in type 2 diabetes-relevant cell types.
To identify potentially functional SNPs at the CDC123/CAMK1D locus, we considered variants in high LD (r2≥.7, EUR, 1000G Phase 1 release) with GWAS index SNP rs12779790. To further prioritize variants for functional follow up, we used genome wide maps of chromatin state (Figure 1) in available type 2 diabetes-relevant cell types including pancreatic islets, liver hepatocytes, skeletal muscle myotubes and adipose nuclei. Variant position was evaluated with respect to DNase- and FAIRE-seq peaks and several histone modifications, including H3K4me1 and H3K9ac. DNase and FAIRE are established methods of identification of nucleosome depleted regulatory regions [20], while H3K4me1 and H3K9ac are post-translational chromatin marks often associated with enhancer regions [21], [22]. We also assessed chromatin occupancy by transcription factors using available genome wide ChIP-seq data sets. Of 11 variants meeting the LD threshold, two SNPs were found to overlap chromatin signals. One SNP, rs11257655 (r2 = .74 with GWAS index SNP rs12779790), located 15 kb from the 3′ end of CDC123 and 84 kb from the 5′ end of CAMK1D, was a particularly plausible candidate overlapping islet, liver and HepG2 cell line DNase peaks, islet and liver FAIRE peaks, H3K4me1 and H3K9ac chromatin marks, and FOXA1 and FOXA2 ChIP-seq peaks in HepG2 cells (Figure S1). A second SNP, rs34428576 (r2 = .71 with rs12779790), overlapped a HepG2 DNase peak and displayed occupancy by FOXA1 and FOXA2 binding in HepG2 cells (Figure 1). No SNPs overlapped with DNase peaks in skeletal muscle myotubes.
To evaluate transcriptional activity of the SNPs in predicted regulatory regions, 150–200 bp surrounding each SNP allele was cloned into a minimal promoter vector and luciferase activity was measured in two beta cell lines, 832/13 rat insulinoma and MIN6 mouse insulinoma cells, and in HepG2 liver hepatocellular carcinoma cells. Four to five independent clones for each allele were generated and enhancer activity was measured in duplicate for each clone. A 151-bp region including rs11257655 (and rs36062557 due to proximity, r2 = .38 with rs11257655) showed differential allelic enhancer activity in both orientations in all three cell lines (Figure 2). The risk allele rs11257655-T showed significantly increased luciferase activity compared to the non-risk allele rs11257655-C (forward: 832/13 P = 6.3×10−3, MIN6 P = 1.7×10−5; HepG2 P = 8.0×10−5; reverse: 832/13 P = 2.2×10−3, MIN6 P = 9.9×10−5; HepG2 P = 2.0×10−3). Enhancer activity represents greater than a 1.4-fold (HepG2, MIN6) to 2.1-fold (832/13) increase in transcriptional activity relative to the non-risk allele in both the forward and reverse orientations. Compared to an empty vector control, enhancer activity was greatest in the islet cell lines (risk allele: 832.13, 4-fold; MIN6, 10-fold; HepG2, 1.6-fold).
A 179-bp region surrounding the second candidate SNP rs34428576 showed only moderate allele-specific activity, and only in the reverse orientation, in HepG2 cells (P = .02) and no allele-specific activity in islet cells (Figure S2).
To verify that rs11257655 and not rs36062557 accounted for allele-specific effects, we used site-directed mutagenesis to construct the remaining haplotype combinations. The T risk allele of rs11257655 exhibited >1.8 fold increased transcriptional activity compared to the non-risk allele C independent of rs36062557 genotype (Figure 3A, B). In contrast, altering alleles of rs36062557 on a consistent rs11257655 background showed no significant effect on transcriptional activity. Taken together, these data confirm that rs11257655 exhibits allelic differences in transcriptional enhancer activity and suggest it functions within a cis-regulatory element at the CDC123/CAMK1D type 2 diabetes-associated locus.
To assess whether alleles of rs11257655 differentially affect protein-DNA binding in vitro, biotin-labeled probes surrounding the T (risk) or C (non-risk) allele were incubated with 832/13, MIN6 or HepG2 nuclear lysate and subjected to electrophoretic mobility shift assays (EMSA). Band shifts indicative of multiple DNA-protein complexes were observed for both rs11257655 alleles (Figure 4A, 4B, 4C). In EMSAs from all three cell nuclear extracts, protein complexes were observed for the probe containing the T allele that were not present for the probe containing the C allele (832/13, arrow a; MIN6, arrows b, c, d; HepG2, arrows e, f) suggesting differential protein binding dependent on the rs11257655 allele. Competition of labeled T-allele probe with excess unlabeled T-allele probe more efficiently competed away allele-specific bands than excess unlabeled C-allele probe, demonstrating allele-specificity of the protein-DNA complexes (Figure 4A, 4B, 4C). rs11257655 did not show a differential protein binding pattern in EMSA using 3T3-L1 mouse adipocytes. To examine transcription factor binding to rs11257655, we used a DNA-affinity capture assay. We observed one protein band showing allele-specific binding to the T allele (Figure 4D) that was identified as transcription factor FOXA2 using MALDI TOF/TOF mass spectrometry.
A search in the JASPAR CORE database provided further evidence that the rs11257655 SNP is located within predicted binding sites for FOXA1 and FOXA2, with only the T risk-allele predicted to contain a FOXA1 and FOXA2 consensus core-binding motif (Figure 4E) [23]. To assess binding to FOXA1 and FOXA2, we performed supershift experiments incubating DNA-protein complexes with antibodies for these factors. Incubation of the T allele-protein complex with FOXA1 antibody resulted in a band supershift in 832/13 and HepG2 cells (asterisk, Figure 4A, 4C) A FOXA2-mediated supershift was observed in 832/13, MIN6 and HepG2 cells (asterisk, Figure 4A, 4B, 4C). Differences in antibody species reactivity may account for the lack of a visible FOXA1-mediated supershift in MIN6 cells. Collectively, these results suggest that rs11257655 is located in binding sites for a transcriptional regulator complex including FOXA1 and/or FOXA2, which bind preferably to the rs11257655-T allele in beta cell and liver cell lines.
To evaluate whether FOXA1 and FOXA2 bind differentially to rs11257655 in a native chromatin context, we performed allele-specific ChIP in human islets with different rs11257655 genotypes. FOXA1 was enriched 7.2-fold compared to IgG control in islets carrying a T allele while FOXA1 was not enriched in islets homozygous for C allele (Figure 5A). Although less robust, FOXA2 was enriched 4.2-fold in islets carrying a T allele compared to IgG control (Figure 5B). This direction of enrichment is consistent with the EMSA data (Figure 4). A region 28 kb downstream of rs11257655 with no evidence of open chromatin (chr10 control) was used as a negative control (Figure S3). These findings strengthen the conclusion that rs11257655 is part of a bona fide cis-regulatory complex binding FOXA1 and/or FOXA2 in human islets.
To determine whether CDC123 or CAMK1D are expressed in type 2 diabetes-relevant tissues, we measured and confirmed expression of both transcripts in human islets and hepatocytes (Figure S4A, S4B). These data are supported by RNA-seq evidence that both genes are expressed in islets [24]. Based on our results showing islet beta cells as a target tissue of risk variant regulatory activity, we assessed whether glucose treatment regulated CDC123 and CAMK1D transcript level. Glucose-mediated transcriptional changes in one of these genes might point to the more plausible candidate important in beta cell biology. In MIN6 cells treated with low (3 mM) and high (20 mM) concentrations of glucose for 16 hours, CAMK1D expression increased (P = .004; Figure S4C) while CDC123 expression remained unchanged (P = .22; Figure S4D). In 832/13 cells, CDC123 levels were significantly higher in cells stimulated with high glucose (P = 1.6×10−5; Figure S4E). We could not assess the effect of glucose on CAMK1D levels in 832/13 cells because this transcript level was below detection limits. While we confirm expression of CAMK1D and CDC123 in islets and hepatocytes, future studies over-expressing the target gene(s) in these tissues would be necessary to establish the mechanisms by which increased expression leads to diabetes risk.
Integration of genome-wide regulatory annotation maps with disease-associated variants identified through GWAS has great potential for elucidation of gene-regulatory variants underlying association signals. In this study, we expand the lexicon of disease-associated functional regulatory variation by examining the type 2 diabetes-association signal at the CDC123/CAMK1D locus. We prioritized candidate cis-regulatory variants and tested whether prioritized variants exhibited allele-specific transcriptional enhancer activity. We provide transcriptional reporter and protein-DNA binding evidence that rs11257655 is part of a cis-regulatory complex differentially affecting transcriptional activity. Additionally, we validate FOXA1 and FOXA2 as components of this regulatory complex in human islets.
In recent years, progress has been made in following up mechanistic studies of GWAS type 2 diabetes-association signals [6], [7], [9], [25]–[30], but challenges remain in sifting through the many associated variants at a locus to identify those influencing disease. We hypothesized that a common variant with modest effect underlies the association at the CDC123/CAMK1D locus and evaluated the location of high LD variants (r2≥.7; n = 11) at the locus relative to known transcripts and to putative DNA regulatory elements. We identified two variants that overlapped putative islet and/or liver regulatory regions and none located in exons. We did not assess variants in lower LD (r2<.7), and additional functional SNPs may exist at this locus acting through alternate functional mechanisms untested in the current study.
Based on our observation of type 2 diabetes-associated SNPs in regions of islet and liver open chromatin, we measured transcriptional activity in two mammalian islet cell models, rat 832/13 and mouse MIN6 insulinoma cells and in one hepatocyte cell model, human HepG2 hepatocellular carcinoma cells. In agreement with our previous observations [7], we found good concordance in allelic transcriptional activity of human regulatory elements across the two rodent islet cell types. Of the two SNPs predicted to be located in predicted enhancer regions, rs11257655 but not rs36062557 demonstrated allele-specific effects in islets and liver, suggesting that rs11257655 is a lead functional candidate. The rs11257655-T allele associated with type 2 diabetes risk displayed increased enhancer activity relative to the C allele, suggesting that increased expression of one or more genes, possibly CAMK1D or CDC123, may be associated with type 2 diabetes. Our subsequent analysis of protein binding revealed complexes that favored the rs11257655-T allele in 832/13, MIN6 and HepG2 cells. Consistent with predictions that the rs11257655-C allele may disrupt binding to the FOXA1 and FOXA2 transcription factors, we demonstrated that only the T allele of rs11257655 leads to FOXA1- and FOXA2-mediated supershifts. The ChIP enrichment of FOXA1 and FOXA2 in human islets from carriers of the T allele is concordant with EMSAs using nuclear extract from mouse and rat cell lines, further demonstrating the utility of rodent islet cell models to characterize human regulatory elements. Our results suggest that a cis-regulatory element surrounding rs11257655 may act in both islet and liver cells. Although we provide evidence that rs11257655 alleles differentially bind FOXA1 and FOXA2 in vivo, it is important to note that this enrichment was detected in isolated human islets. Future experiments will be needed to validate effects of rs11257655 within a whole organism environment. For example, recently zebrafish have been used to assay the regulatory potential of DNA sequences [31], [32].
FOXA1 and FOXA2 are members of the FOXA subclass of the forkhead box transcription factor family and are essential transcriptional activators in development of endodermally-derived tissues including liver and pancreas [33], [34]. In mature mouse β-cells, ablation of both transcription factors compared to ablation of FoxA2 alone leads to more pronounced impaired glucose homeostasis and insulin secretion, indicating that both factors are important in maintenance of the mature beta cell phenotype [35]. In addition, FoxA2 integrates the transcriptional response of mouse adult hepatocytes to a state of fasting [36]. FOXA1 and FOXA2 are thought to act as pioneer transcription factors, scanning chromatin for enhancers with forkhead motifs and opening compacted chromatin through DNA demethylation and subsequent induction of H3K4 methylation, epigenetic changes that likely render enhancers transcriptionally competent by allowing subsequent recruitment of transcriptional effectors [37]–[39]. Our data demonstrate increased transcriptional activity and increased binding of FOXA1 and FOXA2 to the rs11257655-T allele, suggesting that rs11257655 may be functioning as part of a transcriptional activator complex. Recent experiments in pancreatic islets support a role for FOXA transcription factors in activation of islet enhancers [40]. This same study also showed that FOXA2 binds in pancreatic islets in the T2D-associated region surrounding rs11257655. Further experiments, such as ChIP-seq of additional transcription factors, may identify other key factors present in the activator complex.
Both CAMK1D and CDC123 are candidate transcripts affected by variation at this locus. Cis-eQTLs in both blood and lung support an effect on CAMK1D but not CDC123. In blood, initial eQTL evidence for both genes were further analyzed by conditional analyses on the T2D lead SNP or rs11257655. The conditional analyses abolished the cis-eQTL signal for CAMK1D but not for CDC123, providing evidence that the T2D GWAS signal and the CAMK1D cis-eQTL signal are coincident [18]. In lung, the GTEx consortium identified an eQTL for CAMK1D with rs11257655 as a lead associated variant (P = 1.1×10−7); this and other T2D GWAS variants are the strongest cis-eQTLs for CAMK1D, while no significant eQTL is observed for CDC123 [19]. For both eQTLs, the rs11257655 type 2 diabetes risk allele is associated with increased CAMK1D transcript level, consistent with the direction of transcriptional activity we observed for this allele in islet and liver cells. Many eQTLs are predicted to be shared among tissues [41], and a recent study of the beta cell transcriptome reports good concordance of eQTL direction (R2 = .74–.76) between beta cells and blood-derived lymphoblastoid cell lines, fat and skin [42], suggesting that the CAMK1D eQTL may also exist in islets. Some eQTLs differ across tissues, and evidence of a consistent eQTL in islets would be valuable. Knockout mice provide further evidence supporting CAMK1D as a target gene. In FoxA1/FoxA2 beta cell-specific knockout mice, Camk1d expression was reported to be slightly reduced (1.8 fold, P = 0.13) [35], consistent with our conclusion that rs11257655 is part of a transcriptional activator complex that includes FOXA1 and FOXA2. Together, these data suggest that CAMK1D is a more plausible target for differential regulation by rs11257655 alleles.
The mechanism by which CAMK1D may act in type 2 diabetes biology is unclear. CAMK1D is a serine threonine kinase that operates in the calcium-triggered CaMKK-CaMK1 signaling cascade [17], [43]. In response to calcium influx, CAMK1D activates CREB- (cAMP response element-binding protein) dependent gene transcription by phosphorylation [17]. CREB is a key beta cell regulator important in glucose sensing, insulin exocytosis and gene transcription and β-cell survival [44], and FOXA2 has been shown to be necessary to mediate recruitment of CREB in fasting-induced activation of hepatic gluconeogenesis [36]. CAMK1D also has been reported to regulate glucose in primary human hepatocytes [45]. It is important to note that we cannot rule out cell cycle regulator CDC123 as a target for regulation by rs11257655.
In conclusion, we extend follow up studies of GWAS-identified type 2 diabetes-associated variants to the CDC123/CAMK1D locus on chromosome 10. We identify rs11257655 as part of a cis regulatory complex in islet and liver cells that alters transcriptional activity through binding FOXA1 and FOXA2. These data demonstrate the utility of experimentally predicted chromatin state to identify regulatory variants for complex traits.
Variants were prioritized for functional study based on linkage disequilibrium (LD) and evidence of being in an islet or liver regulatory element based on data from the ENCODE consortium [46]. Of 11 variants meeting the LD threshold (r2≥.7, EUR, with the GWAS index SNP rs12779790, 1000G Phase 1 release), two SNPs showed evidence of open chromatin [6], [9], [20], [47], histone modifications [21], [22], [48] or transcription factor binding and were tested for evidence of differential transcriptional activity.
Two insulinoma cell lines, rat-derived 832/13 [49] (C.B. Newgard, Duke University) and mouse-derived MIN6 [50] were maintained at 37°C with 5% CO2. 832/13 cells were cultured in RPMI 1640 (Cellgro/Corning) supplemented with 10% FBS, 1 mM sodium pyruvate, 2 mM L-glutamine, 10 mM HEPES and 0.05 mM β-mercaptoethanol. MIN6 cells were cultured in DMEM (Sigma), supplemented with 10% FBS, 1 mM sodium pyruvate, 0.1 mM β-mercaptoethanol. HepG2 hepatocellular carcinoma cells were cultured in MEM-alpha (Gibco) supplemented with 10% FBS, 1 mM sodium pyruvate and 2 mM L-glutamine.
Fragments surrounding each of rs11257655 (151 bp) and rs34428576 (179 bp) were PCR-amplified (Table S1) from DNA of individuals homozygous for risk and non-risk alleles. Restriction sites for KpnI and XhoI were added to primers during amplification, and the resulting PCR products were digested with KpnI and XhoI and cloned in both orientations into the multiple cloning site of the minimal promoter-containing firefly luciferase reporter vector pGL4.23 (Promega, Madison, WI). Fragments are designated as ‘forward’ or ‘reverse’ based on their orientation with respect to the genome. Two to five independent clones for each allele for each orientation were isolated, verified by sequencing, and transfected in duplicate into 832/13, MIN6 and HepG2 cell lines. Missing haplotypes of rs36062557-rs11257655 constructs were created using the QuikChange site directed mutagenesis kit (Stratagene).
Approximately 1×10−5 cells per well were seeded in 24-well plates. At 80% confluency, cells were co-transfected with luciferase constructs and Renilla control reporter vector (phRL-TK, Promega) at a ratio of 10∶1 using Lipofectamine 2000 (Invitrogen) for 832/13, and using FUGENE-6 for MIN6 and HepG2 cells (Roche Diagnostics, Indianapolis, IN). 48 h after transfection, cells were lysed with passive lysis buffer (Promega), and luciferase activity was measured using the Dual-luciferase assay system (Promega). To control for transfection efficiency, raw values for firefly luciferase activity were divided by raw Renilla luciferase activity values, and fold change was calculated as normalized luciferase values divided by pGL4.23 minimal promoter empty vector control values. Data are reported as the fold change in mean (± SD) relative luciferase activity per allele. A two-sided t-test was used to compare luciferase activity between alleles. All experiments were carried out on a second independent day and yielded comparable allele-specific results.
Nuclear cell extracts were prepared from 832/13, MIN6, and HepG2 cells using the NE-PER nuclear and cytoplasmic extraction kit (Thermo Scientific) according to the manufacturer's instructions. Protein concentration was measured with a BCA protein assay (Thermo Scientific), and lysates were stored at −80°C until use. 21 bp oligonucleotides were designed to the sequence surrounding rs11257655 risk or non-risk alleles: Sense 5′ biotin- GGGCAAGTGT[C/T]TACTGGGCAT 3′, antisense 5′ biotin- ATGCCCAGTA[G/A]ACACTTGCCC 3′ (SNP allele in bold). Double-stranded oligonucleotides for the risk and non risk alleles were generated by incubating 50 pmol complementary oligonucleotides at 95°C for 5 minutes followed by gradual cooling to room temperature. EMSA's were carried out using the LightShift Chemiluminescent EMSA Kit (Thermo Scientific). Binding reactions were set up as follows: 1× binding buffer, 50 ng/µL poly (dI•dC), 3 µg nuclear extract, 200 fmol of labeled probe in a final volume of 20 µL. For competition reactions, 67-fold excess of unlabeled double-stranded oligonucleotides for either the risk or non-risk allele were included. Reactions were incubated at room temperature for 25 minutes. For supershift assays, 4 µg of polyclonal antibodies against FOXA1 (ab23738; Abcam) or FOXA2 (SC6554X; Santa Cruz Biotechnology) was added to the binding reaction and incubation proceeded for a further 25 minutes. Binding reactions were subjected to non-denaturing PAGE on DNA retardation gels in 0.5× TBE (Lonza), transferred to Biodyne nylon membranes (Thermo Scientific) and cross-linked on a UV-light cross linker (Stratagene). Biotin labeled DNA-protein complexes were detected by chemiluminescence. EMSAs were carried out on a second independent day and yielded comparable.
DNA affinity capture was carried out as previously described [7]. Briefly, dialyzed nuclear extracts (300 µg) were pre-cleared with 100 µl of streptavidin-agarose dynabeads (Invitrogen) coupled to biotin-labeled scrambled control oligonucleotides. For DNA-protein binding reactions, 40 pmol of biotin labeled probe for either rs11257655 allele (same probe as for EMSA) or for a scrambled control were incubated with 300 µg nuclear extract, binding buffer (10 mM Tris, 50 mM KCL, 1 mM DTT), 0.055 µg/µL poly (dI•dC) and H20 to total 450 µL at room temperature for 30 minutes with rotation. 100 µL (1 mg) of streptavidin-agarose dynabeads were added and the reaction incubated for a further 20 minutes. Beads were washed and DNA-bound proteins were eluted in 1× reducing sample buffer (Invitrogen). Proteins were separated on NuPAGE denaturing gels and protein bands stained with SYPRO-Ruby. Protein bands displaying differential binding between rs11257655 alleles were excised from the gel and subjected to matrix assisted laser desorption time-of-flight/time-of-flight tandem mass spectrometry (MS) and analysis at the University of North Carolina proteomics core facility. For peptide identification, all MS/MS spectra were searched against all entries, NCBI non-redundant (NR) database, using GPS Explorer Software Version 3.6 (ABI) and the Mascot (MatrixScience) search algorithm. Mass tolerances of 80 ppm for precursor ions and 0.6 Da for fragment ions were used. In addition, two missed cleavages were allowed and oxidation of methionine was a variable modification.
Human islets from non-diabetic organ donors were provided by the National Disease Research Interchange (NDRI). Use of human tissues was approved by the University of North Carolina Institutional Review Board. Islet viability and purity were assessed by the NDRI. Islets were warmed to 37°C and washed with calcium- and magnesium-free Dulbecco's phosphate-buffered saline (Life Technologies) prior to crosslinking. For chromatin immunoprecipitation (ChIP) studies, approximately 2000 islet equivalents (IEQs) were crosslinked for 10 min in 1% formaldehyde (Sigma-Aldrich) at room temperature. Islets were lysed and chromatin was sheared on ice using a standard bioruptor (Diagenode; 20–22 cycles of 30 s sonication with 1 min rest between cycles) to a size of 200–1000 bp. IP dilution buffer (0.01% SDS, 1.1% Triton X-100, 1.2 mM EDTA, 16.7 mM Tris at pH 8.1, 167 mM NaCl, protease inhibitors) was added, 5% of the volume was removed and used as input, and the remainder was incubated overnight at 4°C on a nutating platform with FOXA1 or FOXA2 antibody or a species-matched IgG as control. Antibodies used for ChIP were the same as for EMSA; FOXA1 (Abcam) and FOXA2 (Santa Cruz). Protein A agarose beads (Santa Cruz) were added and incubated for 3 h at 4°C. Beads were then washed for 5 minutes at 4°C with gentle mixing, using the following solutions: Low Salt Buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris, 150 mM NaCl); High Salt Buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris, 500 mM NaCl); LiCl buffer (1 mM EDTA, 10 mM Tris, 250 mM LiCl, 1% NP-40, 1% Na-Deoxycholate), twice; and TE buffer (Sigma-Aldrich), twice. Chromatin was eluted from beads with two 15-minute washes at 65°C using freshly prepared Elution Buffer (1% SDS/0.1 M NaHCO3). To reverse crosslinks, 5 M NaCl was added to each sample to a final concentration of 0.2 M, and incubated overnight at 65°C; to remove protein, samples were incubated with 10 uL 0.5 M EDTA, 20 uL 1 M Tris (pH 6.5) and 3 uL of Proteinase K (10 mg/mL) at 45°C for 3 hours. DNA was extracted with 25∶24∶1 phenol:choloform:isoamyl alcohol, precipitated with 100% ethanol with 1 µl glycogen as a carrier, and resuspended in TE (Sigma). qPCR was performed in triplicate using SYBR Green Master Mix. Primers were designed to amplify a 99-bp region surrounding rs112576555; 5′-CTACTGCTTCTCCGGACTCG ′3′ and 5′- TGGCCTCAAGAGG GAGATAA -3′. Primers for a 133-bp control region not overlapping open chromatin and located 27 kb away were 5′-GCACCCATGGTACTGAAACC -3′ and 5′- CTTTTCCCG AGGAAGGAACT -3′. Dissociation curves demonstrated a single PCR product in each case without primer dimers. Fold enrichment was calculated as FOXA1/FOXA2 enrichment divided by IgG control. A one-sided t-test was performed to compare enrichment based on the direction of binding observed using EMSA.
To measure effects of glucose on expression of Cdc123 and Camk1d, 832/13 cells and MIN6 cells were washed with PBS and preincubated for 2.0 h in secretion buffer (114 mm NaCl, 4.7 mm KCl, 1.2 mm KH2PO4, 1.16 mm MgSO4, 20 mm HEPES, 2.5 mm CaCl2, 0.2% BSA, pH 7.2. For GSIS, cells were incubated in secretion buffer for an additional 2 hours or 16 hours in the presence of 3 mM or 20 mM glucose and then harvested for RNA.
Total cytosolic RNA was isolated using the RNeasy Mini Kit (Qiagen). RNA concentrations were determined using a Nanodrop 1000 (Thermo Scientific, Wilmington, DE, USA). For real-time reverse transcription (RT)–PCR, first-strand cDNA was synthesized using 8 ul of total RNA in a 20 µl reverse transcriptase reaction mixture (Superscript III First strand synthesis kit; Life Technologies). cDNA was diluted to contain equivalent to 20–55 ng/µl input RNA. To measure total human mRNA levels of CDC123, CAMK1D and B2M, gene-specific primers and fast SYBR Green Master Mix (Life Technologies) were used (Table S2). TaqMan designed gene expression assays (Life Technologies) were used to measure Cdc123, Camk1D and Rsp9 (housekeeping gene) mRNA levels of mouse and rat cells. All PCR reactions were performed in triplicate in a 10-µl volume using a STEPOne Plus real-time PCR system (Life Technologies). Serial 3-fold dilutions of cDNA from pooled human tissues, 832/13 or MIN6 cells as appropriate were used as a reference for a standard curve. Statistical significance was determined by two-tailed t-tests.
|
10.1371/journal.pntd.0002766 | Trypsin- and Chymotrypsin-Like Serine Proteases in Schistosoma mansoni – ‘The Undiscovered Country’ | Blood flukes (Schistosoma spp.) are parasites that can survive for years or decades in the vasculature of permissive mammalian hosts, including humans. Proteolytic enzymes (proteases) are crucial for successful parasitism, including aspects of invasion, maturation and reproduction. Most attention has focused on the ‘cercarial elastase’ serine proteases that facilitate skin invasion by infective schistosome larvae, and the cysteine and aspartic proteases that worms use to digest the blood meal. Apart from the cercarial elastases, information regarding other S. mansoni serine proteases (SmSPs) is limited. To address this, we investigated SmSPs using genomic, transcriptomic, phylogenetic and functional proteomic approaches.
Genes encoding five distinct SmSPs, termed SmSP1 - SmSP5, some of which comprise disparate protein domains, were retrieved from the S. mansoni genome database and annotated. Reverse transcription quantitative PCR (RT- qPCR) in various schistosome developmental stages indicated complex expression patterns for SmSPs, including their constituent protein domains. SmSP2 stood apart as being massively expressed in schistosomula and adult stages. Phylogenetic analysis segregated SmSPs into diverse clusters of family S1 proteases. SmSP1 to SmSP4 are trypsin-like proteases, whereas SmSP5 is chymotrypsin-like. In agreement, trypsin-like activities were shown to predominate in eggs, schistosomula and adults using peptidyl fluorogenic substrates. SmSP5 is particularly novel in the phylogenetics of family S1 schistosome proteases, as it is part of a cluster of sequences that fill a gap between the highly divergent cercarial elastases and other family S1 proteases.
Our series of post-genomics analyses clarifies the complexity of schistosome family S1 serine proteases and highlights their interrelationships, including the cercarial elastases and, not least, the identification of a ‘missing-link’ protease cluster, represented by SmSP5. A framework is now in place to guide the characterization of individual proteases, their stage-specific expression and their contributions to parasitism, in particular, their possible modulation of host physiology.
| Schistosomes are blood flukes that live in the blood system and cause chronic and debilitating infection in hundreds of millions of people. Proteolytic enzymes (proteases) produced by the parasite allow it to survive and reproduce. We focused on understanding the repertoire of trypsin- and chymotrypsin-like Schistosoma mansoni serine proteases (SmSPs) using a variety of genomic, bioinformatics, RNA- and protein-based techniques. We identified five SmSPs that are produced at different stages of the parasite's development. Based on bioinformatics and cleavage preferences for small peptide substrates, SmSP1 to SmSP4 are trypsin-like, whereas SmSP5 is chymotrypsin-like. Interestingly, SmSP5 forms part of a ‘missing link’ group of enzymes between the specialized chymotrypsin-like ‘cercarial elastases’ that help the parasite invade human skin and the more typical chymotrypsins and trypsins found in the nature. Our findings form a basis for further exploration of the functions of the individual enzymes, including their possible contributions to influencing host physiology.
| Schistosomiasis caused by Schistosoma blood flukes is a chronic disease with more than 200 million people infected [1]. Schistosome larvae (cercariae), released into an aquatic environment from snail intermediate hosts, penetrate human skin and subsequently develop into adult worms. Adult worms reside in the host vascular system as male/female pairs, and survive for many years, if not decades [2], producing hundreds of eggs per day. Morbidity arises from the host immune responses to eggs in tissues [3]. Treatment relies on one drug, praziquantel, and no effective vaccine has yet been developed [4]. During its complex life cycle, the parasite survives in various environments by presenting or releasing bioactive molecules that aid survival and modulate host physiology [5], [6]. Disruption of these potential mechanisms by specific drugs/vaccines may provide therapeutic benefits.
Proteolysis is a fundamental physiologic process [7], [8]. Proteases (proteolytic enzymes) are crucial to parasitism, including by schistosomes, in facilitating invasion, nutrient intake, hatching, excystment, immune evasion [9], [10] and modulation of host physiology [10]–[15]. Most schistosome research has focused either on cysteine and aspartic proteases (MEROPS database Clans CA and AA, respectively [8]), which are responsible for digesting the blood meal [16], [17] or on the serine proteases (SPs), known as cercarial elastases (CEs; Clan PA, family S1) that facilitate active penetration of the mammalian host [18]–[20].
Regarding the nomenclature for eukaryotic SPs, whereas members of the S1 or ‘chymotrypsin’ family of SPs share a similar tertiary structure, their substrate cleavage specificities differ [8]. Thus, substrate preferences at the P1 subsite [21] may be divided into trypsin-like (P1 preference for basic residues), chymotrypsin-like (bulky hydrophobic residues) and elastase-like (small aliphatic residues) [7].
Despite their name, which was derived from their ability to cleave insoluble elastin, the S. mansoni CEs have a chymotrypsin-like P1 specificity [22] due to preferences for phenylalanine and leucine. In contrast to these well-studied CEs [18]–[20], there are fewer descriptions of ‘non-CE’ Clan PA, family S1 serine proteases in S. mansoni (SmSPs) [6], [12]–[15], [23], [24].
Among these, SmSP1 (S. mansoni serine protease 1, GenBank AJ011561), has been partially described [13], [14]. The open reading frame (ORF) of SmSP1 comprises two non-proteolytic domains, followed by a C-terminal trypsin protease domain. Expression of the trypsin domain (mRNA and protein) was noted in adult worms with a significant accumulation in the tegument (surface) of males [13]. Another SmSP was identified (under TC16843 code) by microarray analysis with a remarkably elevated expression in post-infective larvae (schistosomula) that had been maintained in vitro [23]. Two additional biochemical studies support a function for schistosome SPs in modulating host physiology. Specifically, a protein fraction of S. mansoni adult worm extracts was shown to possess kallikrein-like protease activity [12]. The isolated native enzyme, termed sK1, cleaved kallikrein substrates and processed kininogen to bradykinin which induced strong vasodilatation and decreased arterial blood pressure in experimental rats; sK1 was found in higher abundance in males [12]. Both, sK1 and SmSP1, are proposed to regulate host vascular functions [6]. In the second study, SP activity in extracts of S. mansoni eggs induced significant fibrinolytic activity and was associated with a 27 kDa protein [15]. This protease activity had a similar cleavage pattern to human plasmin and it was hypothesized that the enzyme blocks the intravascular deposition of fibrin by platelets activated by schistosome eggs [15].
In the present study, we sought to understand the gene repertoire of non-cercarial elastase SmSPs by employing a series of genomic, transcriptomic, proteolytic and phylogenetic approaches. In addition to SmSP1, we identified and re-annotated four distinct SmSPs in the S. mansoni GeneDB genome database [25], [26] and term them SmSP2 through SmSP5 according to a previous terminology [13]. The data reveal intriguing expression profiles and phylogenetic relationships that stimulate further study of the individual proteases involved, and their contributions to modulating host physiology.
Mice are kept in the animal facility of the Biology Center (Academy of Sciences of the Czech Republic) in Ceske Budejovice and all animal experiments are carried out as approved by the Animal Rights Ethics Committee under protocol no. 068/2010 issued according to the national regulation 246/1992 Sb.
A Liberian isolate of S. mansoni has been maintained in the laboratory by cycling between CD-1 mice and the freshwater snail, Biomphalaria glabrata. Mice were subcutaneously injected with 200 cercariae and sacrificed 6–7 weeks post-infection by intra-peritoneal injection of thiopental (50 mg/kg). Adults, eggs and miracidia were isolated as described previously [27]. Cercariae were obtained from infected snails induced to release the parasite under a light stimulus. Cercariae were chilled on ice, collected and transformed mechanically to schistosomula [27], [28], which were then cultured for five days under a 5% CO2 atmosphere at 37°C in Basch Medium 169 [29] containing 5% fetal calf serum and 1% ABAM (antibiotics/antimycotics; Sigma-Aldrich). Daughter sporocyst material was isolated by excision of the hepato-pancreases from two month-infected B. glabrata snails. The hepato-pancreases from uninfected snails were used as a negative control when evaluation gene expression.
Adult worms, eggs, miracidia, daughter sporocysts, cercariae and schistosomula were re-suspended in 500 µl of Trizol reagent (Life Sciences) and processed [30]. Single-stranded cDNA was synthesized from total RNA by SuperScript II reverse transcriptase (Life Sciences) and an oligo dT18 primer, and then stored at −20°C.
Genes encoding complete SmSPs or their specific domains were retrieved from the S. mansoni genome database (S. mansoni GeneDB, available at http://www.genedb.org/Homepage/Smansoni) through BLAST searches. Amino acid sequences of vertebrate family S1 SPs were used as queries. Specific PCR primers were employed to amplify each of the sequences retrieved, and the respective amplicons cloned into the TOPO TA 2.1 vector (Life Technologies) for propagation in TOP10 E. coli cells. For SmSP4 and SmSP5, full-length sequences were obtained by 5′ and 3′ RACE (Rapid Amplification of cDNA Ends, Life Technologies).
Based on more recent annotations, the original sequence information for SmSP4 and SmSP5 (GenBank XM_002572739 and XM_002574902) were corrected in the S. mansoni GeneDB database. All newly described SmSP sequences were deposited in GenBank under the accession numbers listed in Table 1. For genes with multi-domain structures, PCR analysis was performed using domain-specific primers in order to detect possible differential expression.
Gene expression of the SmSPs was assessed using RT-qPCR. For genes with multi-domain structures (SmSP1 and SmSP3), the expression levels of individual domains were evaluated separately. cDNA for various life stages was generated using the mRNA isolation protocol described above and previously [30]. For mRNA isolation, 3 infected B. glabrata hepatopancreases and approximately 20 adult pairs, 500 hundred eggs, cercariae and schistosomula were used. Primers for quantitative PCR analysis were designed using the Primer 3 software (http://frodo.wi.mit.edu/ [31],), in order to amplify 150–250 bp regions of the targeted genes or their domains. Primer efficiency was evaluated by serial dilutions of both the primers and the cDNA template as described [32], [33]. Two to three primer pairs were generated per target from which one primer set with optimal efficiency and generating only a single dissociation peak was used (see Supporting Information Table S1).
Reactions, containing SYBR Green I Mastermix (Eurogentech), were prepared in final volumes of 25 µL in 96-well plates [30]. The amplification profile consisted of an initial hot start (95°C for 10 min), followed by 40 cycles comprising 95°C for 30 s, 55°C for 60 s and 72°C for 60 s, and ended with a single cycle of 95°C for 60 s, 55°C for 30 s and 95°C for 30 s. PCR reactions were performed in duplicate for each cDNA sample. At least one biological replicate, i.e., samples from a different RNA isolation was performed for each gene target. Analysis of the cycle threshold (CT) for each target was carried out as described [30] and employed S. mansoni cytochrome C oxidase I (SmCOX I, GenBank AF216698, [33]) as the sample normalizing gene transcript [27]. Finally, the resulting transcript values were calculated as a percentage of the expression of the normalizing gene (SmCOX I) which was set as 100%. Transcript levels were expressed as log functions and as a percentage relative to that of SmCOX I in order to compare variable expression patterns. The threshold for significance of expression was set to 0.01% of the expression of SmCOX I.
The amino acid sequences of 96 vertebrate and invertebrate members of the S1 serine protease family were aligned in MAFFT [34] using the E-INS-i method, and gap opening (–op) and extension penalties (–ep) of 5.0 and 0.0, respectively. The non-catalytic domains and N-terminal extensions were excluded from the resulting alignment in BioEdit (v7.0.5.2; [35]). The bacterial trypsin from Streptomyces griseus was used as an outgroup. The list of family S1 proteases (SPs sequences) used for the phylogenetic analysis is in the Supplementary Table S2. The Maximum Parsimony analysis was performed in PAUP* (v4.b10; [36]), using a heuristic search with random taxa addition, the ACCTRAN option, and the TBR swapping algorithm. All characters were treated as unordered whereas gaps were treated as missing data. Maximum Likelihood analysis was performed in RAxML under the WAG model [37]. Clade support values were calculated from 1000 bootstrap replicates with random sequence additions for both analyses. All trees were displayed using the TreeView32 program [38].
Fifty pairs of adult worms, 1 000 eggs or 1 000 schistosomula were washed five times in Basch Medium 169 containing 1% Fungizone (Gibco) and allowed to stand for 1 h at 37°C in 5% CO2. Samples were washed 10 times and then incubated in the same Basch Medium overnight (adults and eggs) or for five days (schistosomula) at 37°C in 5% CO2. Parasite material was then washed 10 times in M-199 medium (alternative medium for schistosoma cultivation without serum and proteins,Gibco) containing 1% ABAM and incubated in the same medium for 16 h at 37°C in 5% CO2. Medium containing E/S products was removed and filtered using an Ultrafree-MC 0.22 µm filter (Millipore). Filtered medium was buffer exchanged into ice-cold 1× PBS (pH 7.4) and concentrated at 4°C to a 2 ml final volume by centrifugation at 4000 g using an Amicon 10000 Ultra-15 Centrifugal Filter Unit (Millipore). The total volume of PBS used for buffer exchange was 40 ml. Samples (0.04–0.37 mg protein/ml) were frozen in liquid nitrogen and stored at −80°C.
Soluble protein extracts (1–5 mg protein/ml) from S. mansoni adults, eggs and 5 day-old schistosomula were prepared by homogenization in 50 mM Tris-HCl buffer, pH 8.0, containing 1% CHAPS, 1 mM EDTA and 10 µM of the cysteine protease inhibitor, E-64, in an ice bath. The extracts were cleared by centrifugation (16,000 g, 10 min, 4°C), filtered with an Ultrafree-MC 0.22 µm and stored at −80°C.
Proteolytic activities were measured in a kinetic continuous assay using the following peptidyl fluorogenic, 7-amino-4-methylcoumarin (AMC) substrates (Bachem) at a 50 µM final concentration: Z-F-R-AMC (Z, Benzyloxycarbonyl), Bz-F-V-R-AMC (Bz, Benzoyl), Z-G-P-R-AMC, P-F-R-AMC, Boc-L-R-R-AMC (Boc, t-Butyloxycarbonyl), Boc-Q-A-R-AMC, Boc-V-L-K-AMC Suc-A-A-F-AMC (Suc, Succinyl), Suc-A-A-P-F-AMC, Suc-L-Y-AMC, MeOSuc-A-A-P-V-AMC (MeOSuc, 3-Methoxysuccinyl), Z-G-G-L-AMC and Z-V-K-M-AMC. Assays were performed at 37°C in 96-well black microplates in a total volume of 100 µl. Parasite extracts (1–3 µg) or E/S products (0.05–1 µg) were pre-incubated for 10 min in 150 mM Tris-HCl, pH 8.0, containing 10 µM E64, 1 mM EDTA in the presence or absence of 0.5 mM of the serine protease inhibitors, Pefabloc SC and PMSF. E64 was included routinely in extract preparations in order to inhibit Clan CA cysteine protease activity that is present in the life-stages examined [30], [39], [40]. Hydrolysis of substrate was measured continuously using an Infinite M1000 microplate reader (Tecan) at excitation and emission wavelengths of 360 and 465 nm, respectively. All measurements were performed in triplicate and results normalized to protein concentration.
A spatial model of SmSP1 was constructed using the template X-ray structure of bovine trypsin in complex with the peptidyl inhibitor leupeptin (PDB entry 1JRT) and utilizing a pairwise sequence alignment generated by the BLAST program (BLOSUM62 substitution matrix). The homology module of the MOE program was used for modeling the SmSP1 structure (MOE: Chemical Computing Group; http://www.chemcomp.com). The conformation of leupeptin was refined by applying the LigX module of the MOE. The final binding mode of the inhibitor was selected by the best fit model based on the London dG scoring function and the generalized Born method [41]. Molecular images were generated with UCSF Chimera (http://www.cgl.ucsf.edu/chimera/). The electrostatic surface potential was calculated using the APBS software [42] and input data were prepared using PDB2PQR [43].
Genes were selected in silico based on a proteolytic domain organization that matched with family S1 serine proteases: cercarial elastases were excluded because of their detailed studies previously [20], [22]. The five remaining SmSP genes, including the previously sequenced and partially characterized SmSP1 [13], [14], were cloned and sequenced. The other four gene sequences named SmSP2 through SmSP5 (Table 1) were significantly corrected and re-annotated in the primary database (S. mansoni GeneDB) due to various sequence inaccuracies. The sequences of SmSP2 through SmSP5 were deposited into the GenBank as KF510120, KF510121, KF510122, KF939306, respectively. The sequence of SmSP1 defined here was also deposited (KF535923) because of sequence differences from the original description (CAA09691 [13]) and from the information in S. mansoni GeneDB (Smp_030350; Figure S1). A search of the Schistosoma japonicum genome [44] indicates that orthologs for each of the SmSPs are present; SjSP1 (GeneDB Sjp_0012180, GenBank N/A), SjSP2 (Sjp_0100980, CAX74751), SjSP3 (Sjp_0023390, CAX73257), SjSP4 (Sjp_0047680, N/A) and SjSP5 (Sjp_0114710, CAX73292).
The sequence domain organization for the particular proteases is represented in Figure 1. Based on sequence homology analysis, we describe SmSP1 as a multi-domain protein comprising a matriptase-like structure made up of Complement-Uegf-BMP-1 (CUB) extracellular and plasma membrane-associated domains, a LDL-binding receptor domain class A (LDLa domain) and a S1 family serine protease domain. However, the full gene product has been detected only in the eggs, whereas in other parasite stages, the CUB and protease domains are expressed as separate spliced products, as demonstrated by PCR and sequencing (Figure S2).
Primary sequence homology analysis shows that SmSP2 to SmSP5 are distinct molecules with the same family S1 type catalytic protease domain at the C-terminus, but with different N-terminal extensions which include a potential pro-peptide, i.e., a peptide that is removed during zymogen activation. The N-terminal extensions vary from 201 residues in SmSP2 to just a seven residues in SmSP5 (Figure 1). SmSP1, SmSP3 and SmSP5 do not contain a predicted signal sequence for the secretory pathway as identified by the SignalP program [45]. In contrast, SmSP2 and SmSP4 are synthesized as pre-pro-proteins with a typical N-terminal signal peptide preceding an N-terminal extension region containing a putative pro-peptide (‘activation peptide’) that is then followed by the protease domain (Figure 1). The pro-peptide is separated from the protease domain of SmSPs by a basic residue, Arg or Lys (Figure 2) which constitutes a potential activating cleavage site, i.e., is hydrolyzed during protease maturation as is known for other S1 family proteases [7]. For SmSP3, the N-terminal extension contains an incomplete CUB domain. PCR and sequencing revealed that, as found for SmSP1, the CUB and the protease domains of SmSP3 are only co-expressed in eggs whereas they are separate spliced gene products in the other stages (Figure S2). SmSP5 contains a Thr/Asn rich C-terminal sequence extension not present in orthologous SPs from other trematodes (Figure S4).
The catalytic protease domains of SmSP1 to SmSP4 share significantly greater sequence identity (about 30%) with each other than with SmSP5 (about 20%; Figure S3). All five SmSPs have a catalytic triad in the order of His, Asp and Ser that is typical for S1 family proteases; also, the regions surrounding the catalytic triad residues have the most notable sequence identity (Figure 2). The protease domains of SmSP1 to SmSP4 contain cysteine residues at positions 28, 44, 130, 160, 173, 184, 194, and 212 (SmSP1 protease domain numbering), which are conserved in other trypsin-like proteases. They form four disulfide bonds that can be predicted from the alignment with the crystal structures of bovine trypsin and bovine chymotrypsin (Figure 2). Moreover, the protease domain of SmSP2 through SmSP4 contains an additional cysteine residue, Cys112. By comparison with bovine chymotrypsin, this residue in SmSP2 and SmSP3 is likely to form a disulfide bond with a Cys in the N-terminal extension region (at the positions -p13 and -p9, respectively), whereas in SmSP4 a similar Cys in the N-terminal extension region is lacking (Figure 2).
SmSP5 diverges from the other four SPs in that it contains only six cysteine residues that likely form three disulfide bonds. The first two bonds, Cys28-Cys44 and Cys160-Cys173, are identical to those in trypsin, chymotrypsin and other SmSPs. The remaining cysteine residues (Cys46 and Cys72) are absent, but correspond to Cys46 and Cys77 in SmCE that were predicted to form a disulfide bond by homology modeling [46] (Figure 2). Moreover, both SmSP5 and SmCEs lack the disulfides Cys130-Cys194 and Cys184-Cys212, which are conserved in SmSP1 to SmSP4. Taken together, SmSP5 clearly differs in its disulfide pattern from the other investigated SmSPs. This close structural relationship between SmSP5 and the SmCEs is confirmed for the other analyses performed (see below). In addition, two other splice variants of SmSP5 were detected. Compared to the full-length SmSP5, both are C-terminally truncated and one is missing the crucial His residue from the catalytic triad (Figure S4).
Asp182 determines the trypsin-like specificity of serine proteases for substrates with Arg/Lys in the P1 position [47], and this residue is conserved in all of the SmSPs except SmSP5 (Figure 2), which has Gly. Therefore, it might be the case that SmSP5 displays a substrate specificity similar to that of chymotrypsin/elastase-type proteases which also contain a hydrophobic/uncharged residue in the position 182. The calcium binding site in mammalian trypsins is formed mainly by Glu70 and Glu80 (trypsin numbering, corresponding to Glu60 and Glu70 in SmSP1) [48]. This motif is not strictly conserved in the analyzed SmSP sequences; however, it might be present in a modified functional form in SmSP2, SmSP3 and SmSP4 that contain acidic residues in the close proximity of those locations (Figure 2).
Messenger RNA transcript levels for the five SmSPs were evaluated in eggs, miracidia, daughter sporocysts, cercariae, schistosomula and adults using RT- qPCR (Figure 3). For SmSP1 and SmSP3, we determined gene expression for both the protease and non-protease domains (Figure 4).
For SmSP1, the greatest expression was recorded in eggs at 2.5% of the expression level of the reference gene, SmCOX I. Low expression was recorded in adult worms, five-day old schistosomula and daughter sporocysts at around 0.1% or below relative to SmCOX I. Expression in the other stages was below significance, i.e., less than 0.01% of SmCOX I. As described above, the ORF of SmSP1 consists of 3 domains and their individual expression was evaluated by RT-qPCR and PCR (Figure 4A; Figure S2). The data show a differential expression pattern for the CUB, LDLa and protease domains of SmSP1: expression of the CUB domain is mostly in eggs and sporocysts, whereas LDLa is only expressed in eggs with an expression level about 20-fold lower than that of the protease domain (Figure 4A). As stated above, only in eggs is the whole ORF amplified by PCR suggesting that some SmSP1 is expressed as the full-length multi-domain protein (Figure S2).
Among the SmSPs, SmSP2 is the most abundantly expressed SmSP (Figure 3). In fact, expression in schistosomula and adults is on a similar level to that previously measured for the well-characterized S. mansoni cysteine and aspartic proteases [27]. In adults, SmSP2 expression is equivalent to that of SmCOX I, whereas in five-day old schistosomula expression is even greater - 150% that of SmCOX I. Significant expression, i.e., 10% that of SmCOX I, is also detected in eggs. In the other stages, expression is close to or below 1% of the SmCOX I level.
The expression pattern of SmSP3 across all life stages is similar to that of SmSP1 (Figure 3), with minor variations regarding expression in cercariae and schistosomula. Most expression is found in eggs at 2.5% of the SmCOX I expression level. Interestingly, the CUB and protease domains are only co-expressed in eggs and adults (Figure 4B), whereas differential expression is seen for the other developmental stages (Figure S2). SmSP4 is expressed predominantly in eggs (around 10% of SmCOX I level). For the other stages, approximately 1–2% of the SmCOX I level is detectable in cercariae, adults and five-day old schistosomula. Finally, SmSP5 is expressed predominantly in the eggs (2% of the level of SmCOX I) with low expression in the other life stages (0.02–0.05% of SmCOX).
The maximum likelihood analysis of a wide spectrum of vertebrate and invertebrate S1 family SPs based on amino acid sequences revealed that SmSPs clustered with related trematode proteases into five distinct and well-supported clades (Figure 5). Identical results were obtained using maximum parsimony analysis (data not shown). The clades did not create a monophyletic group. Thus, SmSP1 and SmSP3 were placed as two closely related but independent clades (trematode SP clade 1 and 3) and clustered with a large group of vertebrate SPs, including regulatory- and epithelial-derived effector trypsin-like proteases such as plasminogens, plasma kallikreins, tryptases, matriptases and transmembrane SPs (Figure 5). SmSP2 and SmSP4 also segregated into two separate but related trematode clades (numbers 2 and 4), which clustered with cestode SPs and a group of insect plasminogen-like and trans-membrane SPs (Figure 5). Finally, SmSP5 clustered with S. japonicum and Clonorchis sinensis (Chinese liver fluke) orthologs and created a sub-clade that grouped with a sub-clade of CEs within the trematode SP clade 5. This clade also clustered with chymotrypsin-like proteases from invertebrates. Accordingly, SmSP5 and its trematode orthologs associate more with the divergent schistosome CEs [22] than with other S1 family proteases [18].
S1 family SP activities in soluble extracts of S. mansoni adults, five day-old schistosomula and eggs were profiled for proteolytic specificity using peptidyl fluorogenic substrates. Two sets of specific protease substrates were used; (i) substrates with a basic amino acid residue (Arg, Lys) in the P1 position that are cleaved by trypsin-like SPs, and (ii) substrates containing bulky hydrophobic (Phe, Tyr) or aliphatic residues (Val, Leu, Met) at P1 that are cleaved by chymotrypsin- or elastase-like SPs [49]. The measured activities were further authenticated as S1 family SPs by their sensitivity to the small molecule inhibitors, Pefabloc SC and PMSF.
The results indicate that trypsin-like activities predominate over chymotrypsin/elastase-like activities in the analyzed extracts (Figure 6). The trypsin substrates were hydrolyzed with variable efficiencies giving distinct cleavage patterns for the individual life stages. The prominent activity in all extracts was best measured with the Boc-L-R-R-AMC substrate, hence making this substrate a useful probe to detect and measure SmSPs. Extracts of eggs displayed a particularly complex profile by cleaving an additional two substrates, Bz-F-V-R-AMC, and Z-G-P-R-AMC. This suggests that this life-stage possesses additional, possibly stage-specific, trypsin-like proteases. In contrast to the major trypsin-like activities, chymotrypsin/elastase-like activity was relatively weak being measured only in schistosomula and adults.
Subsequently, we tested whether SmSPs is measurable in the E/S products from eggs, schistosomula and adults. For this purpose, we used the substrate Boc-L-R-R-AMC, which was identified as the most efficient substrate for homogenates of all the life stages (Figure 6). The specific activities of the E/S products, which were inhibited by the SP inhibitors, Pefabloc SC and PMSF, were 1.05±0.10, 1.38±0.05, and 0.11±0.01 RFU/µg protein for eggs, schistosomula and adults, respectively.
A spatial homology model of the protease domain of SmSP1 was constructed to analyze its binding pocket and substrate specificity. The X-ray structure of bovine trypsin in complex with the small-molecule inhibitor, leupeptin (PDB code 1jrt), was used as a template. We used SmSP1 as representative of SmSP1 to SmSP4, which have substantial sequence identity, a similar disulfide pattern and homology in active site regions (Figures 2 and S3). Figure S5 shows that the SmSP1 protease domain displays the conserved architecture of S1 family proteases which consists of two six-stranded β-barrel domains packed against each other. The catalytic amino acid residues are located at the junction between the domains. The major insertion/deletion variations between SmSP1 to SmSP4 (such as the SmSP2 insertion at residue 140, Figure 2) are located at surface-exposed loops.
The primary substrate specificity determinant of S1 family proteases is the S1 binding subsite. In SmSP1, this subsite forms a deep and narrow negatively charged pocket that contains Asp182 at the bottom (Figures 7A and 7B). Leupeptin, the transition state analog protease inhibitor, was docked into the active site of SmSP1. The arginal residue of leupeptin forms a covalent linkage with the catalytic Ser188, a salt bridge with Asp182 in the S1 subsite and hydrogen bonds with the carbonyl oxygen of Ala183 and Asp211 (Figure 7C). An additional hydrogen bond is formed between the side chain nitrogen of Gln185 and the carbonyl oxygen Leu2 residue of leupeptin. The putative interaction pattern of leupeptin at the S1 subsite of SmSP1 is similar to that found in bovine trypsin [50]. This demonstrates that SmSP1 has a substrate binding preference for basic residues at the P1 position, the positive charge of which compliments the negatively charged Asp182, i.e., trypsin-like activity. This conclusion can be generalized to SmSP2 to SmSP4 which also contain the critical Asp182 residue.
Much has been reported on the genetic, biochemical and functional characterization of cysteine and aspartic protease activities in schistosomes [16], [17] and flatworms in general [16], [51], and of the schistosome CE SPs [20] that putatively facilitate parasite invasion of the mammalian host [18]–[20]. By comparison, relatively little detail is available for non-CE SPs. There are, however, indications that non-CE S1 family SPs contribute to successful infection [6]. Thus, kallikrein-like protease activity from S. mansoni adults [12] and plasmin-like fibrinolytic activity from S. mansoni eggs [15] have been recorded previously. Both activities displayed trypsin type cleavage specificities and both may contribute to the phenomenon, whereby large occlusions of veins by schistosomes are not associated with intravascular deposition of fibrin and thrombus formation [52]–[54]. At the gene and primary sequence levels, however, only two SmSPs, namely SmSP1 [13], [14] and another [23], [24], which we term SmSP2, have been described.
The S. mansoni GeneDB currently contains 16 unique sequences that belong to Clan PA family S1 SPs. This number is significantly lower than the 135 family S1 proteases found in the human genome [8], [25] and may be due to the lack of need to regulate the more complex and expanded physiological processes found in vertebrates [55]. In our study and apart from SmSP1 [13], [14], we identified four additional SmSP genes encoding typical sequence features of the S1 family [7], [8] and which we term SmSP2 through SmSP5. Two further genes (Smp_194090 and Smp_06530 in GeneDB) were identified in the S. mansoni GeneDB as putative proteolytically inactive SmSPs as they lack the catalytic serine or histidine residue in the catalytic triad. The remaining nine of the 16 family S1 SPs comprise eight CEs (encoding both putative proteolytically active and inactive products) and a gene (Smp_174530) that encodes an S1 family SP ORF fused downstream of an M01 family metallo-protease. This protease that was not known to us at the beginning of our study and because of its domain complexity and sequence size was not described further.
Our phylogenetic analyses of trematode SPs displayed interesting evolutionary trends. The SmSPs segregate into five clusters of family S1 proteases. The protease domains of SmSP1 and SmSP3, forming clades 1 and 3, respectively, cluster with a large group of vertebrate trypsin-like SPs including regulatory and effector epithelial-derived proteases. In addition to a protease domain, the ORFs for SmSP1 and SmSP3 include non-catalytic CUB domains and SmSP1 LDLa domain. Several vertebrate matriptases that also contain CUB domains are present in our phylogenetic analysis including those belonging to the ‘suppressor of tumorigenicity’ group. As judged by the domain organization, SmSP1 resembles mammalian matriptases (a.k.a. epithin, MT-SP); however unlike conventional matriptases with multiple CUB and LDLa domains, SmSP1 has only one of each. CUB domains were first described in the complement proteins C1r and C1s and are modules of approximately 110 amino acids with four conserved cysteine residues [56]. These domains mediate protein-protein interactions and are generally associated with proteins that have diverse, usually regulatory, functions in the extracellular space and/or plasma membrane [56]. CUB domains can also interact with heparin and glycoproteins [56] and are often associated with metallo-proteases, in addition to serine proteases [8].
Based on the RT-qPCR analysis, the complete ORFs of SmSP1 and SmSP3 molecules share a similar expression profile (quantitatively and, to a smaller degree, qualitatively) across the developmental stages tested. However, it is also clear that the individual protease, CUB and/or LDLa domains are differentially expressed across the developmental stages tested being only co-expressed in eggs and, for SmSP3, adults. The particular functions of these enzymes and their component domains are unknown and their importance to parasite vitality and/or survival might be tested via specific RNA interference (RNAi), which has been shown to operate in schistosomes [30], [57], [58]. According to our phylogenetic analysis, the closest vertebrate orthologs to SmSP1 and SmSP3 are those associated with regulatory cascades such as fibrinolysis and vasodilation. This, together with the fact that SmSP1 was detected apparently on the surface area of worms and secreted into the cultivation media [13], suggests a possible function at the host-parasite interface.
The presence in the ORF of SmSP1 of an LDLa domain (positioned between the CUB and catalytic domains) deserves a note. Schistosomes and other flatworms do not synthesize cholesterol (found within LDL) and must therefore scavenge it from the environment, particularly for the energy-intensive work of producing eggs [59], [60]. There is also a report that the presence of S. mansoni eggs is connected with decreased circulating levels of cholesterol in the host [61], however, we can only speculate about the real function of the SmSP1 LDLa domain.
SmSP2 and SmSP4 form two other separate clades and cluster with trypsin SPs from insect and other invertebrates. Both proteases are characterized by their longer but different N-terminal extensions that lack homologies to known proteins but which are shared in orthologous SPs from S. japonicum [44] and C. sinensis [62]. Functions as yet are unknown, however, it is certainly remarkable that SmSP2 is massively expressed in schistosomula and adults (150% and 60% of SmCOX I expression levels, respectively) and, therefore, conceivably contributes significantly to host and/or parasite protein hydrolysis, perhaps in modulating of host physiologic processes [6], [12]. The presence also of close orthologs of SmSP2 in Fasciola gigantica [63] and C. sinensis [62] suggests a general role for SP2 during infection in the mammalian host. The impressive expression levels for SmSP2 are consistent with high levels of SmSP2 expression from microarray [23] and transcriptome data [24]. Also, the expression levels are close to those for the gut-associated, digestive cysteine and aspartic proteases, SmCB1 and SmCD, respectively, for which expression is close to that of SmCOX I [27].
Finally, for SmSP5, phylogenetic analysis identified its relative position in what we term clade number 5. This clade is most closely related to chymotrypsins from invertebrates and comprises SP5 orthologs in S. japonicum [44] and C. sinensis [62], and the CE genes in S. mansoni, S. haematobium [20], [22], S. japonicum [44] and Schistosomatium douthitti [20]. Clade 5 is particularly significant for phylogenetic relationship studies of schistosome proteolytic enzymes as it contains sequences that bridge the outlier CE group and other ‘more typical’ S1 family SPs. Specifically, our previous phylogenetic work [18] had highlighted that the CEs coalesce as a tight group that diverges from other family S1 protease sequences. At that time the SmSP5 sequence was incomplete and not amenable to analysis [18]. The current sequence analysis suggests that SmSP5 and its trematode orthologs are ‘a missing link’ between the outlier CE group and the common ancestor. CEs initially evolved from chymotrypsin regulatory proteases and may provide an evolutionary advantage in contributing to host invasion [22].
For the SmSP protease domains, we investigated the structure-function relationships using primary structure analysis, homology modeling and protease activity profiling with peptidyl substrates. The sequence alignment shows that all the SmSPs except SmSP5 share a conserved Asp182 residue. This residue defines the specificity for the S1 binding site and drives a strong preference for Arg and Lys residues at the P1 position of protein/peptide substrates, as demonstrated for vertebrate trypsins [47]. The homology model of SmSP1 reveals that the S1 pocket with its critical Asp182 residue has an architecture analogous to vertebrate trypsins. In contrast, the S1 binding pocket of SmSP5 has a Gly182. Also, SmSP5 lacks the disulfide Cys184-Cys212 which is present in the other four SmSPs and known to stabilize the S1 binding site in vertebrate trypsins. Interestingly, this disulfide is also absent in schistosome CEs, which contain non-polar residues (Ile or Leu) at the bottom of the S1 binding pocket resulting in elastase and chymotrypsin-like activities [22].
Consistent with the number of trypsin-like sequences in all of the life-stages studied, major trypsin-like activities could also be measured using peptidyl fluorogenic substrates in eggs, schistosomula and adult extracts. Eggs, in particular, presented the most diverse and active profile compared to adults and schistosomula suggesting they express more than one highly active SP. Schistosomula, in contrast, displayed an activity profile restricted to one substrate, and one might suppose that this activity is in fact due to SmSP2 which was, expressed at higher levels than other SPs as measured by RT-qPCR (see above). Finally, the finding that significant trypsin-like activity was found in the E/S products of the three life stages tested suggests that one or more SmSPs are secreted into the host environment where they may interfere with relevant proteolytic cascades such as blood coagulation, complement or blood pressure regulation [6], [12].
In contrast to the trypsin-like activities measured, chymotrypsin/elastase-like activities were absent in eggs, and in schistosomula were at least one order of magnitude weaker. It is possible that the activity in schistosomula is, in whole or part, due to residual CE activity carried forward after mechanical transformation of cercariae and in vitro culture of schistosomula. In adults, however, this possibility seems remote and the minor activities measured may be contributed to by SP5.
To conclude, the present study provides a comprehensive phylogenetic, transcriptomic and functional framework illustrating the heretofore unknown complexity of schistosome S1 family SPs, other than the well-studied CEs [20], [22]. The individual enzymes underlying the activities measured remain ‘undiscovered country’ both in terms of their functional characterization and, not least, their possible contributions to successful parasitism by the schistosome, including at the host-parasite interface.
|
10.1371/journal.ppat.1006191 | Selective expansion of high functional avidity memory CD8 T cell clonotypes during hepatitis C virus reinfection and clearance | The dynamics of the memory CD8 T cell receptor (TCR) repertoire upon virus re-exposure and factors governing the selection of TCR clonotypes conferring protective immunity in real life settings are poorly understood. Here, we examined the dynamics and functionality of the virus-specific memory CD8 TCR repertoire before, during and after hepatitis C virus (HCV) reinfection in patients who spontaneously resolved two consecutive infections (SR/SR) and patients who resolved a primary but failed to clear a subsequent infection (SR/CI). The TCR repertoire was narrower prior to reinfection in the SR/SR group as compared to the SR/CI group and became more focused upon reinfection. CD8 T cell clonotypes expanding upon re-exposure and associated with protection from viral persistence were recruited from the memory T cell pool. Individual CD8 T cell lines generated from the SR/SR group exhibited higher functional avidity and polyfunctionality as compared to cell lines from the SR/CI group. Our results suggest that protection from viral persistence upon HCV reinfection is associated with focusing of the HCV-specific CD8 memory T cell repertoire from which established cell lines showed high functional avidity. These findings are applicable to vaccination strategies aiming at shaping the protective human T cell repertoire.
| In this study we examined the diversity and dynamics of the repertoire of receptors of CD8 T cells that are selected and enriched upon real-life multiple exposures to viral infections. Using hepatitis C virus (HCV) infection in a cohort of high risk people who inject drugs, we demonstrate that protection upon two subsequent infections was associated with a narrow repertoire of virus-specific CD8 T cells and selective expansion of cells with high polyfunctionality (increased TNFα production and cytotoxic potential). Our results have important implications in vaccination programs aiming at shaping the CD8 T cell repertoire against viral infections and cancers.
| The capacity of CD8 T cells to recognize and respond to various pathogen-derived antigens is dictated by the diversity of their T cell receptor (TCR) repertoire. The TCR is a heterodimer of two chains, α and β, that comprise constant and variable regions. The most variable region in both chains is generated by somatic recombination involving variable (V), junction (J) and diversity (D) gene segments that could theoretically generate ~1015–20 unique TCRs or T-cell clonotypes capable of recognizing peptide-MHC (pMHC) complexes [1]. Positive and negative selection in the thymus leaves ~ 2x107 T-cell clonotypes with unique TCR amino acid sequences that constitute the naïve human T-cell repertoire. Hypervariable complementarity-determining regions 1 and 2 (CDR1 and CDR2) are formed by the germline V region sequences and interact mainly with MHC. The CDR3 of the TCR α and β chains, the most variable region of the TCR, are encoded by the V(D)J junction and interact primarily with peptide, thus determining antigenic specificity of the TCR [1]. Upon exposure to a viral infection, particular clonotypes recognizing virus-derived epitopes/pMHC are selected and expand into primary effectors that then contract to form a pool of long-lived memory T cells that are able to respond rapidly upon virus re-exposure. The size and diversity of the expanding effector T cell repertoire can vary according to the initial germline repertoire of naïve CD8 T cells and strength of interaction with pMHC (affinity and avidity). In contrast, factors governing the size and diversity of the antiviral memory CD8 T cell pool are not well understood. Most importantly, determinants for selection and maintenance of CD8 T cell clonotypes exerting an efficacious and protective anti-viral immune response upon virus re-exposure or reactivation remain elusive. There is evidence to suggest that the memory CD8 T cell repertoire can be modulated by heterologous infections and age [2, 3] but other host and viral factors could be involved.
One key characteristic of the virus-specific TCR repertoire is diversity, which defines the number of unique clonotypes forming the repertoire. The repertoire can be characterized as «narrow» or «broad» depending on the number of unique clonotypes it contains. Distinctive molecular properties of T cell clonotypes that determine their functionality include affinity, avidity, functional avidity and flexibility. Affinity describes the strength of binding of a single TCR to cognate pMCH complexes, whereas avidity (structural avidity) is the sum of binding affinities of multiple TCRs to their pMHC complexes. Functional avidity depends on how this translates into measurable biological functions such as cytokine production [4]. Flexibility is the capacity to recognize multiple variants of the same epitope and cross-react with these variants. Although TCRs are generally very specific or «private» in their response to a pMHC complex, identical TCR sequence usage in response to a specific epitope across multiple individuals termed «public TCRs» were observed in a number of infections, tumors and even autoimmune conditions [5, 6].
Studies in chronic cytomegalovirus (CMV) and Epstein-Barr virus (EBV) infections demonstrated focusing and increased affinities of the virus-specific CD8 TCR repertoire [7]. In human immunodeficiency virus (HIV) infection specific clonotypes dominated the response against an epitope in the p24 Gag (KK10; residues 263–272) restricted by HLA B*2705 in patients who controlled viral replication. These clonotypes exhibited higher avidity and polyfunctionality and superior control of HIV replication in vitro [8–10]. Furthermore, certain public clonotypes were detected in several HIV controllers [11, 12]. However, another study examining different epitopes did not observe preferential use of particular clonotypes [13]. Escape mutations in targeted epitopes could be recognized by some of these highly functional clonotypes [10] yet they also drove the expansion of alternate clonotypes with dual reactivity against both the original and mutated epitopes [14]. This expansion was inversely correlated with residual viral load, suggesting that these alternative clonotypes play a role in limiting replication of mutated viruses [15].
Hepatitis C virus (HCV) infection represents a unique model of a human viral infection with dichotomous outcomes, i.e. spontaneous clearance (~30% of infected individuals) and persistent infection [16]. Control of primary HCV infection in the chimpanzee model was associated with a more diverse CD8 TCR repertoire than infections that became chronic and were associated with escape mutations [17]. In humans, selection of high-avidity CD8 T cells correlated with control of primary infection [18]. Analysis in individuals with long-term clearance or persistent infection demonstrated a biased TCR repertoire towards a common core of public TCRs irrespective of infectious outcome [19]. HCV is also thought to exploit a «hole» in the T cell repertoire to undergo escape mutation and evade detection by the immune system [20]. The dynamics of the repertoire upon re-exposure to HCV are less defined. Rechallenging chimpanzees following clearance of primary infection demonstrated that resolution temporally coincided with the expansion of dominant clonotypes that were associated with clearance of primary infection [21]. This is consistent with data in the lymphocytic choriomeningitis virus (LCMV) model [22]. Higher CDR3 diversity correlated with viral clearance and better control of escape mutations within the targeted epitope and temporary narrowing of the repertoire was observed at the peak of the recall response [17].
The majority of acute HCV infections in North America occur among people who inject drugs (PWID). Individuals who spontaneously clear their primary infection but maintain risk behaviors associated with HCV re-exposure remain at high risk of reinfection. As such, HCV represents a unique model to study correlates of protective immunity in a real life challenge experiment. Follow-up of PWIDs through consecutive episodes of infection and reinfection can provide an insight into the dynamics of the virus-specific TCR repertoire and the functional properties associated with protective immunity. We have previously demonstrated that protection from chronicity upon reinfection with HCV correlated with expansion of polyfunctional HCV-specific effector T cells and increased breadth of T cell responses, suggesting the generation of de novo responses. In contrast, viral persistence was associated with limited expansion of virus-specific T cells and infection with variant viral strains that were not recognized by preexisting virus-specific memory T cells [23].
Here, we examined dynamics of the TCR repertoire during reinfection. Our objectives were to distinguish the role of pre-existing memory versus de novo T cell responses during reinfection and to establish the functional correlates of CD8 T cell clonotypes associated with protective immunity in real life exposure and reinfection. Our results suggest that dominant HCV-specific tetramer+ CD8 T cell clonotypes mobilized during the reinfection episode were exclusively recruited from the preexisting memory pool. Protection from chronic infection was associated with a narrower TCR repertoire that became more focused upon reinfection with preferential selection of TCR clonotypes with high functional avidity and polyfunctionality.
To examine the evolution and dynamics of the HCV-specific memory CD8 T cell repertoire upon virus re-exposure, we performed longitudinal analysis of this repertoire on virus-specific CD8 T cells during HCV reinfection episode in five patients with different outcomes (S1 Table). Three patients resolved two successive HCV infections, hereinafter termed SR/SR. Two patients failed to spontaneously resolve the reinfection episode and developed chronic viremia despite clearing a previous HCV infection, hereinafter termed SR/CI.
Epitope-specific CD8+ T cells identified by three different MHC class I tetramers were sorted and the TCR repertoire sequenced at three distinct time points for each patient: i) pre-reinfection (range: -55 to -20 weeks before detection of reinfection), ii) peak reinfection (range: 3 to 8 weeks post detection of reinfection), and iii) late/post reinfection follow-up (range: 12 to 24 weeks post detection of reinfection). The peak reinfection time point was selected based on our previously published longitudinal analysis of the frequency of tetramer+ CD8 T cells [23]. In addition, we examined the repertoire during primary acute infection in patient SR/SR-3, for whom samples were available early at week 3 post primary infection. In patients SR/SR-1 and SR/SR-3, tetramer frequency was high enough during peak reinfection and primary infection, respectively which allowed us to sort effector (CD127-) and memory (CD127+) HCV-specific CD8+ T cells as well. As a control, total naïve CD8+ T cells from patients SR/SR-1 and SR/CI-2 defined as CD8+CD45RO- were sorted and sequenced for all three time points tested. A summary of the time points tested for each patient and the corresponding tetramer frequencies is presented in S1 Table. The gating strategy and post-sorting purity are presented in S1 and S2 Figs. A summary of the sequencing information including the number of cells, sequences and clonality is presented in S2 Table. TCRBV and TCRBJ gene usage, as well as the CDR3 amino acid sequence of all TCR β chains expressed by the different HCV-specific T cell clonotypes were analyzed (S3–S7 Tables). The top 10 clonotypes for each patient and time point are presented in Fig 1. Longitudinal analysis of the TCR β chain dynamics demonstrated a striking difference between SR/SR and SR/CI patients where there was high level expansion (up to 3.5 fold) of specific clonotypes in two out of three SR/SR patients upon reinfection but limited or no expansion in patient SR/SR-3 and both SR/CI patients. This is consistent with the limited expansion of tetramer+ cells observed in the SR/CI (S1 Table and [23]). The analysis also demonstrated that the dominant (frequency > 1%) and sub-dominant (frequency > 0.5%) clonotypes at the peak reinfection time point were recruited from the pre-existing memory CD8 T cell populations and no new clonotypes were detected. This was also true in patient SR/SR-1 where we managed to examine the repertoire of effector (CD127-) CD8 T cells that could be indicative of a de novo T cell response. Comparison of the effector (CD127-) and memory (CD127+) CD8 T cell repertoire showed the same clonotypes albeit at slightly different frequencies (Fig 1 and S3–S7 Tables). Although, some new clonotypes that were not present at pre-reinfection were detected in all patients at the peak time point as shown in the Venn diagrams in S3 Fig, they were of low frequencies (< 0.5%). Similarly, comparison of the repertoire during primary acute infection and reinfection in patient SR/SR-3 demonstrated that the repertoire generated during primary infection was relatively stable and a precursor to the memory repertoire (S4 Fig).
Although the clonotype composition of the T cell repertoire remained relatively unchanged upon reinfection, there was a change in the dominance and hierarchy of the different clonotypes where a more limited number of clonotypes dominated the response at the peak of reinfection. For example, patient SR/SR-1 exhibited preferential expansion of the V20 clonotype. This is even more evident in patient SR/SR-2 where the two most dominant clonotypes (V19.1 and V11.2) switched hierarchies upon reinfection, suggesting a selection and amplification of particular clonotypes during reinfection.
Our analysis demonstrated no overlap and no common clonotypes within the dominant and sub-dominant Vβ-CDR3-Jβ detected in the tested patients. This lack of common repertoire was observed among patients targeting the same epitope whether they belonged to different groups (SR/SR-1 and SR/CI-2 recognizing the HLA-A2 restricted NS3-1073 epitope) (Fig 1 and S3 and S6 Tables) or the same group (SR/SR-2 and SR/SR-3 recognizing the HLA-B27 restricted NS5B-2841 epitope) (S4 and S5 Tables).
Altogether, these results suggest that the dominant protective CD8 T cell clonotypes expanding at the peak of reinfection episode were recruited from the preexisting memory pool and at least within this set of individuals, no common or public TCRs were detectable. Very limited expansion was observed in HCV-specific CD8 T cells and specific clonotypes in the SR/CI group.
Data from the chimpanzee model suggested that a more diverse repertoire, i.e. the presence of a larger number of unique clonotypes recognizing the same pMHC, was associated with spontaneous resolution of primary acute HCV [17]. Hence, we proceeded to examine repertoire diversity within our study subjects in a memory response. Clonotypes were classified into 4 categories according to their frequencies within the repertoire: (i) dominant clonotypes (frequencies >1%); (ii) sub-dominant clonotypes (0.5–0.99%); (iii) low abundance clonotypes (0.1–0.49%); and (iv) lowest abundance clonotypes (<0.1%). Our analysis demonstrated that the clonotypic profile in SR/SR patients was less diverse than that observed in SR/CI patients at the pre-reinfection time point (Figs 2 and 3). In the SR/SR group, 75–83% of the repertoire was contributed by 14–23 TCR clonotypes. This repertoire became more focused at peak reinfection for patients SR/SR-1 and SR/SR-2, where 92% of the repertoire of tetramer+ CD8+ CD127- cells in SR/SR-1 and 96% of the repertoire of tetramer+ CD8+ T cells in patient SR/SR-2 were contributed by 15 and 2 clonotypes, respectively (Fig 2A and 2B). In these two patients, the repertoire retained a more focused status post reinfection when compared to the pre-reinfection time point, with 77% and 93% of the repertoire contributed by 16 and 4 clonotypes, respectively (Fig 2A and 2B). Patient SR/SR-3, for whom samples were available during primary infection, also demonstrated a focused repertoire of effector (CD127-) and memory (CD127+) HCV-specific T cells early during acute primary infection where 85% and 82% of the repertoire were represented by 27 and 26 clonotypes, respectively (Fig 2C). This repertoire remained more or less stable during the memory phase where 83% of the repertoire was represented by 21 clonotypes but did not sensibly change upon reinfection. This is consistent with our previous results, where no significant expansion of the tetramer+ population was observed in this patient upon reinfection (S1 Table), and IFNγ Enzyme-Linked Immunospot (ELISPOT) assays suggested that he had been reinfected with a different HCV subtype [23].
In contrast, the repertoire in the two SR/CI (unprotected) patients was more diverse at the preinfection time point. Dominant clonotypes represented 37% and 55% of the repertoire (18 and 23 TCR clonotypes, respectively). At the peak of the immune response during reinfection, the repertoire remained more diverse than in the SR/SR patients (the dominant category represented 28–64% of the repertoire) (Fig 3A and 3B). The remaining 40–60% of the repertoire was represented by at least 133 and as much as 568 clonotypes. In contrast, the number of minor clonotypes in the SR/SR patients was much lower (Fig 2). Furthermore, we did not observe increased focusing of the repertoire in SR/CI patients at the late reinfection time point.
The presence of a less diverse repertoire that becomes more focused upon reinfection in the SR/SR group was further supported by the observation that the top three clonotypes in patients SR/SR-1 and the top two clonotypes in patient SR/SR-2 represented 28% and 59% of the repertoire pre-reinfection, respectively (Fig 2A and 2B, dissected pie-charts in the lower rows). Upon reinfection, these clonotypes became more dominant, representing 51% and 96% of the repertoire, respectively. This increased dominance was maintained post-reinfection, where they formed 39% and 90% of the repertoire, respectively. On the other hand, the top three clonotypes in SR/CI patients never represented more than 27% of the repertoire, and were sometimes as low as 9% (Fig 3A and 3B, dissected pie-charts in the lower rows). These results confirm that the SR/SR TCR repertoire is highly focused and becomes more focused upon re-exposure and reinfection.
In order to establish a quantitative measure of the changes observed in the TCR repertoire during reinfection we examined diversity using the Simpson diversity index that takes into account the number of species present as well as the abundance of each species, with 0 defined as infinite diversity and 1 as no diversity [24, 25]. The focusing of the repertoire in the SR/SR group was reflected by an increase in the Simpson diversity index for patients SR/SR-1 and SR/SR-2 indicating a less diverse repertoire but it remained stable or decreased in patient SR/SR-3 and in the SR/CI group where no expansion was observed (Fig 4A and 4B). Next, we examined richness of the repertoire. This parameter measures the number of clonotypes per sample [26]. We observed decreased richness index in the SR/SR group (mostly SR/SR-1 and SR/SR-2) indicating focusing of the repertoire while the same index remained stable in patient SR/SR-3 and slightly increased in the SR/CI group (Fig 4C and 4D). Next, we examined evenness of the repertoire. This parameter measures the relative abundance of the different clonotypes [26]. Again, we observed decreased evenness at peak reinfection in patients SR/SR-1 and SR/SR-2 indicating focusing of the repertoire (Fig 4E and 4F) while the same index slightly increased in patient SR/SR-3 and in the SR/CI group. Finally, we used the Morisita Horn index to compare the repertoires from pre-reinfection and peak reinfection, as well as from peak reinfection and late reinfection. This index is indicative of the overlap between two samples as it takes into account both the number of clonotypes and their frequencies within the repertoire [27]. An index of 1 represents complete overlap or identical repertoire and an index close to 0 represents two very different repertoires with almost no clonotypes shared between the samples. The analysis for SR/SR patients showed a Morisita Horn index around 0.6 when comparing the pre-reinfection and peak reinfection time points, indicating changes in the repertoire between these two time points. The repertoire was stable between peak reinfection and late reinfection as indicated by a Morisita Horn index around 0.9. We observed the opposite trend for SR/CI patients, where the repertoire underwent more changes between the peak reinfection and late reinfection (Morisita Horn index of 0.6) time points as compared to the changes between pre-reinfection and peak reinfection (Morisita Horn index of 0.8) suggesting increased diversification with establishment of chronic infection. Collectively, all four measures demonstrate focusing of the repertoire in the SR/SR group and unchanged or even slight increase in diversity in the SR/CI group.
Next, we examined CDR3 amino acid (aa) length distribution. As a reference of CDR3 lengths distribution from an unselected repertoire, we analyzed TCR sequences of naive CD8 T cells sorted from patient SR/CI-2 (Fig 5C). As shown by the bell-shaped curve, the naive repertoire displayed a normal Gaussian distribution. The distribution was identical in the naive compartment in the three time points tested (preinfection, peak and follow-up). In contrast, this normal distribution was altered in HCV-specific T cells from all patients analyzed (Fig 5A and 5B), reflecting the antigen-specific selected population. Furthermore, this analysis showed an expected bias towards the CDR3 lengths of the most dominant clonotypes. For example, the CDR3 length distribution for patient SR/SR-2 was highly biased towards a length of 14 aa, which reflects the fact that the 2 highly-dominant clonotypes (representing 78–96% of the repertoire at the different time points) possess CDR3 of that length (Fig 5A, middle). In addition, for patient SR/SR-1 the CDR3 length distribution was biased towards CDR3s with a length of 13 aa (clonotype TCRVB20-TCRVJ01.01, with frequencies of 8–35% of the repertoire, Fig 5A, left). Another important observation from this analysis is that, in the SR/SR group, we could clearly see an increase in the dominance of CDR3 of certain length between the pre-reinfection time point and the peak reinfection time point (13 aa in SR/SR-1 and 14 aa in SR/SR-2). This is in agreement with the focusing of the repertoire in those patients. In the SR/CI group (Fig 5B), we also observed a bias towards the length of the most abundant clonotypes (13 and 17 aa for SR/CI-2 and 15 aa for SR/CI-3, Fig 5B) but the CDR3 length distribution remained similar across the different time points. This reflected our earlier observation of limited to no focusing of the repertoire during reinfection in these patients. Next, we compared average CDR3 lengths, nucleotides (NT) addition and Germline index measuring junctional diversity for all patients as described by Yu et al [28], but there was no clear difference between groups or across time points (S5 Fig) further confirming our earlier observation of limited novel diversification in the repertoire upon reinfection.
Studies in HIV infection have demonstrated that TCR avidity towards pMHC correlated with CD8 T cell functionality, better control of viral replication and overall lower viral loads [8, 29, 30]. So we sought to evaluate whether selection and expansion of specific clonotypes during HCV reinfection was associated with higher affinity or functional avidity that would endow them with a superior protective capacity. To establish individual cell lines specific to the HLA A2 restricted NS3 epitope (A2/NS3-1073; CINGVCWTV), HCV tetramer+CD8+ T cell were sorted and cultured under limiting dilution. Cell lines were generated from two patients: SR/SR-1 (69 cell lines generated) and SR/CI-2 (36 cell lines generated). Five cell lines from each patient were selected for detailed analysis termed hereinafter cell lines R1-R5 and cell lines C1-C5 generated from patients SR/SR-1 and SR/CI-2, respectively. We first evaluated the avidity of individual cell lines, a parameter that measures both binding strength and the total number of interactions on the T cell surface using a tetramer dilution assay. Both the percentage of tetramer positive cells and the mean fluorescence intensity (MFI) were equivalent for all cell lines regardless if they originated from a SR/SR or a SR/CI patient (Fig 6).
Next, we evaluated the functionality of individual cell lines or micropopulations in response to stimulation with the cognate NS3-1073 peptide in an intracellular cytokine staining (ICS) assay. We evaluated the surface expression of the degranulation marker CD107a and intracellular expression of TNFα, IFNγ and IL-2 (Fig 7A–7D). Representative intracellular cytokine staining (ICS) data are presented in S6 Fig. As shown in Fig 7A and 7B, CD8 T cell lines generated from the SR/SR patient were more sensitive to lower peptide concentrations and showed enhanced CD107a expression (average EC50 SR/SR = 1.8x10-7; SR/CI:5.5x10-6) and TNFα production (average EC50 SR/SR = 1.5x10-7; SR/CI:3.5x10-5). The capacity to produce IFNγ and IL-2 was clone dependent, and no clear difference was observed between cell lines from the SR/SR versus cell lines from the SR/CI patient (Fig 7C and 7D). In general, production of IL-2 was low (Fig 7D).
The polyfunctionality of CD8 T cells measured as the capacity of single cells to produce multiple functions is a key determinant of spontaneous resolution of HCV infection [32, 33]. Boolean gating and pie chart analysis demonstrated that cell lines generated from the SR/SR-1 patient had a higher degree of polyfunctionality compared to cell lines generated from the SR/CI-2 patient (Fig 7E and Supplementary S6B and S6C Fig). At maximum peptide concentration, we observed that an average of 85% of the cells displayed at least two functions for cell lines generated from SR/SR-1, compared to 30% of the cells for cell lines generated from SR/CI-2 (Fig 7E and S6B and S6C Fig). Furthermore, at limited peptide concentration (0.01μg/ml), an average of 50% of the cells were positive for at least one function for SR/SR-1 cell lines compared to an average of 13% for SR/CI-2 cell lines. Polyfunctionality index was calculated as previously described [31] to compare the overall polyfunctionality of all cell lines. As shown in Fig 7F, the degree of polyfunctionality was higher for all SR/SR-1 cell lines compared to SR/CI-2 cell lines. These results suggest that secondary clearance upon reinfection is associated with the presence of polyfunctional CD8 T cell clonotypes.
To determine whether the functionality of the individual cell lines reflected preferential expansion in vivo during reinfection, we have sequenced the TCR of all 10 cell lines tested from patient SR/SR-1 (cell lines R1-R5) and SR/CI-2 (cell lines C1-C5). As demonstrated in S8 Table, all cell lines isolated from patient SR/SR-1 with the exception of clone R3 carried TCR clonotypes that were detectable at frequencies ranging from 0.7% to 33% at the peak of the immune response during reinfection although some of them were micropopulations composed of two different clonotypes. Nevertheless, cell line R1, one of the best responding cell lines, was partially (7%) composed of TRBV-20 that showed 33% expansion during peak reinfection. Similarly, cell line R5, with the highest TNF-α production was partially (12.5%) composed of TRBV27-01*01 that showed 8% expansion during peak reinfection. Altogether, these data demonstrate that secondary clearance upon HCV re-exposure and reinfection is associated with selective expansion of high functional avidity CD8 T cell clonotypes.
Defining the correlates of protective immunity at the clonotypic level is essential for fine-tuning the design of prophylactic vaccines against chronic viruses with highly variable sequences such as HIV and HCV. This study provides an insight into the dynamics of the virus-specific CD8 T cell repertoire during HCV re-exposure and reinfection in a real-life setting. Our results demonstrate that protective immunity and rapid virus clearance upon reinfection was associated with expansion of a limited number of polyfunctional CD8 T cell clonotypes selected from the memory pool. These CD8 T cells displayed a focused repertoire and cell lines established from one SR/SR patient were characterized by high functional avidity and polyfunctionality. In contrast, very little expansion of HCV specific CD8 T cell was observed in SR/CI patients who developed persistent viremia upon reinfection and their repertoire was more diverse. Cell lines established from one SR/CI patient displayed reduced functional avidity demonstrated by a much weaker production of cytokines and lower cytotoxic potential that could potentially have facilitated virus persistence and chronicity. Our observations with the cell lines reflected the ex vivo characterization of the functionality of HCV-specific CD8 T cells in these patients [23].
Our longitudinal analysis of the TCR repertoire, using three different HCV tetramers, showed conserved clonotype usage within the same individual where the dominant and sub-dominant clonotypes forming the effector population at the peak of the immune response during reinfection were recruited from the pre-existing memory T cell pool generated following clearance of the primary infection. New clonotypes were only detected at low abundance frequencies. Two out of three patients in the SR/SR group had dominant clonotypes following primary infection and the same clonotypes expanded upon reinfection with a homologous variant, suggesting that these particular clonotypes played an important role in viral clearance. The third SR/SR patient displayed a focused repertoire before reinfection but showed limited expansion during reinfection. We have previously shown that this patient was reinfected with a different HCV subtype (genotype 1b after primary infection with genotype 1a) which could affect his immune response. In contrast, there was very limited expansion of HCV specific CD8 T cells in the SR/CI group, even despite reinfection with the same HCV subtype for patient SR/CI-2 (Fig 1). It is noteworthy that our previous analysis of autologous virus sequences demonstrated that the SR/CI patients were infected with variant viruses that were not recognized by pre-existing memory T cells [23]. It was not possible to sequence virus from the primary infection in those patients, preventing a clear comparison between both infection episodes. That potential mismatch between the two infecting viral strains could be responsible for the lack of significant expansion of the pre-existing memory T cells or the generation of new clonotypes. Our results are in concordance with data from the LCMV model demonstrating that the TCR repertoire of the primary epitope-specific CD8 T cell response was conserved in the memory pool, and that after a secondary effector recall response, there was 60–100% identity between the clonotypes of the primary effector, memory and recall responses [22, 34]. Rapid resolution of HCV infection upon rechallenge in chimpanzees also coincided with the expansion of T cell clonotypes that dominated the memory CD8 T cell pool [17, 21].
Cross reactivity between NS3-1073–specific CD8 T cells and an influenza virus epitope (NA-231) was previously reported [35, 36]. In addition, CD8 T cells reactive to this epitope could be amplified from a significant proportion of healthy individuals with no prior exposure to HCV [35, 36]. Thus, it is possible that this may have influenced the specific CD8 T cell repertoire analyzed in our study. However, in these previous studies, the overall HCV-specific immune response was narrowly focused towards the NS3 region, which is not the case in patients SR/SR-1 and SR/CI-2 as we have previously described using ELISPOT assays [23]. Furthermore, it was previously demonstrated that these influenza cross-reactive T cells were of low affinity/avidity, and are unlikely to play a major role against HCV infection [37]. Indeed, in our hands, individual T cell lines generated from either the SR/SR or the SR/CI patient exhibited comparable avidities. Finally, TCR repertoire of cross reactive CD8 T cells was previously shown to be “private” as it varied greatly from one patient to another, suggesting that these cells do not carry a specific dominant receptor [38]. Future analysis using a larger cohort of patients responding to this epitope will be required to clarify this point.
The TCR repertoire was narrower and less diverse in SR/SR patients than in the SR/CI patients at the pre-infection time point. Furthermore, this repertoire became more focused at the peak of the immune response during reinfection in patients SR/SR-1 and SR/SR-2 and retained a highly focused composition post-reinfection. This data was confirmed by examining various measures of diversity, richness and evenness. It is possible that those patients, that were recruited as resolvers had previously cleared more than one infection and may have been exposed to more than one genotype, which could have selected the most efficient repertoire. Focusing of the CD8 TCR repertoire upon re-exposure was reported in various infection and vaccination models including LCMV [34]. Focused CD8 TCR repertoires with selection of high-avidity T cell populations were also reported in HIV-1 slow progressors [39]. In contrast, a previous study of an HCV epitope (NS3-1406) suggested that higher diversity of the repertoire would be more advantageous in order to offset viral escape as a result of epitope mutation [20]. A chimpanzee challenge study reached a similar conclusion, since the majority of HCV epitopes that escaped immune recognition upon infection were targeted by a CD8 T cell repertoire with reduced CDR3 amino acid diversity, suggesting that limited TCR diversity facilitates CTL escape mutations in this animal model [17]. Results from the chimpanzee model may not accurately reflect the real life re-exposure for several reasons. First, chimpanzees in these studies were rechallenged with a specific and homologous viral sequence and therefore were exposed to a less diverse viral population as compared to humans that are typically exposed to a complex mixture of viral variants. Second, the chimpanzee data were generated using CD8 T cell clones derived from the livers of infected animals. Examining the CD8 TCR repertoire in the liver of humans is ethically difficult. Although previous analysis demonstrated that PBMCs are a good reflection of the intrahepatic TCR repertoire [21], certain low frequency clonotypes could be enriched within the liver and the in vitro expansion step employed in these studies may have introduced some bias in the repertoire.
The Morisita Horn index showed that there were more changes in the repertoire of SR/SR patients upon reinfection (from pre-reinfection to peak reinfection) as compared to SR/CI patients. This could reflect the efficient priming and focusing of the T cell response in SR/SR patients, leading to virus clearance as discussed above. Interestingly, although we could not detect significant expansion of HCV-specific tetramer+ CD8 T cells and associated clonotypes in the SR/CI patients between pre-reinfection and peak, the Morisita Horn index reflected more changes later in reinfection between the peak and late reinfection time points. This suggests that the repertoire has undergone changes following the establishment of chronic infection and the accumulation of viral variants that would prime the expansion of different clonotypes as we also observed that the repertoire became more diverse in those patients (S3 Fig) but obviously this diversification was not enough to clear the virus. Additional analysis accompanied by in depth sequencing of the infecting viral strain at each episode and of instances where the variant epitope is still recognized by the specific T cells are essential to elucidate the interaction between the T cell repertoire, the infecting virus sequence and the emergence of escape mutations.
The presence of a common set of TCRs associated with protection, known as public repertoires, was associated with viral control in several infections including HIV and CMV [5, 6, 40]. Miles and colleagues have also reported a biased TCR repertoire towards a common core of public repertoires in individuals with long-term spontaneous clearance or persistent HCV infection [19]. Our longitudinal analysis of the dominant and subdominant Vβ-CDR3-Jβ clonotypes in the SR/SR and SR/CI patients revealed no overlap between patients with the same HLA background and targeting the same epitope. Given the limited number of patients included in this study, it was not possible to draw a definitive conclusion about whether or not specific public clonotypes were associated with secondary clearance. Additional analysis of a larger cohort of patients targeting the same pMHC is required.
Establishing CD8 T cell lines enabled us to characterize the molecular determinants of functionality of the individual clonotypes. We did not observe different avidity patterns among the 10 cell lines that were analyzed using the tetramer titration assay. Testing more cell lines might be necessary to identify some with a range of TCR avidity as was observed in the HIV and LCMV models [30, 41]. It is also possible that this particular epitope selected mostly CD8 T cells with high avidity and that examining other epitopes may yield different results [42]. Furthermore, this study was performed on individuals that have successfully cleared a previous infection. It is thus possible that the clonotypes that were selected during primary infection and formed the memory pool are those with the highest avidity. Indeed, studies in a mouse model of influenza infection and rechallenge demonstrated that the clonotypes expanding in the recall response were those with the highest avidity [43]. Establishing cell lines specific to the same epitope but from early primary infection samples would thus also be informative. Moreover, a more sensitive method such as surface plasmon resonance might provide a more accurate measure of binding affinity and/or avidity that might be different between cell lines [44].
Functional avidity, or the capacity of a particular clone to translate TCR binding into a functional response, was strikingly different between the cell lines from the SR/SR patient compared to the cell lines from the SR/CI patient, especially for the surface expression of the degranulation marker CD107a and TNFα production. The cell lines established from the SR/SR patient responded well to lower peptide concentration. The polyfunctionality level was also greater in the cell lines from SR/SR as compared to the cell lines from SR/CI and is therefore an important correlate of control of viral replication [8, 45]. We have already demonstrated a broad IFNγ response in the SR/SR-1 patient as measured by ELISPOT assays but the polyfunctional CD8 T cell response to different epitopes was dominated by the production of TNFα and the surface expression of CD107a and that the CD8 T cell response had an increased magnitude and polyfunctionality in the SR/SR patients compared to the SR/CI patients [23]. Hence, the data from the individual cell lines reflected well the overall in vivo response. It is possible that in this individual, cytolytic effector functions (CD107a) leading to killing of infected cells provide an overall better antiviral effect as compared to non-cytolytic (IFNγ mediated) functions. Similarly, TNFα systemic levels increase during HCV infection [46] and it can have multiple antiviral and inflammatory effects. Specifically, it can induce the apoptosis of HCV infected hepatocytes and bystander cells in the liver, which could enhance viral clearance [47, 48].
Our study focused on the CD8 T cell response to HCV reinfection. Another important component that remains to be evaluated is the role of the antibody response and the antibody repertoire in protection upon reinfection. Indeed, Osburn et al have demonstrated that reinfection is associated with the development of cross-reactive antibodies [49]. The recent development of novel E2-tetramers that allow sorting and characterization of HCV-specific B cells and antibody repertoire [50] represent an invaluable tool to dissect the role of T cells versus antibodies in protection against reinfection.
In conclusion, our results demonstrate that epitope-specific CD8+ T cell clonotypes expanding at the peak of reinfection are recruited from the memory pool, rather than being de novo clonotypes mobilized from the naïve pool. The repertoire is narrower in the SR/SR patients who were protected against viral persistence in comparison to SR/CI patients, and it becomes more focused upon reinfection in 2/3 patients. Analysis of individual CD8 T cell lines from SR/SR versus SR/CI patients revealed that HCV-specific cells associated with resolution of the reinfection had a better functional avidity and polyfunctionality rather than improved avidity of the TCR. Vaccination strategies aiming at enhancing the expansion and polyfunctionality rather than the diversity of HCV-specific T cells by use of adjuvants or immune modulators could be an interesting strategy to follow.
Study subjects were enrolled among PWIDs participating in the Montreal Acute Hepatitis C Cohort Study (HEPCO) [51]. This study was approved by the Institutional Ethics Committee of CRCHUM (Protocol SL05.014). All samples were anonymized.
Primary HCV infection was identified in HEPCO participants who were initially negative for both HCV RNA and anti-HCV antibodies for at least 6 months, then had a positive HCV RNA and/or antibody test as previously described [32, 51]. Participants who resolved primary HCV infection or participants who tested HCV RNA-negative and HCV antibody-positive at recruitment were enrolled in the reinfection study and followed every 3 months thereafter. HCV reinfection was defined by an HCV-RNA positive test following two negative tests that were performed ≥ 60 days apart. The day of the first positive RNA test was defined as day zero post detection of reinfection. Five cases of reinfection were identified between 2009 and 2012 for whom samples collected prior to reinfection were available. Clinical outcomes and immunological responses in these patients were previously reported [23]. Three patients spontaneously resolved (SR) their second infection (SR/SR group) while two patients became chronically infected (SR/CI group). Patients’ demographics and clinical characteristics are summarized in S1 Table.
MHC class I tetramers were synthesized by the National Immune Monitoring Laboratory (NIML), (Montréal, QC, Canada) or the NIH Tetramer Core Facility (Emory University, Atlanta, GA). The following tetramers were used: HLA-A1 restricted HCV NS3 peptide amino acids (aa) 1436–1444 (ATDALMTGY) [A1/NS3-1436], HLA-A2 restricted HCV NS3 peptide aa 1073–1081 (CINGVCWTV) [A2/NS3-1073], and HLA-B27 restricted HCV peptide NS5B peptide aa 2841–2849 (ARMILMTHF) [B27/NS5B-2841]. Cryopreserved peripheral blood mononuclear cells (PBMC) were thawed and CD8+ T cells were purified using the negative selection MACS CD8+ T cell Isolation Kit (Miltenyi Biotec Inc, Auburn, CA). Tetramer staining and cell surface staining for CD3, CD8, CD45RO and CD127 were performed as previously described [32]. Directly-conjugated monoclonal antibodies against the following molecules were used: CD3–FITC (clone UCHT1), CD8–Pacific Blue (clone RPA-T8) and CD45RO–Alexa Fluor 700 (clone UCHL1) were obtained from BD Biosciences (San Diego, CA). CD127/IL-7Ra–Alexa Fluor 647 (clone HIL-7R-M21) was obtained from eBioscience (San Diego, CA). Live cells were identified using LIVE/DEAD fixable aqua dead cell stain kit (Molecular Probes Thermo Fisher Scientific, Burlington, ON). Multiparameter flow cytometry was performed on a BD Aria II cell sorter equipped with blue (488 nm), red (633 nm), and violet (405 nm) lasers or a BD LSRII instrument equipped with an additional yellow-green laser (561 nm) using FACSDiva version 6.1.3 (BD Biosciences). Data files were analyzed using FlowJo version 9.4.11 for Mac (Tree Star, Inc., Ashland, OR).
Genomic DNA was extracted from sorted cells and the variable (Vβ), diversity (Dβ) and joining (Jβ) regions of the TCR β chain were sequenced using an automated high-throughput method (Adaptive Biotechnologies, Seattle, WA). Briefly, CDR3 regions were amplified using a two-step amplification bias-controlled multiplex PCR approach [52]. Amplified libraries were sequenced using an Illumina instrument according to the manufacturer’s instructions. Demultiplexed reads were then further processed to reduce amplification and sequencing bias [53]. The resulting CDR3 amino acid sequences were classified into correct families according to the IMGT database (www.imgt.org). Data were analyzed using ImmunoSEQ software (v2.0). Clonotype lists (CDR3 sequences and frequencies within the repertoire) were cleaned to remove out of frame sequences and sequences with stop codons within the CDR3 region. Clonotypes with a frequency of less than 0.01% of the total repertoire were excluded from the analysis since the number of events/sequences would correspond to less than one cell. TCR sequences raw data are available at (https://clients.adaptivebiotech.com/pub/shoukry-2017-plospathogens). S2 Table details the number of sorted cells, total / unique productive sequences and clonality for each sample. Repertoire Simpson diversity index, richness, Shannon entropy index and Morisita Horn index were provided by Adaptive Biotechnologies. Richness index was calculated as the observed richness divided by the input cell number. Evenness was calculated as the Shannon entropy divided by the log of the observed richness. Venn diagrams were generated by comparing the amino acids clonotypes sequences across time points for each patient, after excluding clonotypes with frequencies < 0.01% as explained above. CDR3 length and NT addition for each clonotypes were provided by Adaptive Biotechnologies and the Germline index was calculated as: (Total CDR3 length–total NT additions) / Total CDR3 length.
HCV specific, tetramer positive (A2/NS3-1073) CD8+ T cells from two patients (SR/SR-1 and SR/CI-2) were enriched and sorted as described above. Cells were diluted in R10 (RPMI 1640 + 10% heat inactivated fetal bovine serum (FBS; Life Technologies) supplemented with penicillin + streptomycin (pen/strep, 1X, Wisent) and 40U/ml rIL-2 (NIH-AIDS Reagents Program) (R10-P/S-IL2); and plated at a concentration of 5 cells per well in 96 well plates in presence of 5 x 104 non-autologous irradiated (30 Gy) PBMCs as feeder cells and 0.01μg/ml anti-CD3 (Beckman Coulter). Cells were cultured for two weeks in 96 well plates and half of the medium was replenished every 3 days (R10-P/S-IL2). Growing cell lines were then transferred to 24 well plates with a new round of stimulation with feeder cells (2x106 cells/well) and anti-CD3 (0.01μg/ml final). Cell lines were then cryopreserved in freeze mix (FBS + 10% DMSO) at a concentration of 5x106 cells/ml.
Avidity of T cell lines was assessed by staining with serial dilutions of tetramers (A2/NS3-1073; 10μg/ml to 0.02μg/ml, two fold dilutions) for 30 min at room temperature in the dark. Surface staining included live/dead marker, CD3-PB (clone UCHT1), CD4-BV605 (clone RPA-T4), CD8-APC-H7 (clone SK1; all from BD Biosciences) and flow cytometry was performed as above.
T cell lines were stimulated with autologous EBV transformed B cell line (BLCLs) at a ratio of 2:1 (T cell: BLCLs). BLCLs were irradiated at 100 Gy and prepulsed with HCV NS3 peptide (NS3-1073-1081; CINGVCWTV) for 1 h at 37°C in R-10 medium. BLCLs were then washed and incubated with T cell lines for 6 hours in AIM-V medium (Life Technologies) supplemented with 10% human serum (Wisent) and anti-CD107a-BV786 antibody (clone H4A3; BD Bioscience). 10 μg/ml Brefeldin A (BFA, Sigma) and 6 μg/ml monensin (Sigma) were added after 1 h of stimulation. After stimulation, cells were washed and surface staining was performed as described in the tetramer titration section. Cells were then permeabilized with CytoFix/CytoPerm (BD Biosciences) for 15 minutes at 4°C in the dark, washed again, and incubated with anti-IFNγ-PE-Cy7 (clone B27), anti-TNF-α-PerCP-Cy5.5 (clone MAb11) and anti-IL-2-PE (clone MQ1-17H12; all from BD Biosciences) for 30 min at 4°C in the dark. Cells were then washed, fixed and analyzed as above. Polyfunctionality was analyzed using Boolean gating and SPICE software [54].
TCR sequences raw data are available at https://clients.adaptivebiotech.com/pub/shoukry-2017-plospathogens
|
10.1371/journal.pntd.0006433 | Entomopathogenic fungal infection leads to temporospatial modulation of the mosquito immune system | Alternative methods of mosquito control are needed to tackle the rising burden of mosquito-borne diseases while minimizing the use of synthetic insecticides, which are threatened by the rapid increase in insecticide resistance in mosquito populations. Fungal biopesticides show great promise as potential alternatives because of their ecofriendly nature and ability to infect mosquitoes on contact. Here we describe the temporospatial interactions between the mosquito Aedes aegypti and several entomopathogenic fungi. Fungal infection assays followed by the molecular assessment of infection-responsive genes revealed an intricate interaction between the mosquito immune system and entomopathogenic fungi. We observed contrasting tissue and time-specific differences in the activation of immune signaling pathways and antimicrobial peptide expression. In addition, these antifungal responses appear to vary according to the fungal entomopathogen used in the infection. Enzyme activity-based assays coupled with gene expression analysis of prophenoloxidase genes revealed a reduction in phenoloxidase (PO) activity in mosquitoes infected with the most virulent fungal strains at 3 and 6d post-fungal infection. Moreover, fungal infection led to an increase in midgut microbiota that appear to be attributed in part to reduced midgut reactive oxygen species (ROS) activity. This indicates that the fungal infection has far reaching effects on other microbes naturally associated with mosquitoes. This study also revealed that despite fungal recognition and immune elicitation by the mosquito, it is unable to successfully eliminate the entomopathogenic fungal infection. Our study provides new insights into this intricate multipartite interaction and contributes to a better understanding of mosquito antifungal immunity.
| Fungal biopesticides constitute potential alternative methods of vector control to tackle the rising burden of mosquito-borne diseases and the development of insecticide resistance in mosquitoes. Insect-fungi interactions represent an intricate co-evolutionary arms race between the invading pathogen and its arthropod host. New knowledge gathered through such studies can lead to the design of more effective microbial control strategies. Here we explored the temporospatial interaction of the mosquito Aedes aegypti with three different entomopathogenic fungi. Infection assays followed by gene expression studies revealed tissue-specific immune responses that appear to be temporal and fungal strain-specific. Our data shows that fungal infection causes significant reduction in phenoloxidase activity at the later stages of infection. The multifaceted response mounted by the mosquito against the fungal challenge appears to result in the dysregulation of midgut homeostasis, noted by an increase in midgut microbiota, especially in mosquitoes infected with the most virulent strains. Our study demonstrates an intricate mosquito-fungi interaction that, despite fungal recognition and immune response by the mosquito, results in death of the host.
| The global increase in new and re-emergent vector-borne diseases has put into perspective once more the need for effective methods of vector control. Although insecticide-based interventions to manage mosquito populations remain important components of vector control programs, the use of biological control agents, such as entomopathogenic fungi, offer promising alternative control methods [1, 2]. In addition to the mosquitocidal properties of entomopathogenic fungi [3–5] fungal infection leads to sub-lethal effects ranging from reduced fecundity, smaller number of gonotrophic cycles [6] and reduced vector competence, all of which are detrimental for vectorial capacity [7, 8]. Host-pathogen interactions represent an intricate co-evolutionary arms race between the invading pathogen and the host. They often encompass an interplay of biological functions, which include immunity, metabolism, and stress management by the host, and the expression of a range of virulence factors by the pathogen, in an effort to overcome the host defenses [9, 10]
In mosquitoes, recent research has uncovered several players involved in fungal recognition, immune elicitation, and anti-fungal defense. For instance, two important components of the antifungal immune response, CLSP2 and TEP22, have been recently identified from infections with the fungi Beauveria bassiana [11]. CLSP2 is a specific fungal recognition protein and negative modulator of the host antifungal immunity. Beauveria bassiana infection elicits CLSP2 expression and leads to cleavage of the CLSP2 protein, with its CTL-type domain binding fungal cell components. In turn, Tep22, a thioester-containing protein, functions as anti-fungal effector and its activity is negatively regulated by CLSP2 [11].
Following the infection and pathogen recognition, there is an activation of cellular and humoral immune responses mounted against the invading pathogen [12]. These immune responses are vastly regulated by four innate immune signaling pathways (Toll, IMD, JAK-STAT and JNK) that work in unison to control microbial infections [13–15]. Although the Toll pathway is recognized as the primary innate immune pathway that governs the antifungal response [16], the JAK-STAT pathway has also been implicated in antifungal defense [17]. In comparison, the IMD pathway is better known for playing critical roles in antibacterial and antiplasmodial defenses [18] while the lesser known JNK pathway has been shown to be critical in the antiplasmodial defense and in maintenance of cellular immunity [19, 20].
As entomopathogenic fungi penetrate the insect exoskeleton, they encounter potent cellular and systemic immune responses that include encapsulation, melanization, phagocytosis, and exposure to antimicrobial peptides. Of all these, melanization plays a crucial role in the antifungal systemic immune response [21]. In a tightly regulated reaction, melanization is catalyzed by a series of prophenoloxidase enzymes (PPOs), which leads to the deposition of melanin, and encapsulation of the invading microbe [12]. These anti-fungal immune responses are somewhat effective at initially controlling the fungal infection. Nonetheless, entomopathogenic fungi effectively counteract melanization and other immune defenses by producing metabolites that disrupt anti-microbial activities [22, 23]
In this study, we explored the temporospatial effects of fungal infection in the major arboviral mosquito vector Ae. aegypti. Infection assays with diverse fungal strains followed by the molecular assessment of infection-responsive genes revealed an intricate interaction between the fungal entomopathogen and the mosquito immune system. Moreover, this study revealed tissue-specific responses that appear to act in concert to limit fungal dissemination. Some of these anti-fungal responses are time and strain-specific. Finally, we show that fungal infection causes significant changes in phenoloxidase activity and ROS which in turn leads to dysregulation of midgut homeostasis. Studies on host responses to microbial pathogens are crucial and should be incorporated in the design of microbial control strategies for a more targeted and effective approach [9]. Although some advancements have been made, most of the studies have focused on the immune responses to either B. bassiana or Metarhizium anisopliae at the early stages of infection. Thus, understanding the interaction between the mosquito vector and entomopathogenic fungi remains an important yet understudied area. Our study assesses the mosquito molecular interactions with four different fungal strains and contributes to a better understanding of mosquito antifungal immunity.
The Rockefeller strain of the mosquito Aedes aegypti was reared in standard insectary conditions at 28°C, 70–80% relative humidity, with a 12h light/dark cycle. Adults were maintained on a 10% sucrose solution and blood feeding for egg production was conducted via an artificial membrane feeding system using bovine blood (HemoStat Laboratories Inc.). Larvae were reared on a mixture of rabbit food and tropical fish food. Three to five-day old female mosquitoes were used in all experimental assays.
Three entomopathogenic strains, Beauveria bassiana (MBC 076), Beauveria brongniartii (MBC 397) and Isaria javanica (MBC 524) were used in the exposure assays (S1 Fig). These were originally isolated from Anticarsia gemmatalis (Noctuidae, Lepidoptera); Melolontha sp. (Scarabaeidae, Coleoptera) and Hypothenemus hampei (Scolytidae, Coleoptera), respectively. The assays also included exposure to a non-entomopathogenic fungi, Trichothecium roseum (MBC 071), a saprophyte originally isolated from a syrphid fly (Syrphidae, Diptera). Fungal cultures were grown on Sabouraud dextrose agar and yeast extract (SDAY) medium and incubated at 26°C for 15 days. Conidial oil suspensions were prepared by scrapping the sporulating surface with soy bean oil, creating a homogeneous suspension via lightly mixing using an electronic pestle to break up aggregates, and then filtering the suspension through a cheese cloth to remove mycelia. Soy bean oil has been shown to be an effective carrier in oil-based formulations [24, 25], maintaining conidial viability while allowing impregnation to the hydrophobic mosquito cuticle. Conidial concentrations were determined by direct counts using an improved Neubauer hemocytometer and adjusted to a final concentration of 1x 108 conidia/mL. For mosquito exposure assays, 3–5 day old mosquitoes were cold-anesthetized, and 36.8 nl of the adjusted conidial oil suspension was topically applied to the ventral surface of the coxal region of mosquitoes using a Nanoject II micropipet. The small volume and location of the application site was chosen following tests and to avoid occlusion of the mesothoracic spiracles. A control group was exposed to the same volume of soy bean oil devoid of fungal spores. To assess the effect of bacteria on mortality of fungal-infected mosquitoes, insects were maintained on either a 10% sucrose solution containing penicillin (20 units/mL) and streptomycin (20 μg/mL) or in 10% sucrose solution alone (controls). Fresh antibiotic solution was provided daily from the day prior to the fungal infection until the end of the experiment. To assess whether any bacterial-effects on survival would be accentuated by a stronger fungal challenge, antibiotic-treated and untreated controls were exposed to a higher dose of fungal conidia (50.6 nl of a 1x 109 conidia/mL suspension). Fresh conidial suspensions from new batches of conidia were prepared for each experimental exposure assay, and a new batch of mosquitoes were used in each experiment. At least three independent experiments were conducted for each procedure involving fungal infection. Following fungal challenge, mosquitoes were maintained under standard insectary conditions with ad libitum access to 10% sucrose solution. Mortality was evaluated daily, and mosquito cadavers were removed from the cages and checked for fungal growth via maintaining them in sealed petri dishes containing moist filter paper. The survival curves were compared using Kaplan-Meier with Log-rank test to evaluate significance (GraphPad 7.0). LT50 and LT95 values were calculated by probit analysis using SAS 9.4 statistical package.
For fungal identification, DNA was extracted using CTAB extraction buffer (Sigma) while polymerase chain reaction (PCR) amplification of the partial Translation elongation factor (TEF) gene was conducted using 1x AmpliTaq Gold 360 Master Mix (Applied Biosystems), ~100 ng template DNA, and 0.2 μM each of primers 983f and 1567r [26]. The TEF gene PCR reactions were performed with an initial denaturation step of 10 min at 95°C followed by 35 cycles, each consisting of a 30 s denaturation step at 95°C, a 30 s annealing step at 55°C, and a 1 min extension step at 72°C, and then followed by a final extension step for 7 min at 72°C. Amplification products were purified using Montage PCR Cleanup Filter Plates (Millipore). Sequencing reactions were conducted using the ABI BigDye version 3.0 sequencing kit (Applied Biosystems, Foster City, CA) following the manufacturer’s suggested protocol, but at one-tenth the recommended volume. Reaction products were purified using the BigDye XTerminator Purification Kit (Applied Biosystems) following the manufacturer’s suggested protocol and sequenced on an ABI3730 genetic analyzer (Applied Biosystems) using the aforementioned oligonucleotide primers. Resulting DNA sequences were visually edited and assembled using Sequencer 5.4 software (Gene Codes Corporation). Consensus sequences were aligned using CLCBio genomics workbench 9.0 software (Qiagen Inc.). Phylogenetic analysis was performed in MEGA 7 [27] using the Maximum Likelihood method with a General Time Reversible model. The complete deletion option was used, and the level of bootstrap support was calculated from 1000 replicates (S1 Fig).
Conidial spores were collected from two-week-old cultures and stained with lactophenol cotton blue stain (Invitrogen) and imaged using a 40x objective on a Zeiss Axioplan compound microscope. For the imaging of hemocytes and blastospores, infected mosquitoes were cold-anesthetized and their hemolymph perfused with 1x PBS at 6d PI. The drop of hemolymph was collected on a microscope slide, air dried, fixed with cold methanol and stained with Giemsa stain (Invitrogen). The slides were rinsed with water before observation using a 40x objective on a Zeiss Axioplan compound microscope. Images were captured using the Leica DMC2900 microscope camera (Leica)
For gene expression, midgut and abdominal fat body walls from fungal-infected mosquitoes were dissected on a drop of 1x PBS at 3 and 6d PI. The abdominal fat body walls, comprised primarily of trophocytes and oenocytes (ectodermic cells), were devoid of Malpighian tubules, crop or ovaries but are expected to include associated tracheoles, sessile hemocytes, peripheral neurons and pericardial cells [28–30]. Then, midguts and fat body walls were homogenized in TRIzol (Invitrogen) and processed for RNA according to the manufacturer’s instructions. RNA concentration and quality were assessed via NanoDrop (Thermo Scientific), and cDNA synthesis was conducted on normalized amounts of RNA using the QuantiTect reverse transcription kit with DNA Wipeout (Qiagen). Two microliters of the generated cDNA were used in a 10 μl qPCR reaction using the PowerUp SYBR green Master mix qPCR kit (Qiagen) with gene specific primers (S1 Table). The qPCR cycling conditions were those recommended for the master mix and consisted of holding at 95.0°C for 10 min and 40 cycles of 15 s at 95.0°C and 1 minute at 60°C. A melt curve stage at the end of the reaction was included. For each experiment, the relative quantification of transcript abundance was done in two groups of midguts or their corresponding fat body walls dissected from 10 mosquitoes. Each sample was analyzed in duplicates (technical replicates) and the reproducibility of the results were evaluated via three independent experiments conducted on separate dates with different batches of mosquitoes and using fresh conidial suspensions for each infection. The ribosomal protein Rp49 gene was used for normalization of cDNA templates. This gene has been successfully used in expression profiles involving Ae. aegypti [31–33]; including transcriptomic studies on mosquito-fungal interactions [4, 34]. The qPCR assays were ran on Applied Biosystems 5700 Fast Real-time PCR (Applied Biosystems) and the data was analyzed post run using the ΔΔCt method [35]. The statistical significance of fold change values was determined on log2 transformed values via ANOVA with Dunnett’s post-test in Prism (GraphPad). Heat maps were created using Morpheus (Broad Institute, https://software.broadinstitute.org/morpheus/) from the median log2-fold change values of at least 3 independent experiments.
The Amplex Red reagent (Invitrogen) was used to determine the levels of hydrogen peroxide release (ROS) in the midguts of fungal-infected mosquitoes following the manufacturer’s recommendations. In short, midguts were dissected in PBS under a chill-block at 4°C and pooled into five midguts per group for evaluation of H2O2 levels. Samples were incubated at room temperature for 30 min with 100 μM Amplex Red reagent and 2 units of horseradish peroxidase (HRP). Samples were then spun, the supernatant collected, and fluorescence measured with Molecular Devices M5 (Ex: 530 nm; EM: 590 nm) along with a hydrogen peroxide standard curve. A non H2O2 blank sample was also included in the readings.
Mosquito midguts were also stained with dihydroethidium (DHE) (Invitrogen) to assess superoxide levels [36] following the protocol from Xiao et al [37]. In short, mosquito midguts were dissected at 6d PI in a solution of PBS and 2 mg/mL of 3-amino-1,2,4-triazole (Sigma). Dissected midguts were then incubated with 2 μM DHE in 1x PBS at room temperature for 30 min in the dark, washed twice with 1x PBS and then fixed with 4% paraformaldehyde for 30 minutes. Midguts were then washed twice before being placed on a slide. Nuclei was stained with DAPI (Invitrogen) and the slides were imaged using a 4x objective lens on an EVOS FL Auto (Life Technologies) fluorescent microscope. Images were acquired using the same fluorescent levels and under identical conditions (exposure time, microscope, reagents, etc.) for all tested groups.
To assay the level of PO activity in mosquitoes, we followed the protocol from Sadd et al. [38] with modifications. In short, mosquitoes were collected at 3 and 6d PI, and two whole mosquitoes were pooled per sample (10 samples per treatment) and homogenized for 30 s with a 2.4 mm bead and 50 μl of 1x PBS using the TissueLyser II (Qiagen). Homogenates were then centrifuged (3000 rpm, 4°C, 5 min), and then 35 μl of the supernatant was snap frozen in liquid nitrogen and stored at -80°C for subsequent analyses. The enzymatic reaction assay was conducted by thawing the samples on ice and adding 15 μl of the sample or 15 μl of PBS (negative control) to a flat-bottomed 96-well plate containing 20 μl PBS and 140 μl of molecular-grade water. Subsequently, 20 μl of a solution of molecular grade water mixed with L-Dopa (4 mg per mL H2O; 3,4 dihyroxy-L-phenylalanine) was added to each well. Plates were shaken for 5s at 30°C in a spectrophotometer (Multiskan GO, Thermal Scientific), with absorbance read at 490 nm every 15 s with 5 s shaking between reads. PO activity was measured from the slope (Vmax) of the reaction curve in its linear phase over 160 readings. Samples were read in duplicate and the average was used in further analyses. Ten samples per treatment were used in each experiment with three independent experiments conducted.
Bacterial load was determined via qPCR on both the cDNA of mosquito midguts used for gene expression analyses and from genomic DNA extracted from the midguts of separate cohorts of fungal-infected mosquitoes. DNA was extracted from pools of 10 midguts using the DNeasy blood and tissue kit (Qiagen) kit according to the manufacturer’s instructions for bacterial DNA extraction. Amplification was conducted using 16s ribosomal RNA universal primers (S1 Table) following the same conditions as for gene expression assays stated above. The Ae. aegypti ribosomal protein Rp49 and the ribosomal protein Rps17 gene were used for normalization of cDNA and DNA templates respectively. To normalize the amount of bacteria present in our cohort of mosquitoes before fungal infection, mosquitoes were treated with antibiotics (20 μg/mL streptomycin and 20 units/mL penicillin) ad libitum for two days, maintained in sterile sucrose for one more day and then fed on a cocktail of bacteria via sugar meal for 24 h prior to fungal challenge. Bacterial isolates used in this study were the most commonly isolated species from our lab-reared mosquitoes and identified as: Serratia marcescens, Burkholderia spp., Leclercia spp., and two species of Pseudomonas spp. Bacteria were grown overnight, spun at 3,000 rpm for 5 min and washed twice in 1x PBS. They were then adjusted to 1x 104 for each bacterium and mixed in a 3% sucrose solution prior to their use. Before dissection, mosquitoes were surface-sterilized with 3% bleach for 2 minutes, 70% EtOH for 2 minutes and washed twice with 1x PBS. Midguts were dissected on a drop of sterile 1x PBS and then pooled in groups of 10 midguts per sample. Midguts were homogenized using the TissueLyser II (Qiagen) and DNA was extracted with the DNeasy Blood and Tissue kit (Qiagen). Their concentration and quality were analyzed via NanoDrop (Thermo Scientific), and employed in subsequent Illumina sequencing.
The midgut microbiome of control and fungi-challenged mosquitoes following the reintroduction of bacteria post-antibiotic treatment was analyzed via 16s rRNA sequencing. Ten midguts were pooled for each treatment with the analysis conducted in four to five replicates. PCR was performed using Kapa HiFi PCR mastermix (Kapa Biosystems Willington, MA) using the following parameters; 95°C, 10 min, 35 cycles of 95°C, 30 s; 58°C, 30 s; 72°C, 60 s. PCR primers for the bacterial community (341f and 806r) targeted the V3-V4 region of the 16S rRNA genes [39]. The loci-specific primers were incorporated into fusion primers for Illumina dual indexing and incorporation of Illumina adapters [40]. After sequencing, the amplicons were cleaned and normalized using a SequalPrep normalization plate (Thermo Fisher Scientific, Waltham, MA). The samples were pooled and the library quantified with a Kapa library quantification kit (Kapa Biosystems Willington, MA). The samples were sequenced using an Illumina MiSeq system with a MiSeq V3 2 x 300 bp sequencing kit. The demultiplexed reads were quality trimmed to Q30 using the CLC genomics workbench v9.5 (Qiagen inc., Valencia, CA). Read pairing, fixed length trimming and OTU clustering was conducted using the CLC Bio Microbial Genomics module (Qiagen inc., Valencia, CA) using the reference sequences from the Greengenes ribosomal ribonucleic acid gene database (97% similarity) [41]. OTUs with five or more sequence reads were used in subsequent analyses. Alpha diversity was evaluated using Chao1, Shannon, and Simpson indices; while the beta diversity was analyzed by using the principal coordinate analyses (PCoA) based on the Bray-Curtis dissimilarity matrix. The diversity indices were tested for normality via the Shapiro-Wilk normality test, those passing the normality test were analyzed via ANOVA with the Dunnett’s multiple comparison test. Otherwise, they were analyzed via the Kruskal-Wallis test for multiple comparison. Diversity measurements were performed on PAST 3.15 software and the statistical analysis was conducted with GraphPad Prism 7 (GraphPad).
All statistical analyses to assess significance of Kaplan-Meier survival curves, gene expression analyses, enzymatic activity, and biodiversity index comparisons were performed using GraphPad Prism 7 (GraphPad). Significance was assessed at P<0.05 with asterisks indicating the strength of the significance (*P< 0.05; **P<0.01; ***P<0.001). Error bars represent the standard error of the mean and the type of test used is indicated in the respective figure legend.
Mosquito survival differed significantly following challenge with the four fungal species (log-rank Mantel–Cox test, X2: 181.4, P<0.0001) (Fig 1A). Mosquitoes topically infected with B. bassiana had the highest mortality (94.9%, X2: 86.33, P<0.0001). The second and third most virulent species were B. brongniartii with 88% (X2: 81.46, P<0.0001) and I. javanica with 71.8% mortality (X2: 47.66, P<0.0001). The low mortality (6%) observed in mosquitoes challenged with the saprophytic isolate T. roseum was not significantly different from the control group (4%, X2: 0.2409, P = 0.6236). The values for LT50/LT95 were determined for each pathogenic strain and varied with 7.39 days for B. bassiana, 9.25 days for B. brongniartii and 12.44 days for I. javanica (S2 Table). LT50 could not be estimated for T. roseum due to mosquito survival exceeding 50% or 95% throughout the experiment. Slide preparations from fungal spores confirmed both conidial morphology and purity of the samples (Fig 1B, top panel). Fungal bodies were present in the hemolymph of mosquitoes challenged with B. bassiana, B. brongniartii and I. javanica (Fig 1B, middle panel) at 6d PI, confirming the success of fungal infection. We did not observe any blastospores or fungal bodies in the hemolymph preparations of the T. roseum-challenged mosquitoes. Cadavers collected soon after death showed mycelial growth and sporulation for all of the fungal-challenged groups (Fig 1B, lower panel), but those were absent in the few control group cadavers.
To assess successful fungal infection and recognition by the mosquito immune system, we evaluated the expression of two genes, CLSP2 and TEP22, which have been found to be significantly expressed in response to B. bassiana infection at 24h post-exposure [11]. CLSP2 is composed of a serine protease and a C-terminal galactose-type C-type lectin domain and acts as a negative regulator of the antifungal immune response [11]. CLSP2 is upregulated at 3d PI in the fat body of mosquitoes infected with B. bassiana and B. brongniartii, and shows a slight increase, albeit not significant, in mosquitoes challenged with I. javanica or T. roseum (Fig 2). This increase in fat body CLSP2 expression became more prominent at 6d PI for all fungi-infected mosquito groups except for T. roseum-challenged mosquitoes (ANOVA, with Dunnett’s test, P = 0.874).
Although, CLSP2 and TEP22 have been observed to primarily be expressed in the fat body, our expression analysis shows an upward trend at 3d PI in the midgut of all infected mosquitoes (Fig 2). While not significant, by 6d PI it was no longer observed in infected mosquitoes; with all CLSP2 upregulation occurring exclusively in the fat body. TEP22 regulation in the midgut was observed at 3d and 6d PI, but overall it was significantly upregulated only in the midgut of mosquitoes infected with B. bassiana. In comparison, a significant increase in fat body TEP22 expression was observed at 3d and 6d PI in mosquitoes infected with B. bassiana, B. brongniartii and I. javanica (Fig 2). The fat body of T. roseum-challenged mosquitoes showed no significant increase in TEP22 expression.
To assess fungal elicitation of the innate immune pathways, we evaluated the expression of the transcription factors REL1, REL2, STAT, and pathway component JNK, in the midgut and fat body at 3 and 6 days PI. These pathway components belong to the four-main mosquito immune pathways Toll, IMD, JAK-STAT and JNK respectively. REL1 expression was significantly downregulated in the midgut of B. brongniartii-infected mosquitoes at 3d PI but returned to the control levels at 6d PI (Fig 3A). No other modulation of this transcription factor was observed in the midgut at 3d or 6d PI with the other fungal strains. In contrast, REL1 expression was significantly upregulated in the fat body at 3d PI in B. bassiana and I. javanica-infected mosquitoes with stronger expression in the fat body of mosquitoes infected with B. bassiana and B. brongniartii at 6d PI. No significant REL1 regulation was observed in T. roseum-infected mosquitoes (Fig 3A). In comparison to REL1 (Toll pathway), the IMD pathway transcription factor REL2 had no modulation of expression at 3d PI but was strongly upregulated in the midguts of B. bassiana, B. brongniartii and I. javanica-infected mosquitoes at 6d PI. No significant modulation of REL2 expression was observed in the midgut of T. roseum-infected mosquitoes.
As for the fat body tissues, REL2 expression at 3d was only slightly upregulated in mosquitoes infected with B. bassiana but had a robust upregulation at 6d PI in mosquitoes infected with either of the two Beauveria species or I. javanica, but with no change in the T. roseum-challenged group. STAT showed no significant regulation in the midgut at 3d PI and a slight upregulation at 6d PI in B. bassiana and B. brongniartii-infected mosquitoes. The fat body showed similar trends, with no significant fat body STAT expression at 3d but with a significant increase only in B. bassiana and B. brongniartii-infected mosquitoes at 6d PI. JNK expression was not significantly modulated at 3d PI in the midgut or fat body but presented a strong significant upregulation at 6d PI only in the midgut and fat body of B. brongniartii-infected mosquitoes (Fig 3A).
We then evaluated the expression of four antimicrobial peptides, cecropin G (CECG), defensin C (DEFC), attacin (ATTA) and lysozyme C (LYSC). We observed a dynamic modulation of expression that was also tissue-specific, time and fungal-isolate dependent (Fig 3B, S2 Fig). Cecropin G, DEFC and ATTA showed no significant changes in expression in midguts or fat body tissues at 3d PI but a strong CECG upregulation by 6d PI only in B. bassiana-infected mosquitoes. The expression of attacin in midgut and fat body tissues presented an irregular modulation across the biological replicates; and as a group it was not statistically significant. We also observed weak modulation of DEFC with a significant upregulation at 6d PI only in the fat body of B. bassiana-infected mosquitoes. Lysozyme C was the only effector that consistently show downregulation in the midgut and upregulation in the fat body of fungal-infected mosquitoes. In particular, LYSC was significantly downregulated in the midguts of B. brongniartii and I. javanica-infected mosquitoes at 6d PI. In contrast, LYSC expression was significantly upregulated at 3d and 6d PI in the fat body of mosquitoes infected with the entomopathogenic strains B. bassiana, B. brongniartii and I. javanica. No significant LYSC modulation was observed in T. roseum-challenged mosquitoes (Fig 3B and S2 Fig).
We investigated the expression of two important gene members of the phenoloxidase cascade (PPO3 and PPO5). We observed a decrease in PPO3 and PPO5 expression in the midgut and fat body of B. bassiana and B. brongniartii-infected mosquitoes at 3d PI with a more pronounced downregulation in fat body tissues at 6d PI (Fig 4A). The PPO3 and PPO5 expression levels in mosquitoes infected with T. roseum and I. javanica was not-significant relative to the control (Fig 4A and 4B). We next measured the whole body phenoloxidase (PO) enzymatic activity at 3d and 6d PI. We observed a significant decline in PO activity in mosquitoes infected with B. bassiana and B. brongniartii, and non-significant decline in I. javanica-infected mosquitoes at 3d PI (Fig 4C). This reduction in PO activity was stronger at 6d PI in all mosquitoes exposed to either B. bassiana, B. brongniartii or I. javanica (Fig 4D). No change was observed in mosquitoes challenged with the saprophytic isolate T. roseum (Fig 4C and 4D).
We decided to evaluate whether the modulation of the mosquito immune system by the fungal challenge could be reflected in the midgut homeostasis and affect microbial load. Measuring the bacterial 16s rRNA relative to mosquito Ribosomal protein 49 (Rp49), at 3d and 6d PI, showed an upward trend in the midgut bacterial load of mosquitoes infected with B. bassiana and B. brongniartii, with less of an increase in mosquitoes infected with I. javanica (Fig 5A). This increase, was significant in the midgut of B. bassiana-infected mosquitoes at 6d PI. To corroborate our transcript results, we evaluated the levels of 16s rDNA via qPCR. The results were in accord to our transcript analysis with significant increase in bacterial 16s rDNA copies in mosquitoes infected with B. bassiana and B. brongniartii (Fig 5B).
We next evaluated whether these changes were due to an overall increase in the midgut bacterial load or to a specific bacterial taxon, one benefiting from the midgut physiological conditions created in response to the fungal-challenge. Our microbiome study indicated no significant difference in the midgut microbial composition among all treatments. Although there was a slight increase in microbial diversity in control mosquitoes at the family and genera level (Fig 5C and S3A Fig), the OTU calculated species richness (Chao1), Shannon’s diversity index and Simpson’s diversity index did not change significantly from the controls (P > 0,05) (S3B Fig) The microbial community was dominated by two Enterobacter species with Serratia marcescens being the most abundant of all OTUs across all sample treatments. Other represented genera included Pseudomonas spp., Burkholderia spp., Pseudoxanthomonads spp., and Xanthomonads spp. (Fig 5C). For beta diversity analysis, we evaluated the relationship between entomopathogenic fungal infection and mosquito midgut microbiota using the principal coordinate analysis (PCoA) based on the Bray-Curtis dissimilarity matrix. The PCoA test showed that the community composition did not differ among treatments (Fig 5D).
To understand the physiological changes behind the increase in gut microbiota of infected mosquitoes, we analyzed the expression of known gut-homeostasis-related genes. One such gene is the gut-membrane-associated protein (MESH), which recently has been recognized to be part of the insect gut homeostatic mechanism [37]. Our analysis showed no change in MESH expression in the midgut at 3d PI but a significant increase in the midgut of B. bassiana and B. brongniartii-infected mosquitoes at 6d PI (Fig 6A). These coincided with the increases of bacterial load observed in our microbiome load assessment (Fig 5B). Upregulation of MESH expression due to an increase in midgut microbiota has been reported previously for the Ae. aegypti mosquito [37]. Given that MESH is known to control the expression of DUOX, we next decided to evaluate DUOX1 and DUOX2 expression in the midgut of infected mosquitoes. We observed no changes in the expression of DUOX1 at 3d PI, but a substantial increase in expression at 6d PI in the gut of B. bassiana and B. brongniartii-infected mosquitoes (Fig 6B). No significant regulation of DUOX1 was observed in the midguts of T. roseum or I. javanica at 3d or 6d PI. In addition, no change in the DUOX2 expression profiles were observed in the midgut of mosquitoes from any of the treatments. Since DUOX1 is directly related to ROS production in the midgut we assessed ROS activity levels by measuring hydrogen peroxide (H2O2) and superoxide production by the midgut. Our assessment shows a significant decrease of H2O2 release (Fig 6C) as well as a decrease in superoxide staining by the midgut epithelium (Fig 6D) at 6d PI. This decrease was independent of whether the mosquito was challenged with an entomopathogenic or non-entomopathogenic fungal isolate, since T. roseum-infected mosquito midguts released 30% less H2O2 than controls, compared to the 20% reduction seen in mosquitoes exposed to B. bassiana or B. brongniartii or 24% in mosquitoes challenged with I. javanica (Fig 6C). These results were consistently observed in three independent experiments. Next, we assayed the expression of antioxidant genes that would counter act any effects from ROS levels. No significant changes were observed in the midgut expression of CuSOD2, GPX, and Catalase gene (CAT) at 3d or 6d PI, while a significant TPX up-regulation at 6d PI was observed only in B. brongniartii-infected mosquito midguts (Fig 6E). Overall, fat body tissues had a more robust modulation of DUOX genes as well as antioxidant genes especially at 6d PI with a significant DUOX2 upregulation in the fat body of B. brongniartii-infected mosquitoes (Fig 6F and S4 Fig). Of the four different antioxidant genes evaluated, CuSOD2 and TPX were the only two genes whose expression was significantly upregulated in either B. bassiana or B. brongniartii-infected mosquito fat bodies (Fig 6F and S4 Fig). There was no significant regulation of any of these genes in fat bodies of I. javanica or T. roseum-infected mosquitoes.
To assess whether the gut bacteria had any influence in the mosquito survival to fungal infection, mosquitoes were treated with antibiotics from the day prior to infection until 15d PI. The antibiotic treatment was successful in clearing the mosquito gut bacteria (S5 Fig), but it did not change the survival rate of the fungal-infected mosquitoes (S6 Fig). This was observed even when mosquitoes were exposed to a higher concentration of fungal conidia (S7 Fig).
Our comparison study with fungal strains of diverse virulence shows a dynamic interaction of immune signaling pathways and effectors working in concert to control the infection. The observation of blastospores in mosquito hemolymph coupled with the presence of mycelial growth in mosquito cadavers confirmed the successful infection by these entomopathogenic strains. In comparison, the absence of blastospores but presence of mycelial growth on mosquito cadavers from the T. roseum-challenged group indicated the true nature of this fungus, since saprophytic fungi are known to grow on dead organic matter, including dead insects.
Upon immune challenge, pathogen recognition is the first critical step in the elicitation of immune responses mounted to control pathogen invasion. The fungal recognition protein CLSP2 and the anti-fungal effector Tep22 have been recently identified as part of the mosquito anti-fungal response [11]. In a series of experiments Wang et al [11] described elevated levels of expression for these two molecules at 24h post-infection with B. bassiana. Our results corroborate the elicitation of CLSP2 and TEP22 and show increasing fat body expression of these two genes at the later times of infection, 3d and 6d PI. Fungal recognition by these two molecules was independent of the fungal strain and the magnitude of their regulation appears to reflect the virulence of the infecting fungi, with B. bassiana-infected mosquitoes showing the strongest CLSP2 upregulation at 3d and 6d PI. This indicates that the degree of immune elicitation is fungal strain-dependent.
The strong expression of the anti-fungal effector TEP22 in the fat body of infected mosquitoes at 3d PI indicates the attempt by the mosquito’s immune system to control the infection. It is interesting to note that while TEP22 expression increases with time, as blastospores start to disseminate throughout the mosquito body, its negative regulator CLSP2 also increases in expression. This suggests a tight control of the response as the infection progresses. Furthermore, while B. bassiana is eliciting the strongest expression of this antifungal effector, mosquitoes are nevertheless unable to limit the infection and succumb two to five days earlier than mosquitoes exposed to B. brongniartii or I. javanica. This could mean that while TEP22 is somewhat effective at limiting I. javanica and B. brongniartii proliferation and dissemination; it is less so with regards to B. bassiana.
Pathogen recognition leads to the elicitation of immune signaling pathways that would direct anti-microbial immune responses. Our study shows a degree of compartmentalization for each of the immune signaling pathways in response to fungal infection. While the transcription factor REL1 (Toll pathway) is elicited in the fat body especially at the later stages of infection, it remained unchanged in the midgut compartment. This is in contrast to REL2 expression (IMD pathway), which largely showed no significant expression at 3d PI, but displayed a robust expression in the midgut and fat body of infected mosquitoes at 6d PI. These results are in line with what has been reported previously in Drosophila in that the Toll pathway is mostly expressed in hemocytes and fat body for cellular and humoral immunity [42], while IMD pathway acts both at the midgut and fat body for epithelial and humoral immunity [43–45]. Whereas the Toll pathway is known to be involved in the antifungal defense in both fruit flies [46] and mosquitoes [16, 47], the IMD pathway is generally regarded as part of the antibacterial and anti-plasmodial defense [16, 48]. Interestingly, our results indicate that the IMD pathway may be playing a bigger role in the anti-fungal defense, since its transcription factor was highly elicited with all three entomopathogenic fungi. Alternatively, the IMD pathway activation in the midgut might occur in response to the increase in the microbial gut flora observed at the later stages of infection, controlling or preventing a systemic bacterial infection. Another interesting observation was that, with exception to the activation of the Toll pathway at 3d PI in the fat body, the IMD pathway was the only immune signaling pathway elicited in I. javanica-infected mosquitoes. This suggests that while the mode of invasion is relatively similar among all entomopathogenic fungi, its dissemination and development as blastospores inside the hemocoel might interact differently with the mosquito immune system.
In addition, the expression of STAT and JNK, from the JAK-STAT and JNK pathways respectively, showed additional modulation of pathway activation in relation to the invading fungal strain. The strong elicitation of the STAT pathway in B. bassiana and B. brongniartii-infected mosquitoes, but not in I. javanica, is likely the result of specific fungal-mosquito interactions. For instance, it could be in response to fungal secondary metabolites modulating the mosquito immune response. Prior research has found that JAK-STAT pathway is crucial in the anti-fungal defense against B. bassiana [47]. Although further analyses are needed, our results might indicate a potential divergence in this response, given the lack of elicitation of this pathway in I. javanica-infected mosquitoes. Although less is known about the importance of the JNK pathway in the anti-fungal response, the fact that this pathway was strictly elicited in B. brongniartii-infected mosquitoes, reflect specific interactions that occur between each fungal strain and the mosquito. Pathogen-specific elicitation of immune signaling pathways is not uncommon and has been observed in the Drosophila-bacteria [43, 49] and in the Anopheles-Plasmodium system [50]. Hence, a level of pathogen-specific modulation of immune signaling pathways is likely to occur with different fungal strains.
One important picture that emerged in our expression analysis is that the main signaling pathways appear to be elicited consistently at the later stages of infection, while largely remaining unchanged relative to the controls at the early 3d PI time point. This could be the result of early active-immune suppression by the entomopathogenic strains or the result of immune evasion by fungal morphological changes (blastospores). Indeed, prior studies on entomopathogenic fungi metabolism have described the existence of bioactive factors with potent immune suppressive activity [9, 51]. B. bassiana as well as B. brongniartii are known to produce several of these compounds [52]. Whether any of these bioactive compounds are directly affecting these immune signaling pathways in the mosquito remains to be investigated. Thus, while immune suppression is possible, the lack of immune signaling activation might be due to the lack of sufficient recognition of blastospores inside the mosquito body [53]. This would agree with what is currently known about blastospores in that they are devoid of most of the structural cell wall found in conidia or mycelia that would elicit the immune system and hence, go largely unnoticed by hemocytes [9]. In addition, the weak and irregular modulation of three of the four antimicrobial peptides tested may suggest a level of immune suppression by the fungal entomopathogen. For instance, CECG and DEFC were only significantly expressed at 6d PI in B. bassiana-infected mosquitoes but presented an irregular and not significant expression in mosquitoes infected with the other two fungal entomopathogens. Alternatively, it may indicate that the expression of these effectors is not directly modulated by the fungi but it is an indirect effect of fungal infection, for instance, they could be elicited in response to the increase in gut microbiota. The only antimicrobial peptide consistently upregulated across all the different entomopathogenic fungal infections was LYSC. The significant regulation of this AMP in the fat body at 3d and 6d PI might suggest that it is playing a bigger role in the antifungal defense compared to the other two AMP genes tested.
Another important immune response mounted against invading fungal pathogens is the melanization cascade [21, 54, 55]. Although phenoloxidase activity has been attributed primarily to hemocytes [56], it has also been localized in tissues such as the hindgut [57], trachea, and adult cuticle [58]. Detailed analysis during the acute phase (24h) of B. bassiana infection has shown the elicitation of several prophenoloxidase genes [11]. Our analysis at 3d and 6d PI revealed the downregulation of two important members of the melanization cascade (PPO3 and PPO5). In addition, we observed a significant reduction in phenoloxidase activity in the whole body of B. bassiana and B. brongniartii-infected mosquitoes at these same time points of infection. These indicates active modulation of this antifungal defense mechanism. While the reduction in PO activity could be due to active immune suppression by the entomopathogenic fungi, it is plausible that it is regulated by the mosquito itself. In fact, CLSP2 has been found to negative modulate the transcription of PPO genes during fungal-infection [11]. Although, Wang et al [11] did not report downregulation of PPO genes, silencing CLSP2 led to an increase in the expression of these same phenoloxidase genes at 24h PI. In our studies, significant CLSP2 upregulation coincided with significantly lower PPO3 and PPO5 expression at the later stages of B. bassiana and B. brongniartii infection. A similar reduction in PO activity at the later times of infection has been observed following infection with B. bassiana in the moth Spodoptera litura, the fruit fly Drosophila melanogaster and the wax worm Galleria mellonella [55, 59, 60]. Another line of evidence for lower PO activity resides in the upregulation of lysozyme and REL2 expression post-fungal infection. Our analysis indicates a significant increase in lysozyme expression in the fat body at 3d and 6d PI. Lysozyme has been found to negatively regulate PPO activation upon an immune challenge [61, 62]. In addition, REL2 has been shown to negatively regulate melanization in the mosquito Anopheles gambiae [18]. Our expression analysis showed a significant increase in REL2 expression at the later stages of infection when PO activity was the lowest particularly for B. bassiana and B. brongniartii-infected mosquitoes. Thus, this potential tradeoff between PO activity and lysozyme and/or REL2 regulation could be a mosquito strategy to avoid the burst of the most costly and toxic PO activation. Alternatively, the drop in PO activity, especially at the later time points, could be the result of an exhaustion of the melanization pathway components, as they are being used to counteract the continuous proliferation of fungal bodies. However, a study by Matskevich et al. [59], who also observed suppression of PO activity, found similar amounts of prophenoloxidase proteins in naive and B. bassiana-infected D. melanogaster fruit flies. Hence, whether this reduction in PO activity is the result of immune suppression by the entomopathogenic fungi or a protective process developed by the mosquito and exploited by the entomopathogenic fungi, remains to be elucidated.
The infection-modulated expression of some immune genes led us to investigate the microbial flora of infected mosquitoes since their elicitation might indirectly affect the midgut microbiome. To our surprise we observed an increase in the midgut bacterial load of infected mosquitoes at the later stages of fungal infection. The level of midgut microbiota increase appears to reflect the magnitude of the immune response, since mosquitoes that had the most immune elicitation (B. bassiana and B. brongniartii-infected mosquitoes) also had the highest increase in gut microbiota. Hence, this increase in midgut bacterial load is most likely the indirect result of the multifaceted response mounted by the mosquito against the fungal challenge. A microbiome analysis across all treatment groups showed no large differences on the commensal gut bacteria composition. Thus, the bacterial load increase in the midgut of infected mosquitoes likely occurs independent of bacteria taxa. A recent article by Wei et al. [63] also describes an increase in gut bacteria loads following B. bassiana infection in Anopheles stephensi mosquitoes. Furthermore, the authors found that the increase in bacteria accelerated the death of infected mosquitoes. Our comparative study using antibiotic-treated and untreated mosquitoes did not show any significant difference with regards to survival. The difference between the results described by Wei et al. and our work could be attributed to the use of different mosquito species and different fungal strains, or due to differences in bacterial composition in the gut of mosquitoes. Our study evaluated the commensal midgut bacteria during the course of an entomopathogenic fungal infection, and it could be expected that pathogenic bacteria would produce potentially different phenotypes. Although our results show no detrimental effect of the microbial load increase in the survival of infected mosquitoes, it could still alter vector competence. This could be a potential phenotype given that the midgut microbiota has been shown to influence arboviral and parasitic infection of the mosquito midgut [64–66].
To decipher any mechanisms that were leading to the increase in midgut microbial load, we evaluated the expression of genes involved in maintaining gut homeostasis [37, 49]. The dual oxidase (Duox)-generation of reactive oxygen species (ROS) has been identified to be a critical mechanism controlling the proliferation of gut bacteria. This mechanism, in turn, has been shown to be under the control of MESH, a gut-membrane-associated protein [37]. In addition, the Mesh-mediated DUOX induction is known to be elicited in response to bacteria proliferation in the gut. Hence, MESH upregulation in the gut of B. bassiana and B. brongniartii-infected mosquitoes at 6d PI is likely a secondary response to the increase in gut bacteria, especially since MESH-DUOX expression were absent at the early time of 3d PI.
The significant reduction in soluble H2O2 (a potent ROS) being released by the midgut of fungal-challenged mosquitoes most likely explains the increase in bacterial loads. In addition, the gene expression analysis of the Duox pathway and other detoxifying enzymes suggests that there is a post-translational regulation of the DUOX enzyme. Studies have found that peptidoglycan from commensal bacteria did not induce DUOX enzyme activity even though it did induce DUOX gene expression [67, 68]. This is a controlled regulation that prevents overstimulation and generation of potentially deleterious ROS in the absence of pathogenic gut bacteria [68]. Alternatively, it could also be part of the antifungal immune response profile, directing and localizing ROS production where the fungal pathogen is detected (hemocoel) and not in the gut, where commensal bacteria are located. Further studies are needed to discern the significance of this phenotype.
In comparison to the midgut, the fat body showed high elicitation of both oxidant-generating and detoxifying enzymes in mosquitoes infected with the entomopathogenic fungal strains. This expression profile most closely represents the systemic immune response that might be expected at the later stages of infection. Interestingly, T. roseum-challenged mosquitoes also had a significant drop in H2O2 release in three independent experiments. The limited transcriptome analysis conducted in this study does not allow us to determine a possible mechanism for this phenotype in T. roseum, but it might indicate that this is part of the antifungal response when challenged with fungal spores, independent of whether they are pathogenic or not. Further studies are needed to determine the implications of this response in T. roseum.
In summary, our study shows that the antifungal immune response of the mosquito Ae. aegypti displays a degree of compartmentalization that changes as the fungal entomopathogen disseminates throughout the body of the mosquito during the infection process. Our results further show a tissue-specific immune modulation that in turn varies according to the fungal entomopathogen infecting the mosquito. The concomitant differential elicitation of immune pathways along with the modulation of redox signaling pathways appear in turn to alter midgut homeostasis. This indicates that fungal infection have far reaching effects to other microbes that naturally reside in mosquitoes, which could in turn alter the vector competence of infected mosquitoes.
Our study demonstrates an intricate mosquito-fungi interaction, which, despite fungal recognition and immune elicitation by the mosquito, results in the death of the host. This observation, in turn, indicates that the fungi may play a more complex role in suppressing or modulating mosquito immunity than has been previously recognized. In addition, this study contributes to a better understanding of mosquito-fungi interactions that may facilitate the development of novel fungal-based biocontrol strategies, for instance, through the selection of entomopathogenic fungi with different modes of action.
|
10.1371/journal.pcbi.1002435 | Hemodynamic Traveling Waves in Human Visual Cortex | Functional MRI (fMRI) experiments rely on precise characterization of the blood oxygen level dependent (BOLD) signal. As the spatial resolution of fMRI reaches the sub-millimeter range, the need for quantitative modelling of spatiotemporal properties of this hemodynamic signal has become pressing. Here, we find that a detailed physiologically-based model of spatiotemporal BOLD responses predicts traveling waves with velocities and spatial ranges in empirically observable ranges. Two measurable parameters, related to physiology, characterize these waves: wave velocity and damping rate. To test these predictions, high-resolution fMRI data are acquired from subjects viewing discrete visual stimuli. Predictions and experiment show strong agreement, in particular confirming BOLD waves propagating for at least 5–10 mm across the cortical surface at speeds of 2–12 mm s-1. These observations enable fundamentally new approaches to fMRI analysis, crucial for fMRI data acquired at high spatial resolution.
| Functional magnetic resonance imaging (fMRI) experiments have advanced our understanding of the structure and function of the human brain. Dynamic changes in the flow and concentration of oxygen in blood are observed experimentally in fMRI data via the blood oxygen level dependent (BOLD) signal. Since neuronal activity induces this hemodynamic response, the BOLD signal provides a noninvasive measure of neuronal activity. Understanding the mechanisms that drive this BOLD response is fundamental for accurately inferring the underlying neuronal activity. The goal of this study is to systematically predict spatiotemporal hemodynamics from a biophysical model, then test these in a high resolution fMRI study of the visual cortex. Using this theory, we predict and empirically confirm the existence of hemodynamic waves in cortex – a striking and novel finding.
| Functional magnetic resonance imaging (fMRI) experiments have substantially advanced our understanding of the structure and function of the human brain [1]. Hemodynamic responses to neuronal activity are observed experimentally in fMRI data via the blood oxygenation dependent (BOLD) signal, which provides a noninvasive measure of neuronal activity. Understanding the mechanisms that drive this BOLD response, combined with detailed characterization of its spatial and temporal properties, is fundamental for accurately inferring the underlying neuronal activity [2]. Such an understanding has clear benefits for many areas of neuroscience, particularly those concerned with detailed functional mapping of the cortex [3], those using multivariate classifiers that implicitly incorporate the spatial distribution of BOLD [4], [5], and those that focus on understanding and modeling spatiotemporal cortical activity [6]–[10].
The temporal properties of the hemodynamic BOLD response have been well characterized by existing physiologically based models, such as the balloon model [11]–[14]. Although the spatial response of BOLD has been characterized experimentally via hemodynamic point spread functions [15]–[18], it is commonly agreed that the spatial and spatiotemporal properties are relatively poorly understood [19], [20].
Many studies work from the premise that the hemodynamic BOLD response is space-time separable, i.e. is the product of a temporal HRF and a simple Gaussian spatial kernel. The latter assumed as a simple ansatz or ascribed to diffusive effects, for example [21]. This approach raises the following concerns: (i) since the temporal dynamics of the HRF is the focus of most theoretical analyses, e.g. the balloon model, this precludes dynamics that couple space and time, dismissing whole classes of dynamics, such as waves; (ii) in practice, employing a static spatial filter then convolving with a temporal HRF on a voxel-wise basis neglects non-separable interactions between neighboring voxels; and (iii) calculating temporal correlations between voxels then assumes that the hemodynamic processes responsible for the signal occur on scales smaller than the resolution of the measurements. In summary, neglecting spatial effects such as voxel-voxel interactions and boundary conditions (e.g., blood outflow from one voxel must enter neighboring ones) ignores important phenomena and physical constraints that could be used to increase signal to noise ratios and to improve inferences of neural activity and its spatial structure. These constraints are becoming increasingly relevant, as advances in hardware and software improve the spatial resolution of fMRI by reducing voxel sizes.
While treating spatial hemodynamics as a Gaussian is a reasonable first approximation [15]–[20], [22], this requires spatiotemporal BOLD dynamics, such as spatially delayed activity to be attributed solely to the underlying neuronal activity, without hemodynamic effects from neighboring tissue, an assumption that may not be valid. In this limit, BOLD measurements would simply impose a spatial low-pass filter of neuronal activity [20]. Several studies have already presented results that challenge this assumption, most strikingly by demonstrating reliable classification of neuronal structures such as ocular dominance or orientation columns [4], [5] on scales significantly smaller than the resolution of the fMRI protocols used [4], [5], [19], [23], [24]. Although there are suggestions that the organization of orientation columns may have low spatial frequency components, hemodynamics may also contribute to this effect. Going in the opposite direction, as voxels decrease in size, they must eventually become smaller in linear extent than the hemodynamic response, and thus become highly interdependent. Recent studies [20], [25]–[27], have highlighted how BOLD responses involve active changes in cortical vasculature, and hence reflect their mechanical and other spatiotemporal response properties, with spatial scales that are at least partly distinct from the scales of the underlying neuronal activity [20].
Given the above points, the mapping between neuronal activity and the spatially extended BOLD response cannot be assumed to be a spatially local temporal convolution [20], but should rather be treated in a comprehensive framework that accounts for both spatial and temporal properties and their interactions. A recent theoretical approach [25] treats cortical vessels as pores penetrating the cortical tissue and draws on a rich framework of methods developed in geophysics [28] to derive a physiologically based approximation of the hemodynamics. This model is expressed in terms of a closed set of dynamical equations that reduces to the familiar balloon model [25] in the appropriate limit where spatial effects are averaged over each voxel. It analyzes the spatiotemporal hemodynamic response by modeling the coupled changes in blood flow, pressure, volume, and oxygenation levels that occur in response to neural activity.
The objective of the present work is to make and empirically test novel predictions of the model, focusing particularly on spatiotemporal dynamics. We first predict the quantitative spatiotemporal hemodynamic response function (stHRF) for physiologically plausible parameters. We find that the model predicts a local response and damped traveling waves, whose speed and range are potentially observable with current high resolution fMRI. Second, we acquire and characterize such high resolution fMRI data from subjects viewing a visual stimulus designed to excite spatially localized neuronal activity in primary visual cortex. We observe hemodynamic waves in these experimental data, whose characteristics confirm our theoretical predictions of wave ranges, damping rates, and speeds, and constrain the physiological parameters of the model.
Cortical hemodynamics and the resulting BOLD signal are modeled by incorporating the physiological properties of cortical vasculature into the theory of fluid flow through a porous elastic medium. The pores are the dense elastic cortical vasculature that penetrate the bulk cortical tissue [25]. In response to a rise in neural activity, local arterial inflow increases, deforming surrounding tissue and thus exerting outward pressure on neighboring tissue. The model predicts coupled dynamical changes in pressure, blood volume, and deoxyhemoglobin (dHb) content in the two-dimensional sheet comprising the cortex and its vascular layer (Figure 1). We refer the reader to the Methods for full details of the model.
Here we linearize this model and derive the stHRF, which is the BOLD response due to a spatially point-like, brief increase in neuronal activity. The main stages of this response are as follows: An increase in neuronal activity z(r,t) occurs, as a function of time t and position r on the cortical sheet. This causes relaxation of the smooth arterial muscles (mediated by astrocytes), inducing an increased influx of blood (Figure 1A) increasing local vessel volume and pressure. Blood then flows to regions of lower pressure, resulting in a redistribution of pressure through the medium (Figure 1B). These pressure changes induce further blood volume changes in adjacent cortical tissue. As a consequence of conservation of mass and momentum, this results in local changes of blood velocity in this adjacent tissue (Figure 1C). As these coupled changes propagate outwards, the viscous properties of blood lead to damping of their amplitude (Figures 1C, D). This is accompanied by outflows to draining veins, reducing vessel volume, and thus yielding further dissipation. Although these two sources of dissipation occur at small scales, they are reflected in the larger scale dynamics of the system.
Together, the above processes result in the propagation of pressure changes that travel beyond the ∼1 mm range of direct blood flow between adjacent arterioles and venules. In our theory, the time required for changes in directly coupled adjacent tissue to occur is comparable to the time it takes for the blood to transit the gray matter (∼1 s). Consequently, these considerations predict propagation speeds of order 1 mm/s.
During the change in vessel volume, local deoxyhemoglobin (dHb) content also changes (Figure 1D). The influx of arterial blood increases the content of oxygenated hemoglobin (oHb), hence reducing local dHb concentration. Further reductions occur by the removal of dHb by local outflow. As this process occurs, oxygen diffuses passively into the cortical tissue, converting oHb to dHb. Together, these processes result in an initial decrease of local dHb concentration followed by a delayed increase.
The BOLD signal thus reflects the net change of blood volume and dHb content. In the case of a spatiotemporally localized neural activation, the predicted BOLD response is given by the stHRF, expressed in Eqs. 1–5 of the Methods. The response to a more general stimulus is obtained by convolving the stimulus with the stHRF, as discussed in Text S1. Critically, the predicted stHRF (Eq. 5) implies that the hemodynamic response contains a component that propagates as damped traveling waves over spatial scales potentially far greater than those of the neural signal that generated them. In other words, our model predicts that even if neuronal activity is restricted to a very small patch of cortex, it will cause changes in the BOLD signal that propagate for several millimeters over a few seconds.
The precise quantitative properties of the predicted BOLD signal depend on several key physiological parameters that can be experimentally determined. Our analysis (Figure 2) indicates that the average vascular stiffness and the rate of damping due to blood viscosity and outflow at boundaries are the most critical parameters (see Table S1 in Text S1 for a complete list of parameters). High stiffness results in a rapid return to equilibrium, thereby increasing the wave speed and range. Conversely high blood viscosity results in strong damping, thereby reducing the range of the waves. Figure 2 shows a range of predicted spatially extended BOLD responses spanning physiologically realistic ranges of these two parameters (Text S1). A combination of strong damping and low stiffness (Figure 2, top left panel) is predicted to lead to localized responses whose spatial extent is mostly confined to that of the neuronal signal. Although there would still be some weak signal propagation, it is unlikely this would be detectable given typical levels of measurement noise and voxel sizes. This parameter set corresponds to cases where the stHRF spatial scale is smaller than a typical voxel size, where the approach of treating voxels independently would be justified.
The opposite extreme of weak damping and high elasticity (Figure 2, bottom right) yields predicted responses that propagate rapidly and far across the cortical sheet. These parameters are unlikely to be relevant experimentally because such extensive waves would likely have already been reported. Between these two extremes there is a broad region of physiologically plausible values of these parameters (medium damping and/or vascular stiffness) for which traveling hemodynamic waves are predicted to have properties that are potentially detectable in current experiments but with ranges that would not likely have led to their detection to date.
To test the predictions for the theory, high-resolution fMRI data were acquired in primary visual cortex, V1, from four healthy subjects. The well-known retinotopic mapping of the visual field to V1 allowed us to design simple visual stimuli that resulted in a spatiotemporally localized neural response [29]–[32]. Subjects viewed visual stimuli consisting of three dashed, time-varying (4 reversals of the light and dark dashes per second; i.e., a 2 Hz cycle) concentric rings at eccentricities of 0.6°, 1.6°, and 3° (Figure 3A). Subjects were instructed to focus at the center of the screen and perform a simple fixation task (see Methods).
These concentric visual stimuli resulted in strong BOLD modulations in early visual cortex, as seen in the example in Figure 3C. As a result of the retinotopic projection from visual field to early cortex (Figure 3B), these concentric rings are projected to lines on the cortical surface, one for each eccentricity. We find that the maximum BOLD modulation occurs on these lines, and that signal modulations weaken away from these peak responses. The three rings were presented at the same time. However, since the hemodynamic responses of the last two rings were not sufficiently separated on the cortex (in all subjects), we focus on the responses to the most inner ring, which was clearly separated (see Methods) from the responses of the other rings. The spatiotemporal evoked response is shown in a series of snapshots at different times t with respect to the stimulus onset (Figure 3C). Early responses (t = 2.5 s) are restricted to the central region (red) on these surface patches. As time progresses the BOLD signal near the center rises, and locations successively further from the center demonstrate increasingly delayed rises at t = 2.5–7.5 s. Finally, outward propagation of positive modulation in the periphery continues, while central responses decrease until they display the well known negative phase of the local post-stimulus undershoot at t = 12.5–17.6 s.
The outward propagation (Figure 3C), discussed in the previous section, suggests the possibility of a traveling wave response, propagating normal to the centerline of the underlying neuronal response. We now focus on characterizing this response. Because stimulation of an isoeccentric curve in the visual field excites an approximately straight line of neurons in V1 [33], normal directions are clearly identifiable on a flattened cortex. We thus estimate the centerline of the primary isoeccentric response in a flattened representation of V1 (Figures 4A,B), and average the signal change over all points the same distance from this centerline (Figure 4B) (see Methods). Repeating this at various distances x and times t reveals the average spatiotemporal response (Figure 4C). Although the signal has been low-pass filtered in the time domain, and averaged along the direction parallel to the stimulus centerline, it is crucial to note that no spatial smoothing has been performed in the x direction.
The response has two characteristic spatial scales. Near the center (|x|<1 mm), a local response occurs, with a range similar to that of the expected neural response and whose time variation is similar to that predicted by the (purely temporal) balloon model. However, outside this region, the response propagates outward for several mm, with the peak response occurring steadily later as |x| increases (Figure 4D). At |x| = 5 mm the response is delayed, reaching its peak just as the central response |x|<1 mm reverses sign (Figures 4D,E). Furthermore, the amplitude of the propagating response decreases with |x| until it reaches the background level beyond |x| = 5 mm.
The above features of the BOLD signal modulations confirm the qualitative theoretical predictions of the spatiotemporal model. The theory also makes quantitative predictions of the waves' propagation speed, range, and damping rate. We next test these predictions and estimate the corresponding parameters through more detailed quantification of the empirical response.
To estimate the speed of wave propagation, phase fronts of the BOLD signal were estimated (see Methods). As the hemodynamic disturbance propagates, the spatially dependent time delay is evident (Figures 3B, 4D). This delay also appears as a change in the phase of each frequency component of the response. Analysis of the instantaneous phase, at the frequency of maximum response (0.1 Hz), confirms propagation of phase fronts (Figure 5A). Points on the phase front that intersects the peak of the BOLD activity at x = 0 are overlaid on the spatiotemporal response in Figure 5B. This highlights the different behaviors of the nonpropagating local (|x|<1 mm) and propagating (|x|>1 mm) components of the BOLD response.
As depicted in Figure 5, our analysis shows that: (i) Near the centerline there is a localized response at approximately |x|<1 mm, corresponding to the spatial scale of the expected neural response [34], including the lateral spreading of thalamocortical projections from the lateral geniculate nucleus to V1 [35]. (note that all that is required to estimate the properties of the waves, in the following analysis, is that the central region, i.e. Δx, be small compared to the propagation distance of the waves so that we separate the propagating from the local component). (ii) Propagation occurs away from the center at a roughly constant speed (constant slope of the phase front) in both directions. Straight-line fits towards the fovea (F, red) and toward the periphery (P, black) yield propagation speeds vF = 2.3±0.2 mm s−1 and vP = 1.8±0.2 mm s−1 (s.d.), respectively for Subject 1 (Figure 5B). (iii) The signal is attenuated as it propagates, with fits to the log-linear plots (Figure 5C) yielding spatial damping constants KP = 0.33±0.02 mm−1 and KF = 0.39±0.02 mm−1 (s.d.). Characteristic ranges are the reciprocals of these constants; i.e., about 3 mm for the signal to decrease by a factor of e. Equivalently, the fits vs. t in (Figure 5D) imply temporal damping rates ΓP = KPvP = 0.56±0.04 s−1 and ΓF = KFvF = 0.86±0.08 s−1 (s.d.).
Data of sufficient quality to enable segmentation and primary response identification were obtained in seven of the eight available hemispheres. Data from the left hemisphere of one subject contained signal drop-out and artifact, most likely due to head movement, which prevented artifact-free surface-based reconstruction. Clear evidence of propagating waves in BOLD signal was observed in all seven of these usable data sets (Figure S5).
A total of 14 sets of wave responses were found, from which parameter estimates were able to be made for 12 (Figure 5E). Two cases of propagation toward the periphery in one subject showed interference from the second stimulus ring and were not used. These 12 responses yielded group averages v = 4±2 mm s−1, K = 0.28±0.03 mm−1, and Γ = 0.8±0.2 s−1 (s.e.m). The scatter plot (Figure 5E) shows a correlation (R2 = 0.22) between the temporal damping rate and the velocity. As discussed above, the ratio between these two quantities yield secondary estimates for spatial damping and are consistent with the substantially smaller relative uncertainty in the spatial damping constant. Furthermore, this shows that the spatial extent is relatively fixed regardless of the wave properties. This is consistent with the observation that the mean FWHM of the responses is clustered around 4.7±0.4 mm (s.e.m).
To recap, our biophysical theory shows that spatiotemporal hemodynamic responses obey a wave equation, whose key parameters are the wave velocity and a damping constant affecting decay in time and space. These parameters can be estimated from the empirical responses using simple regression analyses. We find that the parameter estimates here all lie within the a priori ranges estimated independently of the model (Text S1).
The increasing resolution of functional MRI and the development of analysis methods that depend on spatial patterns in these data underline the need for a systematic, quantitative approach to the spatial and temporal properties of the BOLD signal that is based on the properties of the underlying tissue and vasculature. Here we apply a recent physiologically based theory of hemodynamics in cortical tissue and vasculature to derive the linear spatiotemporal hemodynamic response function (stHRF), which is the response to a spatiotemporally localized stimulus of moderate amplitude. High resolution fMRI data are then used to test the predicted BOLD response to localized neuronal modulation in early visual cortex. Just two extra measurable parameters – vβ and Γ - suffice to characterize the spatial properties of the response.
The theory used is the first to make a mean-field approximation to cortical vasculature and goes beyond spatially point-like multicompartment models [12], [36], [37] as it allows calculation of spatiotemporal hemodynamic responses to general stimuli. It predicts, and the data demonstrate, that the hemodynamic response to a spatially highly localized neuronal drive exhibits traveling waves that propagate over a few mm of cortical tissue. Moreover, the velocity, damping and characteristic range of the observed waves are well within the range of theoretical predictions. These traveling waves have not been previously predicted or reported in human cortex. The central part of the response is non-propagating and has a temporal profile consistent with the standard balloon model [11]–[14], [38].
A further key implication of the spatial effects in our data is that fMRI voxels should only be treated independently if each voxel is larger than the stHRF scale. For interpreting fMRI acquired at sufficiently high spatial resolution the spatiotemporal properties of the stHRF must also be taken into account. Moreover, when voxels are small, we speculate that propagation of hemodynamic waves beyond their boundaries may underlie the observation that some experiments can be sensitive to structures at scales below the voxel size, including ocular dominance and orientation preference columns [4], [5], [24].
The combination of modeling and data allows us to estimate key physiological parameters of the model from observations of individual subjects. This lays the basis for replacing fMRI analysis procedures that rely on purely empirical analysis by ones that relate to the underlying physiology. We have shown how characterization of the spatiotemporal properties of fMRI data allows properties of the cortical tissue and vasculature to be inferred, hence accounting for differences between subjects and, potentially, brain regions. For example, relatively low blood viscosity and/or high tissue stiffness are predicted to lead to longer-range wave propagation. Specific experimental manipulations, such as the use of blood-thinning agents, could be employed to test the predicted changes in wave speed and spatial range. Similarly, the reduction in tissue elasticity that typically occurs with ageing [39], should be able to be probed noninvasively via its effects on wave velocity, and thereby taken into account when making inferences about neuronal activity in cohorts where age may be a confound. Likewise, regionally specific vascular properties have recently been highlighted as an important potential confound in studies of effective connectivity [40], [41], thereby underlining the need for a careful measurement and allowance for hemodynamic effects.
It is worth asking why hemodynamic waves have not been previously observed in fMRI. Some reasons are: (i) If voxel dimensions are large and sampled over a long time period, the hemodynamic response is not sufficiently resolved to detect propagating waves (ii) If the BOLD signal is spatially smoothed, then the spatiotemporal structure of the measured BOLD signal will be averaged out; (iii) Wave propagation is confined to occur within the cortical sheet and will be only be readily apparent in surface-based data reconstruction; (iv) Hemodynamic waves from a point source (e.g., a localized activation in a typical study) in two spatial dimensions decay more rapidly with distance than from the line source our experiment, where net decay can occur only in the direction perpendicular to the cortical locus of our one-dimensional stimulus. Despite these points, as high resolution protocols and surface-rendered data analysis techniques gain widespread use, the need for quantitative spatial analysis will likewise grow.
An important consequence of having hemodynamic traveling waves is that the spatial dynamics of BOLD are not independent of their temporal dynamics. The conventional factorization into spatial and temporal convolution operators is thus not valid in general. A greater understanding of the BOLD response, and brain mapping in general, would come from understanding the spatiotemporal hemodynamic response [20]. The present spatiotemporal HRF provides a solution to this problem, starting from a theory of spatiotemporal hemodynamics.
Several other issues arise from having hemodynamic traveling waves. (i) The existence of hemodynamic waves mean that spatiotemporal hemodynamics, induced by nearby sources, can interact in a nontrivial way, a property that occurs in the temporal domain, concerning the nonlinear interaction between temporally proximate responses [13], [38]. (ii) These findings cast further doubt on those measures of effective connectivity, such as Granger causality, unless they include a careful treatment of hemodynamic effects [40], [42], [43]. (iii) On the other hand, experimental designs could exploit the wave properties of hemodynamics by using stimuli that induce resonant properties of cortical tissue - akin to the temporal domain [14]- enhancing detection of the evoked signal.
Traveling waves of neuronal or glial origin have been described throughout the brain, including in visual cortex [43]–[46], raising the question of whether these waves might be responsible for the hemodynamic traveling waves seen in our data. However, several considerations argue against this: (i) The close match between the theoretically predicted values and the observed data strongly supports the conclusion that the waves in our data are of hemodynamic origin. (ii) Previous studies [45], [47], [48] that reported propagating neuronal waves in V1 of similar spatial extent to those seen here demonstrated that these waves are 1–2 orders of magnitude faster (approximately 200 mm s−1 in cats [47], 100–250 mm s−1 in primates [45] and 50–70 mm s−1 in rats [48]), and waves in cortical white matter travel even faster [43]. Likewise, although the spatial scales may be similar to those presently reported, the diffusion of nitrous oxide - which mediates the coupling between neuronal activity and vasodilation - occurs too rapidly to explain our results [21]. (iii) Another possible source of propagating signal of possible relevance are calcium waves traveling via astrocytes because these mediate the neuronal signal in vasodilation. However, these calcium waves travel at ∼10 µm s−1 [49], which is 2 orders of magnitude slower that the waves reported here.
Although hemodynamic waves have not been characterized, detected, or previously modeled, existing work has detailed some spatiotemporal properties of the BOLD response. Previous studies have demonstrated hemodynamic contributions to spatiotemporal BOLD response, including: the effect of draining veins [50], [51] which induces a latent BOLD signal due to these veins; effects across vascular layers [52], [53] that induce layer dependent delays of the BOLD response; and general effects of the vascular network [43] that cause delayed BOLD responses across extensive brain regions. Studies have also implemented ways to minimize such effects to improve spatial specificity of functional activations [43], [50]. The hemodynamic waves are different from the mentioned phenomena in that they exhibit propagation across the cortical surface. As the waves pass through they induce changes in arterioles, capillaries, and venules – not reliant on overall drainage by large veins. The possible interplay of these effects will be subject to future modeling and experimental work.
In summary, with advances in imaging technology and data analysis, intervoxel effects will become more pronounced, demanding spatiotemporal analyses based on the underlying brain structure and hemodynamics. By verifying a model that enables such analysis, the present paper opens the way to new fMRI probes of brain activity. These new possibilities include experiments using spatial deconvolution to discriminate between neural and hemodynamic contributions to the spatiotemporal BOLD response evoked by complex sensory stimuli. An important potential application would be to disentangle negative components of the BOLD response from surround inhibition in the visual cortex. Our analysis also affords novel insights and physiological information on neurovasculature, a subject of particular significance to ageing and vascular health. Finally, the combination of the present stHRF with spatially embedded neural field models [8] would allow a systematic and integrated computational framework for inferring dynamic activity in underlying neuronal populations from fMRI data.
The methods used in this study are threefold: (i) Derivation of a theoretical spatiotemporal hemodynamic response function (stHRF) from a physiologically based model of cortical hemodynamics; (ii) Execution of a customized experimental protocol for the acquisition and preprocessing of high resolution fMRI data on the response to spatially localized visual stimuli; and (iii) Model-based analysis of these data and explicit comparison with the predicted stHRF, including inference of underlying cortical hemodynamic parameters.
The theoretical prediction of the stHRF is derived from a physiologically-based model for spatiotemporal hemodynamics [25]. This model treats brain tissue as a poroelastic medium, with interconnected pores representing the cortical vasculature. The governing equations are a set of nonlinear partial differential equations that connect blood flow velocity v, mass density contributed by blood (i.e. the part of the total density contributed by blood as opposed to tissue) ξ, deoxygenated hemoglobin concentration Q, and blood pressure P due to an increase in arterial flow F caused by an increase neural activity z, as a function of time t and position on the cortex r. Although we describe changes in blood volume throughout the text, we model changes in ξ - which is closely related to the fractional volume of blood in tissue: ξ/ρf, where ρf is the density of blood itself. These model equations were recently derived and explained in a separate paper [25]. We provide a synopsis here and apply this with appropriate boundary conditions (Text S1) to calculate the stHRF and derive a hemodynamic wave equation.
The dynamics of flow F(r,t) are modeled as a damped harmonic oscillator [13] driven by neural activity z(r,t): (1)The dynamics in Eq. 1 are parameterized by the signal decay rate κ, the flow-dependent elimination constant γ and the resting flow F0. The neural activity z(r,t) drives a distribution of arterial control sites (see Figure S1), as described further in Text S1.
The vascular response due to this increase in arterial flow is then constrained by physical laws, including conservation laws. Firstly, the conservation of blood mass is embodied by(2)where cP is a proportionality constant. This conservation law describes how the rate of change of local blood mass density ∂ξ/∂t is determined by the local divergence of the flow ρf∇•(v), the source of mass due to the average inflow of blood F, and the average venous outflow of blood cPP. These latter source/sink terms are mean-field terms that describe the average spatiotemporal hemodynamic processes (For more details see the Text S1 on the derivation of these terms).
The rate at which blood travels through the vasculature depends on the elastic response of cortical vessels. This process must conserve momentum, expressed as(3)where P is the average pore pressure, D parameterizes damping due to blood viscosity, and c1/ρf is the constant of proportionality between pressure gradient and acceleration in the porous medium. This equation describes how forces are directed down pressure gradients, causing blood to accelerate toward regions of lower pressure. These velocity changes are resisted by blood viscosity leading to the resistive term D(v-vF - vP) where vF and vP are the blood velocities at inflow and outflow, respectively (Text S1). The average pressure is related to the elastic properties of blood vessels by the constitutive equation, (4)where the elasticity of blood vessels is parameterized by the Grubb exponent 1/β (see Table S1 in Text S1) and a proportionality constant c2, as in previous empirical studies of cerebral blood flow [54].
As changes in local blood volume occur, oxygen diffuses into cortical tissue because of the increased partial pressure of oxygen. This process produces blood deoxyhemoglobin (dHb) - whose concentration is represented by Q - from oxygenated hemoglobin. The local concentration of hemoglobin in tissue is a fixed proportion ψ (in mmol kg−1), of local blood density in tissue, and is thus expressed as ψξ. Hence, the difference ψξ - Q is the amount of oxygenated hemoglobin. If η is the fractional rate at which oxygen passes from oxygenated hemoglobin to cortical tissue, the flow of dHb obeys a conservation equation, similar to Eq. 2 for the conservation of blood mass:(5)where the difference term (ψξ - Q)η on the right hand side introduces the source of dHb, in which dHb is convereted to oHb at a rate η, and the term –QPcP/(ξρf) represents the rate of reduction of dHb concentration due to blood outflow. This assumes that the blood is well mixed so the concentration leaving a vascular unit is Q/(ξρf) at a rate of average venous blood outflow cPP (Text S1). As in Eq. 2, the net outflow rate, here dHb ∇•(Qv), is balanced by local changes in content, here ∂Q/∂t.
Finally, the measured BOLD signal y is predicted by a recent semi-empirical relation [55] between tissue blood volume content ξ/ρf and dHb content Q to be(6)where V0 is the resting total volume, Q0 is the resting fraction of dHb, and the constants k1, k2, and k3 depend on the acquisition parameters, including field strength and echo time.
By restricting the fMRI scans to the occipital pole, we achieved a resolution of 1.5×1.5×1.5 mm3 and 2 s TR echoplanar images (EPI). The stimulus onset was dephased by 250 ms per block for 8 blocks to further increase the effective time resolution. These EPI data were then coregistered to a high-resolution T1-weighted anatomical scan, acquired at 0.75×0.75×0.75 mm3, so that the spatiotemporal resolution was effectively resampled to 250 ms×(0.75×0.75×0.75) mm3. Furthermore, these mappings were restricted to the gray matter by segmenting the anatomical data into gray and white matter. Finally, functional retinotopic scans were used to map out the expected cortical positions in the visual cortex of each subject [3]. Data were acquired on a Philips 3 T Achieva Series MRI machine equipped with Quasar Dual gradient system and an eight-channel head coil.
Five healthy subjects (two female) ranging from 21 to 30 years participated in this study. The study protocols were approved by ethics boards of the University of New South Wales and Neuroscience Research Australia (formerly the Prince of Wales Medical Research Institute).
Participants viewed visual stimuli via a mirror mounted on the head coil at a viewing distance of 1.5 m, resulting in a display spanning a diameter of 11° (or 5.5° in eccentricity). The visual paradigm was prepared with Presentation® software. Stimulus duration and fMRI pulse timing were logged with 0.1 ms accuracy. The stimulus consisted of 3 concentric rings simultaneously presented at 0.6°, 1.6°, and 3° eccentricity, presented in a block paradigm. (on for 8 s and off for 12.25 s). Each annulus was only 1 pixel wide (roughly 0.014° visual angle).
As known from retinotopic studies, isoeccentric lines in the visual field map to approximately straight lines in primary visual cortex. This stimulus was chosen to exploit this property, optimizing the identification of the primary response and secondary changes in BOLD signal in the orthogonal direction.
During the ‘on’ state, the annuli were divided into black and gray dashes that reversed roles 4 times per second (i.e., in a 2 Hz cycle). During the ‘off’ state, the annuli remained black (Figure S2). To improve visual fixation, a black fixation cross extended across the entire screen, 4 black circles were permanently present, and a pseudorandomly flickering fixation dot fluctuated between red, green, and blue [56]. Subjects reported that they were able to maintain alertness and attend to the fixation cross throughout data acquisition.
Note that off blocks were 12.25 s in duration, ensuring that the evoked response was effectively sampled at 250 ms. Each fMRI session consisted of 8 stimulus blocks, consisting of 80+1 fMRI volumes (the extra scan was due to the delayed onset) plus an additional 7 ‘off’ scans prior to the first block. Hence each session contained 88 fMRI volumes, a running time of 176 s, 14 such sessions were acquired from each subject.
To improve timing accuracy and synchronization we used a monitor refresh rate of 60 Hz for the visual display, and a TR of 2006 ms, rather than 2000 ms, to compensate for system delays. The remaining variability of the stimulus onset precision was logged and used for modeling the experimental design during data analysis.
To achieve high resolution, speed, and minimize distortions, we used a SENSE [57] accelerated echoplanar imaging (EPI) sequence. Great care was taken to minimize distortion, and each subject's data were carefully investigated to ensure distortion was minimal. Functional data were acquired in 29 1.5 mm slices for all but one subject, for whom there were 28 slices, with a 192×192 matrix, 230 mm field of view, and a SENSE factor of 2.3.
Functional data were motion corrected and slice scan-time corrected using SPM5 (SPM software package, http://www.fil.ion.ucl.ac.uk/spm/), then imported into the mrVista- Toolbox (http://white.stanford.edu/software/) for further processing and analysis. The fMRI data were transformed and analyzed in three different spaces: Firstly, the original planar space of the data acquisition, secondly, the 3 dimensional space defined by the high resolution T1 anatomy scans into which the data were aligned, transformed, and spatially up-sampled and finally for a flattened representation of the visual cortex, with maps of phase of the fMRI signals at the left and right occipital pole. Apart from the spatial up-sampling and mapping, no further preprocessing of the data in the spatial dimension was performed. The temporal time series of each voxel were low-pass filtered with a third order Butterworth filter below 0.1 Hz. Furthermore, although there were three concentric rings, we focus only on the 0.6° ring closest to the fovea when analyzing the spatiotemporal hemodynamic response as this was clearly spatially distinct whereas there was some overlap in responses to the furthest two rings (1.6°, and 3°).
Distance measurements were made along the cortical surface using meshes generated by the segmentation on each subject. A shortest path algorithm (in the VISTA software) was used to determine these distances on the surface.
When a stimulus excites a line of cortex, as in the present case, the hemodynamic response depends only on time and the perpendicular distance x from that line. To analyze these dependences, we estimated the location of the centerline of the primary response on the flattened surface, then measured the average BOLD signal at various distances orthogonal to this response as a function of time since stimulus onset. This was achieved in five steps (see Text S1 for further details):
As a hemodynamic disturbance travels, the BOLD signal phase depends on x and t. Phase fronts enable any wave propagation to be tracked at large |x|. To obtain phase estimates from the signal y(x,t), we first constructed the analytic signal [59],(8)where(9)is the temporal Hilbert transform [59]. The phase φ(x,t) is then given by(10)where arg is the complex argument. Maps of this phase are shown in Figure 5A as well as in the Supplementary text for all the data sets. Constant-phase lines represent the phase fronts of the BOLD signal.
The empirical estimates for the properties of the wave fronts were calculated from the phase fronts emerging from the peak of the BOLD signal at x = 0 (Figure 5B). From here, the two principal characteristics of the spatiotemporal response can be identified, the local response, close to x = 0, as a region of near-uniform phase spanning |x|<1 mm, and the propagating component heading away from these regions at |x|>1 mm (see Figure 5B and Text S1). This is consistent with the expected neural point spread function estimated from independent physiological data [34]. Straight-line fits to this propagating region, as shown in Figure 5B, yielded estimates of the wave velocity νβ.
The BOLD signal was measured at each point on the phase front as a function of time (Figure 5C) and space (Figure 5D). Transformation to logarithmic scales yielded approximately straight line plots, suggesting exponential decay of BOLD signal in space and time. Linear regression then yielded rate constants of temporal and spatial signal decay. Estimation of the standard error of these linear regressions provided error estimates for these parameters (see Text S1).
|
10.1371/journal.pgen.1006700 | Probing the canonicity of the Wnt/Wingless signaling pathway | The hallmark of canonical Wnt signaling is the transcriptional induction of Wnt target genes by the beta-catenin/TCF complex. Several studies have proposed alternative interaction partners for beta-catenin or TCF, but the relevance of potential bifurcations in the distal Wnt pathway remains unclear. Here we study on a genome-wide scale the requirement for Armadillo (Arm, Drosophila beta-catenin) and Pangolin (Pan, Drosophila TCF) in the Wnt/Wingless(Wg)-induced transcriptional response of Drosophila Kc cells. Using somatic genetics, we demonstrate that both Arm and Pan are absolutely required for mediating activation and repression of target genes. Furthermore, by means of STARR-sequencing we identified Wnt/Wg-responsive enhancer elements and found that all responsive enhancers depend on Pan. Together, our results confirm the dogma of canonical Wnt/Wg signaling and argue against the existence of distal pathway branches in this system.
| Our manuscript addresses the question of whether either of the canonical transduction components, beta-catenin or TCF, can be bypassed when the Wnt pathway is activated. By using somatic cell genetics in Drosophila cells (via CRISPR/Cas9 editing) in combination with RNA-seq and STARR-seq (Self-transcribing-active-regulatory-region-sequencing) as functional read-outs, we provide firm evidence against the existence of distal branches in the Wnt pathway.
| Wnt proteins are highly conserved signaling molecules specifying the fate and behavior of cells in multicellular animals ranging from nematodes to humans [1]. They play crucial roles in embryogenesis, pattern formation and tissue homeostasis during development and in adult life. Therefore it is not surprising that aberrant Wnt signaling has been found to be implicated in many human diseases [2].
Following the identification of Wnt proteins nearly 40 years ago [3–5] genetic and biochemical studies have revealed mechanistic details of how the signaling cascade operates when cells receive a Wnt signal [for review see 6]. As a consequence of Wnt/Wg proteins binding their cognate receptors, beta-catenin is no longer marked for degradation and accumulates in the cytoplasm and nucleus [7–10]. In the prevailing model, TCF is targeted through its DNA binding domain to Wnt-responsive elements (WREs) in the promoters or enhancers of target genes [11] and initiates the transcription of Wnt/Wg-responsive genes when complexed with beta-catenin. In the absence of Wnt/Wg ligand, beta-catenin is phosphorylated and degraded while TCF is bound by transcriptional repressors, such as Groucho and Coop [12–15]. In contrast to the well-studied mechanism of gene activation, the mechanisms by which beta-catenin and TCF promote target gene repression are not well understood [16]. Several reports suggest that, in addition to beta-catenin and TCFs, other factors are involved in Wnt-mediated repression, such as Prop1, Mad or Zic [17–19]. Furthermore it is not clear, in which context alternative [20] or traditional TCF binding sites are used for transcriptional repression [21–23].
A recent study showed that TCF4 is a predominant factor in mediating the Wnt response and for recruiting beta-catenin to DNA [24], however ongoing research on the Wnt signaling pathway has repeatedly demonstrated that beta-catenin as well as TCF interacts with various other proteins. Yet it remains to be determined, whether alternative transcriptional complexes also regulate the expression of Wnt/Wg target genes. For example, an interaction between beta-catenin and FOXO-transcription factors in mouse and DLD-1 human colon carcinoma cells has been demonstrated resulting in the activation of genes involved in oxidative stress and colon cancer metastasis [25–27]. Furthermore in mouse embryonic stem cells it was shown that beta-catenin forms a complex with Oct4 to promote Oct4-driven transcription and pluripotency [28]. In addition, studies in Xenopus reported an interaction between beta-catenin and Sox17, promoting expression of Sox17 target genes [29], and more recently it was suggested that beta-catenin complexes with YAP1 and TBX5 in human cancer cell lines [30]. In addition, alternative binding partners have also been reported for TCF, such as Plakoglobin or Mad [31, 18].
In this study, we address the question of whether alternative routes exist that bypass beta-catenin or TCF to promote the transcription of Wnt/Wg target genes in Drosophila cells. Using cells that lack either Arm or Pan and functional read-outs (i.e. RNA-seq and STARR-seq), we show that both, Arm and Pan, are absolutely required for target gene activation and repression. Consistent with these findings, we further demonstrate that Wnt/Wg-responsive enhancers also require Pan, arguing against the existence of distal branches in the Wnt signaling pathway.
Next-generation RNA-sequencing (RNA-seq) was used to identify and quantify the expression of target genes of the Wnt/Wg signaling pathway in Drosophila Kc167 cells. Cells were treated either with Wg-enriched medium (referred to as Wingless-conditioned medium, WCM; [32]), or control-conditioned medium (CM) lacking the Wg ligand. Wg-responsive genes were determined by statistical analysis of gene expression levels in treated samples versus control samples, according to a protocol described in [33]. In order to determine a high confidence set of Wnt/Wg targets, genes had to pass the following selection criteria: exhibit a significantly altered expression profile (WCM vs CM, p-value ≤ 0.0005) and an at least two-fold change of expression upon Wg stimulation (Fig 1A). WCM-treatment resulted in the robust induction of 51 genes. Among them we found previously identified Wnt/Wg target genes such as naked cuticle (nkd), CG6234, frizzled 3 (fz3) and Peroxidasin (Pxn) [34–36, 20], confirming our quality filters. 40 genes were at least two fold up-regulated (positive targets) and 11 genes two fold down-regulated (negative targets) (Fig 1B). 7 positive and 5 negative candidate target genes were confirmed by qRT-PCR (Fig 1C). This high confidence set of Wnt/Wg target genes was used to systematically elucidate potential beta-catenin or TCF-independent branches of Wnt/Wg signaling.
To investigate whether Arm can be bypassed via alternative branches of the pathway, we generated arm knockout cells using the CRISPR/Cas9 technology as described by Bassett and colleagues [37]. In order to generate Drosophila arm null mutant cells we used sgRNAs targeting two different exons that are present in all transcript variants (Fig 2A and 2B). sgRNA-a1 on the reverse strand targets the translational start site residing in exon 2. sgRNA-a2 targets a site in the third exon. The presence of CRISPR-induced mutations generated by NHEJ (non-homologous end joining) was assessed by sequencing of the PCR products spanning the sgRNA target sites (see Material and Methods). The analysis revealed that most of the alleles had indel mutations at the expected cleavage sites, some of which lead to the deletion of the translational start site or to frameshifts in exon 3. To generate an arm-/- cell line, we carried out serial dilutions and searched for cell populations that carried previously identified mutations using allele-specific primers as described in [38]. In this way, we isolated an arm null mutant cell line (named arm-/--AFII7/8) which was a homogenous cell population (see Material and Methods) carrying a deletion of either one or sixteen nucleotides in the second exon, each of them destroys the START codon (ATG), and a deletion of one nucleotide in the third exon (Fig 2B). Importantly no wild-type alleles were present. These mutations, affecting both arm alleles, result in frameshift mutations introducing a premature termination codon that should trigger nonsense-mediated mRNA decay (NMD) [39] (S1A Fig). We confirmed the complete loss of Arm protein in arm-/--AFII7/8 cells by Western blot analysis (Fig 2C, S1B Fig).
Next we investigated whether Arm is absolutely required for the Wnt/Wg-driven transcriptional output. To that end arm-/--AFII7/8 cells were treated either with WCM or CM and target gene responses were monitored by RNA-seq. We found that the induction of the positive Wnt/Wg target genes is dependent on Arm, since their expression was not changed in arm null mutant cells. Similarly all negative target genes are no longer repressed in arm-/--AFII7/8 cells (Fig 3A and 3B). These results demonstrate that Arm is absolutely necessary for both, activation and repression of identified Wnt/Wg targets. We confirmed our results with qRT-PCR analysis of 11 candidate targets genes (S2 Fig).
From the analysis above, we conclude that Arm is absolutely required for both activation and repression of Wnt/Wg target genes and interpret this as evidence against the existence of an Arm-independent Wnt/Wg signaling transcriptional output. Since several alternative interaction partners for beta-catenin have been proposed for the activation and the repression of genes, such as Sox17 [29], Oct4 [28] and Prop1 [17], we next asked whether TCF-independent Wnt/Wg signaling exists. To search for TCF-independent Wnt/Wg signaling, we utilized a similar setup as described above to generate pan null mutant cells. Two distinct sgRNAs were used to target independent loci within the pan gene (Fig 4A). We isolated a population of pan null mutant cells that no longer contain any wild-type allele. Similar to the arm-/--AFII7/8 cells, the selected pan null mutant cells, termed pan-/--AF1AD26, carry three defined mutations that lead to frameshift mutations. Molecular analysis of the alleles revealed no wild-type allele but a large deletion of approximately 9 kb spanning the two selected CRISPR sites (Fig 4B). In addition, pan-/--AF1AD26 cells also harbor two distinct frameshift mutations in the HMG box, both of which result in premature termination codons (S3A Fig) and NMD. Consistent with this qRT-PCR analysis showed a reduction of pan mRNA in knockout cells compared with wild-type cells (S3B Fig). The presence of the three pan mutant alleles suggests that at the pan locus Kc cells are polyploid; segmental polyploidy has been reported for Kc cells [40]. Since no anti-Pan antibodies were available to confirm the absence of functional Pan protein we used the wingful luciferase reporter assay, an artificial built reporter giving a robust and high Wg-response [41]. Consistent with the absence of Pan, in pan-/--AF1AD26 cells the wingful reporter could no longer be induced after WCM-stimulation; responsiveness could be restored by Pan overexpression (Fig 4C and 4D).
To answer the question of whether Pan is dispensable for Wnt/Wg-regulated induction of target genes, we treated pan-/--AF1AD26 cells with either WCM or CM and performed RNA-seq. We observed that pan-/--AF1AD26 cells can no longer transduce the Wnt/Wg signal as expression of none of the identified Wnt/Wg targets was altered. Neither positive nor negative Wnt/Wg-target genes significantly changed their expression profile in pan knockout cells after Wg stimulation providing evidence that Pan is indispensable for the activation and repression of Wnt/Wg target genes (Fig 5A and 5B). The lack of a change in the expression of several candidate Wnt/Wg targets was confirmed by qRT-PCR (S2 Fig).
Like most major developmental signaling pathways, the Wnt/Wg system uses a “transcriptional switch” mechanism to positively regulate target gene expression [42]. In the absence of Wnt/Wg signaling, the transcription of target genes is repressed by Pan via its interaction with co-repressors such as Groucho or Coop [13, 15]. Pan turns into an activator when complexed with Arm following pathway activation. It has been shown that loss of Pan function leads to de-repression of the Wg target genes nkd and CG6234 in the Wg OFF state in vivo and in vitro [35, 43]. To determine whether this mode of action is valid for the entire set of identified Wnt/Wg target genes we compared the gene expression profiles of wild-type and pan-/--AF1AD26 cells in the absence of Wnt/Wg signaling. Interestingly, we found that only a fraction (37.5%) of positive target genes were de-repressed in pan null mutant cells (Fig 5C; fold change ≥ 2; p-value ≤ 0.0005); among them were nkd and CG6234 [35]. We also noted that this set of de-repressed genes is highly induced in the presence of Wg ligand (Fig 5C). In contrast, the absence of Pan had no effect on the basal expression of the other (the majority) target genes. However, we also identified some genes exhibiting reduced levels of expression in unstimulated pan knockout cells (Fig 5C), suggesting that Pan might be required for their transcription in the absence of Wnt/Wg signaling. Blauwkamp and colleagues (2008) proposed this mode of action for Pan in Drosophila Kc cells for several negative target genes, when cells were not exposed to Wnt/Wg [20].
Transcription factors bind to specific signal responsive elements in the promoters or enhancers of target genes in order to regulate their expression [44]. So far we have analyzed in detail the Wnt/Wg-triggered transcriptional output and demonstrated that both, Arm and Pan are absolutely required for the activation and repression of Wnt/Wg target genes in Drosophila cells. However, in order to obtain a more complete understanding of the transcriptional regulation of Wnt/Wg target genes, we carried out Self-transcribing-active-regulatory-region-sequencing (STARR-seq), a genome-wide enhancer activity assay that reveals the identity of DNA sequences that can function as enhancers in a particular cell type [45–46] and in response to external stimuli, such as the insect steroid hormone ecdysone [47]. To identify enhancers whose activity changes in response to the Wnt/Wg signal, we performed STARR-seq under conditions of active Wnt/Wg signaling and under control conditions (Fig 6A, S4A Fig). For technical reasons, we used the Gsk3β-inhibitor CHIR99021 (CHIR)–a widely used alternative inducer of Wnt-signaling to stimulate Wg signaling in the STARR-seq experiments [48, 49] (see Material and Methods), whose activity we compared to WCM by using the wingful reporter (S5A Fig). Furthermore, treatment with CHIR robustly induced expression of known Wg targets in Drosophila cells (S5B and S5C Fig).
Activation of the Wnt/Wg signaling pathway led to robust changes in enhancer activities: we identified 185 STARR-seq peaks (p-value ≤ 0.001) that were at least 3-fold induced in the CHIR-treated versus control sample, and 348 that were at least 3-fold repressed (Fig 6B). Among the induced peaks, 73 (39.5%) were induced more than 5-fold and 32 (17.2%) more than 10-fold (Fig 6C). We found several enhancers, which have already been described as WREs in Drosophila Kc cells. For instance we identified two enhancers close to the TSS of nkd (first intron and 10 kb upstream of TSS) (Fig 6D), the well-studied WRE 2.2 kb upstream of the TSS (transcription start site) of Notum, an enhancer 15.2 kb upstream of pxb and an element in the 5’ intergenic region 178 bp upstream of Ugt36Bc [50, 43, 20] (S4B Fig). We validated activated and repressed STARR-seq enhancers in luciferase reporter assays as described in [47]. Consistent with the STARR-seq results, we found luciferase reporter activities responded as expected to both CHIR treatment and WCM treatment: increased activity for activated enhancers and decreased activities for repressed enhancers (S4C and S6 Figs). Taken together, these results indicate that the activities of STARR-seq detected enhancers are modulated by Wnt/Wg signaling.
To further test that the identified enhancers were directly regulated by Pan, we assessed the enrichment of known transcription factor motifs [46] in Wnt/Wg-responsive STARR-seq enhancers in comparison to negative control sequences (see Material and Methods). The known TCF/Pan motif [51] (Fig 6E) was strongly enriched in induced enhancers (2.7-fold enrichment, p-value = 1.3x10-8), whereas it was not enriched in constitutive or repressed enhancers (p-value = 0.27 and p-value = 0.08, respectively). Using de novo motif discovery (see Material and Methods) we found an additional Helper site motif in induced enhancers (GCCGCC, p-value = 3.4x10-14; Fig 6E), which is a GC-rich element near TCF/Pan binding sites that is critical for Wnt/Wg target gene activation [52–53, 11]. To experimentally validate the necessity of the TCF/Pan motif for Wnt/Wg induced enhancers, we tested wild-type and mutated versions of the TCF/Pan motif in 3 enhancers of the odd, how and lbe genes in luciferase assays. While the wild-type enhancers activated luciferase reporters 31-, 11- and 7-fold after Wnt/Wg induction by CHIR treatment, the Pan motif-mutant sequences did not respond to treatment (<1.2-fold induction), a substantial and significant difference in each case (p-value≤0.01; Fig 6F), indicating that at least these 3 Wnt/Wg-responsive enhancers require the TCF/Pan motif.
Given the enrichment of the TCF/Pan motif in the Wnt/Wg-responsive STARR-seq enhancers and the necessity of this motif for enhancer function, we next examined whether Wnt/Wg-responsive enhancers require the Pan protein. We repeated the STARR-seq experiments in pan null mutant cells (S7A Fig) and again confirmed our findings for a subset of the enhancers by treatment with WCM (S6 Fig). Consistent with our analysis of target gene expression by RNA-seq, we found that enhancer-induction was overall strongly reduced from 26.1-fold the highest induction in wild-type cells to at most 3.8-fold in pan null mutant cells and that the vast majority (80%) of Wnt/Wg-induced enhancers no longer responded to pathway activation (Fig 7A). For example, the enhancers in first intron and 10 kb upstream of TSS in the nkd gene locus that were strongly induced in wild-type cells by Wnt/Wg signaling were not any more induced nor detected in pan knockout cells (p-value>0.001, Fig 7B). We confirmed these findings by testing several of the most strongly activated enhancers in luciferase reporter assays. In agreement with the STARR-seq results, enhancers that were strongly activated by Wnt/Wg signaling in wild-type cells did not respond to Wnt/Wg pathway activation in pan knockout cells (S7B Fig). Taken together, these results argue that Pan is required for the activation of Wnt/Wg-responsive enhancers.
According to the generally accepted dogma the canonical Wnt signaling pathway culminates in the transcriptional induction of target genes via the beta-catenin/TCF complex. During the past decade, several alternative configurations of the Wnt pathway have been proposed in which either beta-catenin or TCF is bypassed. A recent study explored the co-occupancy of TCF4 and beta-catenin using ChIP-seq and showed that TCF4 is the major factor in tethering beta-catenin to DNA [24]. However, the study could not exclude the possibility that other putative factors could compensate the lack of TCF or beta-catenin–an aspect that is still poorly understood in the field of Wnt research. In the present study, we investigate whether and, if yes, to which extent a Wnt response can bypass beta-catenin or TCF. To this aim we used somatic cell genetics in Drosophila cultured cells. As a basis for our analysis, we first carried out a systematic and genome-wide study to explore all Wnt/Wg-related transcriptional outputs in this system. We identified a set of 51 genes that are induced upon Wg stimulation. To probe whether their expression requires Arm or Pan, we generated cells lacking one or the other of these factors using the CRISPR/Cas9 technology. Surprisingly, we found that Arm and Pan are both absolutely required for all Wnt/Wg-related transcriptional outputs in this system. As a transcription factor, Pan binds to DNA regulatory elements up- or downstream of the TSS of its target genes. Thus, next we asked, whether these DNA regulatory elements–enhancers/repressors–are dependent on Pan using STARR-seq. Impressively, consistent with our RNA-seq analysis, we found that the induction of Wnt/Wg-responsive enhancer elements fully depends on Pan.
In our work we identified eleven down-regulated target genes and showed that knockout of Arm or Pan is sufficient to abrogate their repression. We observed the same effect for repressed enhancers in pan null mutant cells. These findings are in line with a previous study in Drosophila Kc cells [20], in which it was shown that Pan and Arm are required for the repression of the negative target genes Pxn, Ugt36Bc, Tig and Ugt58Fa [20]. We also found Pxn in our Wnt/Wg target gene set. However, the other genes were less than 2-fold repressed in our system and thus did not pass our selection criteria. This might be due to technical differences in Wnt-pathway stimulation and/or timing. Blauwkamp and colleagues showed also in their study that the negatively regulated targets exhibited lower expression upon Pan reduction in the Wnt OFF state [20], implicating that Pan normally activates their expression even in the absence of Wg ligand. When analyzing our data, we found that only half of the negative target genes appear to be activated in the Wnt OFF state upon Pan abrogation, the remaining targets did not exhibit a significant change in their expression profile. This suggests that they might be indirect targets or independent of Pan. Furthermore, we found that several repressed enhancers possess neither the traditional TCF/Pan binding motif, nor the previously reported alternative binding site important for repression, indicative for a Pan-dependent indirect regulation of repressed enhancers. It is likely that Pan is tethered to the DNA by other co-factors as it was shown for dpp or CDH1 [21, 23]. Thus, these Pan-dependent enhancers without any known TCF/Pan binding site provide a good starting point for further molecular studies to gain insight into the still incomplete model of Wnt-mediated repression [16].
In sum our results demonstrate that all Wnt/Wg-related transcriptional output in Drosophila cells requires Arm and Pan and that the induction of Wnt/Wg-responsive enhancers is fully dependent on Pan. Hence, collectively our data argue against the existence of distal branching of the Wnt pathway in this system.
Drosophila Kc167 cell lines were cultured in M3+BYPE medium, supplemented with 5% fetal bovine serum (FBS) and 1% penicillin and streptomycin at 25°C.
Wg-CM was harvested from S2 tubulin wingless cells. S2 tubulin wingless cells were seeded 24 h prior collecting the supernatant (1x106 cells/ml) by centrifuging the cells at 3500 rpm for 5 min. For the control medium S2 cells were prepared as described above. WCM or CM was added to Kc cells for 24 h to induce Wnt/Wg signaling. To induce the Wnt/Wg signaling pathway with CHIR99021 (S1263, Selleckchem), 25 μM of the inhibitor was used and added to the medium for 24h. As control DMSO was used. After 24 h of induction, cells were harvested.
Cas9 (49330, Addgene) and empty gRNA vector (49410, Addgene) were obtained from Addgene. Oligo design and cloning was accomplished after manufacturer’s protocol.
CRISPR was performed as described in [37]. Briefly, cells were plated at 2 x 106 cells per well of a 6-well dish and a total of 1.7 μg DNA, Cas9 and gRNA in a 1:1 ratio, was co-transfected into each well using Fugene HD (Promega) at a 1:2 ratio (μg:μl), following manufacturer’s instructions. Both gene loci were targeted simultaneously using a gRNA and Cas9 with integrated gRNA. Transfections were analyzed after 3 days, and selection was performed in 5 μg/ml Puromycin (P8833 Sigma). The genotype was analyzed using PCR primers spanning the cut site. PCR products were cloned in pGEMT-vector system (Promega) and 10–100 clones were analyzed by sequencing. Primers for gRNA cloning and for detection of CRISPR events are available in the S1 Table.
Nuclear protein extraction was performed as described in [54]. For Western blot analysis, monoclonal anti-Arm (1:500; N2(7A1), DSHB) and monoclonal anti-alpha-Tubulin (1:5000; T5168, Sigma) antibodies were used and followed by HRP-anti-mouse IgG (705-035-003, Jackson Immuno Research Laboraties, inc).
Real-time q-PCR analyses were carried out with SYBR Green Supermix (BioRad) on a iCycler iQ real-time OCR detection system (BioRad). For qRT-PCR, total RNA was extracted from 1–2 x 106 cells with NucleoSpin RNA extraction kit from Macherey-Nagel according to the manufacture’s protocol and reverse transcribed with Roche, followed by qRT-PCR. Sequences of the primer pairs used are listed in S1 Table.
All pair-end sequencing was performed on an Illumina HiSeq2500 machine at the Genomics Platform of the University of Geneva. For all experiments we compared three independent biological replicates and merged them for the subsequent analysis. All RNA-seq files are available from SRA NCBI database. Submission code: SUB2472808; Study: PRJNA378604 (Accession Number SRP101692).
All deep-sequencing data were mapped to the Drosophila reference genome dm3 using TopHat and analyzed as described in [34] and using thresholds as indicated above. We used GraphPad Prism for all statistical analysis and R for plotting.
STARR-seq in Drosophila WT cells and pan knockout cells was performed in two biological replicates as described in [47]. To obtain Wnt-responsive enhancers, cells were treated with 25μM CHIR99021 or DMSO for 24h. Data were analyzed as described in [47]. For Fig 7A fold enrichments were calculated directly over DMSO-treated samples at summits of induced enhancers and p-values indicate significance of the fold change. All STARR-seq files are available at the GEO database (GEO number GSE96542).
For TCF/Pan motif enrichment analysis, we used 200 bp regions around the summit of 185 induced, 348 repressed, 1834 constitutive enhancers, and 987 random sequences that were not detected with STARR-seq but followed the same genomic distribution (denoted as negative regions). Enrichments were calculated as described [46]. De novo motif analysis was done with DREME using negative regions as a background set (see S2 Table).
Enhancer candidates were amplified from genomic DNA of Drosophila Kc167 cells (for primers see S3 Table). All candidates were subcloned to either pCR8/GW/TOPO (Invitrogen) or pENTR/TOPO (Invitrogen) and delivered into the firefly luciferase vector [45] using the Gateway LR Clonase II enzyme mix (Invitrogen). Kc cells (1x105) were transfected using Fugene HD (Promega) with a total of 300 ng of various plasmid combinations (1:3 ratio of promoter reporter plasmid to Renilla). Luciferase activities were measured 48 h after transfection and after stimulation with either Wg ligand or CHIR99012 using the Dual-Luciferase Reporter Assay System (Promega). Every experiment was repeated at least twice with three replicates in each independent experiment. Enhancers’ sequences used are listed in S3 Table.
|
10.1371/journal.pbio.2006422 | Neural timing of stimulus events with microsecond precision | Temporal analysis of sound is fundamental to auditory processing throughout the animal kingdom. Echolocating bats are powerful models for investigating the underlying mechanisms of auditory temporal processing, as they show microsecond precision in discriminating the timing of acoustic events. However, the neural basis for microsecond auditory discrimination in bats has eluded researchers for decades. Combining extracellular recordings in the midbrain inferior colliculus (IC) and mathematical modeling, we show that microsecond precision in registering stimulus events emerges from synchronous neural firing, revealed through low-latency variability of stimulus-evoked extracellular field potentials (EFPs, 200–600 Hz). The temporal precision of the EFP increases with the number of neurons firing in synchrony. Moreover, there is a functional relationship between the temporal precision of the EFP and the spectrotemporal features of the echolocation calls. In addition, EFP can measure the time difference of simulated echolocation call–echo pairs with microsecond precision. We propose that synchronous firing of populations of neurons operates in diverse species to support temporal analysis for auditory localization and complex sound processing.
| We routinely rely on a stopwatch to precisely measure the time it takes for an athlete to reach the finish line. Without the assistance of such a timing device, our measurement of elapsed time becomes imprecise. By contrast, some animals, such as echolocating bats, naturally perform timing tasks with remarkable precision. Behavioral research has shown that echolocating bats can estimate the elapsed time between sonar cries and echo returns with a precision in the range of microseconds. However, the neural basis for such microsecond precision has remained a puzzle to scientists. Combining extracellular recordings in the bat’s inferior colliculus (IC)—a midbrain nucleus of the auditory pathway—and mathematical modeling, we show that microsecond precision in registering stimulus events emerges from synchronous neural firing. Our recordings revealed a low-latency variability of stimulus-evoked extracellular field potentials (EFPs), which, according to our mathematical modeling, was determined by the number of firing neurons and their synchrony. Moreover, the acoustic features of echolocation calls, such as signal duration and bandwidth, which the bat dynamically modulates during prey capture, also modulate the precision of EFPs. These findings have broad implications for understanding temporal analysis of acoustic signals in a wide range of auditory behaviors across the animal kingdom.
| Diverse groups of animals, such as electric fish, owls, and echolocating bats show remarkable temporal precision in the processing of sensory events. Specifically, weakly electric fish can detect a stimulus time disparity on the order of 0.4–1 μs [1,2], barn owls can distinguish an interaural time difference of approximately 10 μs, and big brown bats (Eptesicus fuscus) can discriminate differences in echo arrival time on the order of 36–80 μs in target range discrimination tasks [3–5] and <1 μs (and down to 10 ns) in target range jitter discrimination tasks [6–8]. The neuronal basis of the microsecond temporal precision has been identified for both electric fish and barn owls [9–12]. Specifically, single neurons in the prepacemaker nucleus of the weakly electric fish were found to be sensitive to temporal disparity as small as 1 μs [11]. Similarly, the firing rate of many neurons in the midbrain of the barn owls can distinguish an interaural time difference smaller than the behavioral threshold [9]. By contrast, the neural basis for microsecond auditory resolution in echolocating bats remains unknown.
Neurons hypothesized to function in bat sonar target distance measurement show facilitated responses to pairs of sounds, separated by a restricted range of delays, which mimic bat sonar calls and echoes, and this response property is referred to as echo delay tuning [13–15]. Although it has been hypothesized that echo delay–tuned neurons encode target range information in bats [16–19], the tuning widths of echo delay–tuned neurons in echolocating bats are typically several milliseconds wide [13,14,17,20], and this is far beyond the behavioral threshold.
It has also been hypothesized that the variability in response latency of auditory neurons may contribute to the bat’s sonar range resolution [21–25]. In contrast to echo delay–tuned neurons with millisecond delay tuning widths, the response latency of many auditory neurons of echolocating bats varies by only a few hundred microseconds, a reduction in timing errors of up to a factor of 100 [24]. For example, about one-third of neurons in the midbrain inferior colliculus (IC) of the big brown bat show latency variability <1 ms [26]. The most precise neurons in the bat IC show latency variability between approximately 100 and 250 μs [23,25,26]. Of note, neurons of submillisecond latency precision are not exclusive to echolocating bats but have also been reported in other animal models, such as cats and mice [27,28].
How echolocating bats discriminate echo arrival time with microsecond resolution remains an unsolved problem [18,19,29–31]. Here, combining extracellular recordings and mathematical modeling, we show that synchronous neural firing can improve the precision of stimulus event timing an order of magnitude greater than the temporal precision of single neurons. Of note, in this study, we analyzed the stimulus-evoked extracellular field potential (EFP, 20–600 Hz) to characterize neural timing in a band up to 600 Hz, whereas local field potential (LFP) is generally analyzed in a lower band, up to 100–200 Hz. The most precise EFPs (in the 200–600-Hz band) showed a latency variability of 17 μs, based on extracellular recordings in the IC of seven awake big brown bats passively listening to simulated echolocation calls. Moreover, the spectrotemporal structure of the echolocation calls affected the latency variability of the EFP, which is consistent with predictions from sonar receiver models of ranging accuracy [32–34].
We conducted three experiments with a total of seven bats. The first experiment investigated response latency variability measurements with a standard two-harmonic frequency modulated (FM) stimulus of 3-ms duration and approximately 80-kHz bandwidth, the second experiment investigated the influence of stimulus time–frequency structure on response latency variability, and the third experiment investigated the response latency variability measurements with simulated call–echo pairs.
In the first experiment, we took multichannel extracellular recordings with silicon probes in the IC of three head-fixed, awake big brown bats (one male, two female) passively listening to the broadcasts of simulated wideband echolocation calls of 3-ms duration and 70-dB sound pressure level (SPL) amplitude (Fig 1A, top right) (referred to as the “standard call” in this study). Based on a threshold for detection of six times the background noise floor, corresponding to a signal-to-noise ratio (SNR) of 15 dB (Fig 1A, grey dashed line), we recorded stimulus-evoked EFPs (20–600 Hz) of large amplitude (Fig 1A, blue trace) in addition to the spikes from multiunit activity (MUA) (Fig 1A, orange trace). Note that both EFP and MUA were derived from the same wideband neural recording with different cutoff frequencies of the digital filters. Over 20 presentations of the standard 3-ms FM stimulus, the response latency of the first EFP, measured as the time difference between the stimulus onset and the first negative peak of the EFP, was very stable, with a standard deviation (SD) of 150 μs in this example. Fig 1B shows data from another recording site that did not pick up isolated spikes but instead high SNR EFPs, with a latency SD of 59 μs. From both examples and S1 Movie, one can see that EFPs occurred directly after the presentation of the echolocation calls and typically occurred once per sound presentation. The high-amplitude, temporally precise EFPs have not been reported in neurophysiological studies of auditory processing in echolocating bats.
To address the possibility that the EFPs were artifacts of electrical noise from our sound broadcast system, we made neural recordings from the IC of a bat listening to two spontaneously vocalizing conspecifics inside the sound booth. Further, this approach allowed us to study responses to variable, natural stimuli, which animals process in the real world. Fig 1C shows that EFPs were reliably evoked in response to the natural vocalizations of bats in the recording booth. Compared to the broadcasts of acoustic stimuli through an electrostatic loudspeaker, the natural bat vocalizations resulted in slightly longer response latencies and greater variability, which can be explained by differences in call parameters, particularly the amplitude and duration differences between sonar playbacks and natural sounds. In the loudspeaker broadcasts, the stimulus amplitude (70-dB SPL) and duration (3 ms) were fixed. By contrast, the calls produced by the spontaneously vocalizing bats varied in duration by approximately 2.2 ms (3.3–5.5 ms) and amplitude by approximately 30 dB. The fact that EFPs could be evoked by vocalizations of live bats when the electrostatic loudspeaker was not powered eliminates the possibility that EFPs recorded in this study were a result of electrical artifacts. Moreover, EFPs showed selectivity for the fine spectrotemporal structure of the acoustic stimuli. Fig 1D shows EFPs from one recording site: EFPs were reliably detected in response to the top-row stimuli, including the standard call, −20 dB of the standard call, and the first harmonic of the standard call, but not in response to the bottom-row stimuli, which included a time-reversed version (upward FM sweep) of the standard call, white noise, and the second harmonic of the standard call. S1 Fig shows stimulus selectivity by the EFP for data from all three bats (S1 Data). There were 202 recording sites, in which at least one of the six types of the acoustic stimuli evoked ≥5 EFP responses over 20 presentations (i.e., ≥25% response probability). We found that the EFP showed a selective response at 193 (95.5%) and 124 (61.4%) recording sites, based on a 25% and 50% response probability difference between at least two acoustic stimuli, respectively. Thus, the stimulus-evoked EFPs reflect robust responses to biologically relevant acoustic stimuli.
How are the precise EFPs generated? Synaptic inputs and spiking activity are two main sources of EFPs [35–38]. The contribution of synaptic inputs to EFPs declines rapidly above approximately 200 Hz. Spiking activity, on the other hand, contributes to both the 20–200-Hz band and the 200–600-Hz band [39,40]. Thus, the 20–200-Hz EFPs can arise from both synaptic inputs and spiking activity, while the 200–600-Hz EFPs reflect primarily spiking activity. To assess whether the observed EFPs arose specifically from spiking activity, we separated the 20–600-Hz band EFP into 20–200-Hz band and 200–600-Hz band and analyzed them separately. In total, the 200–600-Hz EFP was detected at 528 recording sites from three bats (154 sites from Bat 1, 126 sites from Bat 2, and 232 sites from Bat 3). By contrast, the 20–200-Hz EFP and 20–600-Hz EFP were detected at 162 and 243 recording sites in total, respectively. The fact that the 200–600-Hz EFP was more prevalent in our recordings suggests that spiking activity contributed to the stimulus-evoked EFP, and all following EFP analyses were carried out on the 200–600-Hz EFP with 90% detection reliability. A 90% detection reliability criterion indicates that EFPs or spikes were detected in at least 72 out of 80 stimulus presentations.
In addition to the EFP, we analyzed the first negative peak of multiunit activity (MUA; 600–3,000 Hz). A comparison of the latency variability between the first negative peak of the EFP and MUA revealed that the EFP latency was more stable than the MUA (Fig 2A; Wilcoxon rank sum test, P < 0.001). At the 90% detection reliability, the median SD for the EFP and the MUA were 104 μs (minimum SD: 53 μs) and 425 μs (minimum SD: 62 μs), respectively. A detailed examination of the data sets revealed that the MUA data contained many recording sites with greater latency variability than the EFP data, which might be the basis of the observed population differences. To test this idea, we excluded the sites of the MUA group whose SD were larger than the maximum SD of the EFP data but again found that the EFP response latencies were more stable than the MUA (Fig 2B; Wilcoxon rank sum test, P < 0.001). These results show that EFP latency is more precise in timing stimulus events than MUA.
There is a topographical organization of the EFP across the IC with respect to the stability of response latency (Fig 2C and 2D). Specifically, high-precision EFP latency was most often recorded in the ventral region of the IC, whereas higher variability in the EFP latency tended to appear in the dorsal region of the IC (Fig 2D). Similarly, EFPs of shorter response latency were mainly found in the ventral region of the IC (Fig 2C), which aligns closely with earlier observations of single neurons in the IC [41].
The EFP to the 3-ms standard FM stimulus showed a latency variability as low as 53 μs, which is more precise than the least variable single unit latencies of IC neurons reported in the literature, between 100 and 250 μs [23,25,26]. To test whether our recordings contained single neurons that were far more precise than those previously reported, we identified single spiking units of high SNR using the Wave_clus algorithms [42] and carefully evaluated the waveforms of the sorted single units (see S1 Text for details). In total, we identified 59 single units in our recordings. The most precise single unit in our data set had an SD response latency of 140 μs (with a population median of 1.36 ms), which is within the range of values reported in the literature.
Our analyses revealed that the EFP showed the highest temporal precision in response latency, and single units showed the highest latency variability. This raises a fundamental question: how does the activity of single neurons contribute to EFPs of greater temporal precision? Since action potential waveforms only contain a proportion of energy in the EFP frequency band (<600 Hz), multiple overlapping spikes are required to produce a prominent EFP. We hypothesized that the number of firing neurons and neural synchrony influence the temporal precision of the EFP.
To gain insights into the effect of the number of firing neurons on the properties of the EFP, we counted the number of negative peaks in the wideband neural signal, whose amplitudes were larger than the detection threshold (20–3,000 Hz, Fig 1A and 1B), within ±3 ms of the mean response latency of the EFP. We chose the ±3-ms time window to count the number of firing neurons, as it largely overlaps with the 5-ms period of the 200-Hz high-pass cutoff frequency of the EFP, aiming to obtain counts of neurons potentially contributing to the EFP. We counted the number of firing neurons from the wideband neural signal (20–3,000 Hz) rather than from the 600–3,000-Hz band (i.e., spikes), since spikes from neurons at a distance that cannot be picked up by the electrode also contribute to the low frequencies of the EFP [35,36,39,43]. In other words, the 20–3,000-Hz wideband neural signal may contain more information than the MUA. We found that there was a weak negative correlation between the number of firing neurons and the temporal variability of the EFP (Fig 2E). Moreover, the number of firing neurons correlated negatively with the response latency and positively with the peak amplitude of the EFP (Fig 2F and 2G). These results suggest that the number of firing neurons could influence the properties of the EFP. The weak correlations between the number of firing neurons and the properties of the EFP also imply that the number of firing neurons is not the only factor influencing the properties of the EFP.
To systematically investigate the effect of the number of neurons and neuronal firing synchrony on the EFP properties, we took a simulation approach that allowed us to examine quantitatively the effect of one factor at a time. We started with the simplest scenario in which five neurons with the same extracellular spike shape fired an action potential at a random latency between 17.5 and 22.5 ms, and thus each neuron had a response latency of 20 ± 1.5 ms (mean ± SD, indicated by the red dash-dot line in Fig 3A). Of note, the 20-ms average response latency is arbitrarily chosen for illustrative purposes and does not affect the simulation results. Then, we band-pass filtered this virtual recording to generate simulated MUA and EFP and analyzed the first negative peak of the MUA and EFP for each simulated trial in the same way as for the experimental data. Fig 3A shows the results of 1,000 simulations. Both the MUA and EFP showed responses that were shorter in response latency and more precise in timing than the single neuron, supporting the experimental observations. Subsequently, we quantified the effect of the number of neurons on MUA and EFP for both high-synchrony scenarios in which all neurons fire randomly within a 1-ms time window and low-synchrony scenarios in which all neurons fire randomly within a 10-ms time window. Fig 3B shows that the MUA and EFP become more precise and greater in amplitude with increasing number of firing neurons but only under the high-synchrony simulation. For instance, the first EFP showed a precision of 75 μs and an amplitude of 2.5 mV when 50 neurons fired within a 1-ms time window (Fig 3B), which is approximately four times more precise and approximately 32 times greater in amplitude than the single neuron. Moreover, the effect of increasing the number of firing neurons is stronger for the EFP than for MUA. Thus, this result is consistent with the experimental observation that the EFP can be more precise than MUA (Fig 2A and 2B).
Moreover, one can mathematically demonstrate that when multiple neurons fire within a short time window, about the width of the action potentials, EFPs become shorter in response latency, more precise in timing, and larger in amplitude (Fig 3C, see details in S1 Text). This mathematical analysis leads to three predictions: (1) the peak amplitude of the EFP increases linearly with increasing number of the firing neurons; (2) the variability of EFP latency decreases as the number of firing neurons increases, following a power law relationship; (3) the decrease in response latency of the EFP is positively related to its variability. Importantly, all of these predictions agree with the experimental observations (Fig 2E–2G) and simulations (Fig 3A and 3B).
The simplified simulations and the mathematical model based on multiple copies of a single extracellular spike shape allowed us to directly examine the influence of the number of firing neurons and neuronal synchrony on the properties of the EFP, yet extracellular recordings from neurons naturally vary in spike waveform, which additionally depends on the position and geometry of the electrode. To consider a more realistic scenario, we performed biophysical simulations in which neurons produced varying spike waveforms at a virtual recording electrode. Fig 3D shows that the conclusions drawn from the simplified simulation approach also hold for simulated recordings with different extracellular potential waveforms. For example, the precision of the EFP latency measurement was higher than that of MUA, and the single neurons showed the highest latency variability. Particularly, the EFP (200–600 Hz) measurement showed a precision of 45 μs with 100 neurons each firing with a 250-μs SD. Moreover, the magnitude of response latency reduction of both MUA and EFP latency measurements correlated positively with the variability of single neurons.
One key component of bat echolocation is the dynamic adjustment in the spectrotemporal features of sonar signals with target distance [45–48]. For example, big brown bats initially use long duration calls (up to 15 ms) of narrow frequency bandwidth (fundamental sweeps over 3–5 kHz) to search for insects in open space and then progressively shorten the duration (down to 1–2 ms) and widen the bandwidth of the calls (fundamental sweeps over 30 kHz) while approaching the prey (Fig 4A). It has been shown that bats achieve the highest precision in sonar ranging with short, wideband calls, which they produce when approaching prey [8,32,49]. Thus, we hypothesized that the precision of stimulus registration with the EFP changes with signal duration and signal bandwidth. Specifically, we predicted that wideband, short echolocation calls (i.e., produced by bats in the prey approach phase) should generate EFPs with the highest temporal precision, and narrowband, long echolocation calls (i.e., produced by bats in the prey search phase) should generate EFPs of the lowest precision.
To test the hypothesis that stimulus bandwidth and duration influence the temporal stability of the EFP latency, we conducted a second experiment by taking extracellular recordings from the IC of four awake big brown bats (all females) passively listening to simulated echolocation calls of varying bandwidth and durations. We found that measurements of EFPs at a single site showed the highest precision in registering the timing of acoustic events (a minimum SD of 17 μs at a recording depth of 1,060 μm) when the bats were listening to echolocation calls of the shortest duration (1 ms) (Fig 4C, S2 Fig). The average value of the 10 most precise EFPs evoked by the 1-ms broadband echolocation calls was 21 ± 2 μs. By contrast, EFPs showed the largest variability in registering the timing of acoustic events (a minimum SD of 168 μs at a recording depth of 720 μm) when the bats were listening to calls of longest duration (12 ms) and narrowest bandwidth (5-kHz bandwidth of the first harmonic). The average value of the 10 most precise EFPs evoked by the 12-ms narrowband echolocation calls was 200 ± 20 μs. Overall, the precision of EFP timing with respect to stimulus events became progressively poorer with increasing call duration (two-way ANOVA, F = 158.85, df = 3, P < 0.001, all Padj < 0.001 for pair-wise comparisons). Similarly, the response latency of EFPs became progressively longer with increasing call duration (Fig 4B) (two-way ANOVA, F = 218.9, df = 3, P < 0.001, all Padj < 0.001 for pair-wise comparisons). EFPs to narrowband calls showed the greatest variability in the temporal registration of stimulus events (two-way ANOVA, F = 65.63, df = 2, P < 0.001, all Padj < 0.001 for pair-wise comparisons) and the longest response latency (two-way ANOVA, F = 85.21, df = 2, P < 0.001, all Padj < 0.001 for pair-wise comparisons), except for the shortest duration of 1 ms (Wilcoxon rank sum test, all Padj > 0.05).
The systematic change in EFP latency and precision with stimulus duration and bandwidth suggests that EFP latency itself could serve to code different stimulus parameters. To understand how much information is potentially encoded by the response latency of EFPs, we applied information theory and calculated the sample size–corrected Shannon mutual information [50]. First, we calculated the mutual information for each stimulus bandwidth, namely the wideband (25–55 kHz for the first harmonic), the midband (25–40 kHz for the first harmonic), and the narrowband (25–30 kHz for the first harmonic). Fig 5A shows an example with a relatively high mutual information, and Fig 5B shows an example with a relatively low mutual information. Of note, the theoretical maximum of the mutual information calculated for four stimulus types within each bandwidth category is 2 bits. Fig 5C–5F show the mutual information for all recording sites from all four bats. A maximum mutual information of 2.57 bits suggests that the response latency of EFPs from a single recording site can maximally distinguish six out of the 12 total stimulus types (Fig 5F). Nevertheless, within each bandwidth category, the response latency of EFPs from 185 (22.2%), 143 (18.2%), and 15 (3.6%) recording sites can distinguish at least three out of four stimulus types with a mutual information larger than 1.58 (red dashed line). Moreover, the wideband chirp featured the highest mutual information, followed by the midband chirp, and the narrowband chirp had the smallest mutual information (Wilcoxon rank sum test, all P < 0.001). These data suggested that the response latency of EFPs encodes more information about the duration of the stimulus type with a wider frequency bandwidth.
Echolocating bats use the time delay between an emitted call and returning echo to estimate the distance or range of objects in the environment [5,17,19]. For the big brown bat, the operating range of echolocation for prey detection is up to a few meters, which corresponds to echo delays of a few tens of milliseconds. To test whether the high-precision EFP evoked by single sounds can estimate the time difference of a pair of sounds that simulate the call–echo pairs used in echolocation tasks, we took neural recordings from the IC of three big brown bats (the same bats in Experiment 1) passively listening to simulated call–echo pairs. The calls were the “standard call” (Fig 6A) presented at an amplitude of 75-dB SPL. The echoes were attenuated versions of the standard call and were presented at either 25-, 45-, or 65-dB SPL and at delays between 2 and 30 ms following the call.
Fig 6A shows an example of EFP responses to a call–echo pair over 20 presentations. The simulated echo delay was 28 ms. The echo was 10 dB weaker than the call, i.e., 65-dB SPL. After each presentation of the call–echo pair, there were often two EFPs. The first (black dots) and the second (blue dots) EFP responses occurred at a latency of 7.8 ± 0.075 ms and 35.8 ± 0.082 ms, respectively. The estimated echo delay (red dots) from the EFPs was 27.99 ± 0.08 ms, which is not only very precise across stimulus presentation but also aligns with the actual 28-ms echo delay. Fig 6B shows that across the 684 tested stimulus conditions from all three bats, the time difference estimated by EFPs showed a submillisecond precision in 411 (60%) stimulus conditions and showed a precision greater than 100 μs in 155 stimulus conditions (15%). Fig 6C and 6D shows that the precision of the stimulus time difference estimation by EFPs was largely constrained by the precision of the EFP response to the weaker echo (Fig 6D), not by the precision of the first EFP response to the more intense call (Fig 6C). In response to simulated call–echo pairs, the first EFP responses to the calls, with a median SD of 0.1 ms, were more precise in timing than the second EFP responses to the echoes, with a median SD of 0.29 ms (Wilcoxon signed-rank paired test, P < 0.001). Fig 7 shows the EFP responses at a single recording site to simulated call–echo pairs at three different echo amplitudes (25-, 45-, and 65-dB SPL) and at varying echo delays between 2 and 30 ms. These data show that at each echo amplitude, the estimated time difference of EFPs increases with the echo delay of the simulated call–echo pairs (from the bottom to the top) and achieves a precision greater than 100 μs for many stimulus conditions.
In this study, we have shown that the response latency of the EFP evoked by simulated echolocation calls shows high precision in the temporal registration of stimulus events, with latency variability as low as 17 μs. The EFPs, recorded in the IC of the awake, passively listening big brown bat, were generated from the synchronous firing of a population of neurons, and both the number of firing neurons and the tightness of neuronal synchrony influenced the EFP properties of response latency, precision (variability), and amplitude. We also provide evidence that the precision of the EFP depends on the spectrotemporal features of bat echolocation calls, and EFPs evoked by simulated call–echo pairs can precisely estimate echo delay.
There is a general consensus that spiking activity represents an important source of LFPs at frequencies higher than approximately 100 Hz, at least in the hippocampus [35,38,43,51,52]. Yet sources other than spiking activity, such as synaptic events, can also contribute to the LFP at higher frequencies [43,53]. Since the term LFP has been traditionally used to refer to EFPs at frequencies below 100–200 Hz, we adopted the term EFP [54] in this study to describe neural activity in a band between 200 and 600 Hz. Combining extracellular recording and mathematical modeling, we provide evidence that the EFPs in our IC recordings are generated by a population of neurons firing synchronously. Although the neurons that generate the EFP are likely located within the IC, one recent study raised the possibility that the EFP can be generated via axon bundles projecting from one brain region to another [54]. Specifically, the authors found that EFPs could be reliably recorded from axon bundles, and the properties of the EFPs were affected by the properties of the axons. This work implies that the source of EFPs recorded in the IC of bats in our study could be either from the local IC neurons, from axons passing through the IC, or a combination of the two.
One intriguing feature of the EFP is its remarkable precision in registering the timing of stimulus events. The most precise EFP evoked by the same acoustic stimulus over 30 trials showed an SD of 17 μs in response latency. Here, the reported 17-μs temporal precision might be an underestimation of the highest temporal precision achievable by EFPs in the IC of the big brown bat. Since the temporal precision of the EFP is determined by both neural synchrony and the number of firing neurons, based on Equation 2 in Fig 3C, one can predict a temporal precision of 3.2 μs from 1,000 neurons that each has 100-μs variability in response latency. Single neurons showing 100-μs temporal variability in response latency have been reported for the IC of the big brown bat [26]. Note that all the measured and predicted temporal precision reported above was for single EFP sites. Considering the possibility that the nervous system can combine information across multiple brain sites [28], submicrosecond temporal precision could be achieved, although the exact mechanisms of such population coding have yet to be identified.
Our study revealed two potential mechanisms that the nervous system can exploit to represent the timing of stimulus events with high precision: (1) many neurons responding to a given stimulus and (2) high synchrony of firing among these neurons. Increasing the tightness of neuronal synchrony is a general computational principle for modulating nervous system functions. At the single neuron level, precise spike timing can propagate or be maintained across synapses when multiple neurons fired in high synchrony [55–57]. At the synaptic level, highly synchronized presynaptic inputs are more effective at driving a neuron to fire [58] and generating spikes with higher temporal precision [59,60]. There is also evidence that neuronal synchrony is functionally linked to the behavioral performance of animals [61–63]. For example, Gutierrez and colleagues (2010) found that within the taste–reward circuit of rats, neurons that fired in synchrony with licking behavior exhibited greater cue discrimination than nonsynchronized neurons and that the magnitude of this effect increased with learning. On the other hand, the potential importance of a larger population of neurons for specific brain functions is a matter of debate [64]. Here, we identified one potential advantage of an increased neuronal population size in improving the temporal precision of the EFP. Nevertheless, it is worth noting that without neural synchrony, the population size of firing neurons itself contributes very little to improving the temporal precision of the EFP (Fig 3B, the low-synchrony scenario).
One crucial finding of this study is the functional relationship between the precision of registering the timing of stimulus events in the EFP and the spectrotemporal features of the echolocation calls of bats. During a foraging task, big brown bats use echolocations of long duration and narrowband frequency range to search for prey. A main function of the search-phase echolocation calls is prey detection, and long duration narrowband signals are well suited for this task [32,65]. Once a prey item is detected, insectivorous bats rapidly change the echolocation call structure by decreasing the call duration and increasing the frequency bandwidth while approaching the prey. One important function of the approach-phase echolocation calls is to track the position of the prey, which thus requires estimating target distance, i.e., measuring echo arrival time, with great precision. Moreover, there is extensive evidence showing that short broadband echolocation calls are well suited for precise measurements of echo arrival time [5,8,32,49]. Here, we found that the EFP evoked by echolocation calls of short duration and wideband spectrum showed the most precise registration of stimulus events, and by contrast, the EFP evoked by long-duration, narrowband echolocation calls showed the poorest temporal precision. Moreover, we have shown that the precise EFPs evoked by single sounds can also be used to estimate the time difference of simulated call–echo pairs with a precision as great as 45 μs, representing a feasible neural substrate for the 36–80-μs behavioral echo delay discrimination by echolocating bats [4]. The precision of time difference estimation of the simulated call–echo pairs was largely constrained by the second EFP responses to the weaker echoes, which were less precise than the first EFP responses to the more intense calls. This finding suggests that echo amplitude might play a role in influencing the precision of target ranging and points to the potential importance of stabilizing echo amplitude for target ranging. Interestingly, there is an accumulating body of evidence that echolocating bats maintain a relatively constant echo amplitude while approaching a target [66,4,67,68]. Thus, the temporal precision of the EFP, as a marker for acoustic events, provides a window to neural computational principles that might underlie the accurate echo delay discrimination behavior of echolocation tasks.
The findings reported here raise several important questions that can be answered by further research. One critical question concerns the mechanisms by which the brain makes use of the neural events underlying the precise EFPs. At this time, knowledge is lacking about the contribution of the biophysical properties of single neurons and/or biochemical environments to the EFP. Because EFPs are generated by synchronous firing in a population of neurons, the high-precision EFP may be a proxy for population coding across single neurons. Thus, the question of how the brain might extract and use information carried by EFPs could be approached through population analysis of stimulus-evoked activity in pools of single neurons. We also conjecture that inhibition may play an important role in generating the precise EFPs, which is known to underlie precise interaural time difference estimation in mammals [69]. Specifically, many properties of IC neurons, such as selectivity for interaural intensity difference, sound frequency, and sound duration, are generated by convergent inhibitory and excitatory inputs, as revealed by the intracellular recording of postsynaptic potentials [70–72]. Both intracellular recording and application of inhibitory neurotransmitter antagonists in the IC could provide a first step to investigate the role of inhibition in shaping EFP properties.
Going beyond the implications for understanding mechanisms of precise target ranging performance in echolocating bats, the high-precision EFP bears general relevance for understanding stimulus coding in the brain across species. Precise timing of neural signals could encode rich information that is relevant to a variety of brain functions [73]. For example, the temporal precision of neural events may be a key property in categorizing sound features by the auditory system [74,75]. Indeed, the information theoretic analysis showed that EFPs from many recording sites in the bat IC could unambiguously differentiate three out of four simulated echolocation calls of different durations. By extension, EFP analysis could generate insights into mechanisms supporting other auditory behaviors, such as acoustic communication, scene analysis, and spectrotemporal discrimination [76]. Although specialized neural mechanisms for temporal processing have been revealed in species that use stimulus timing in natural behaviors, such as barn owls [9] and electric fish [11], the extent to which diverse species share common mechanisms to achieve high temporal precision is an important open question. We propose that the computational principle of synchronized firing across a population of neurons represents a general mechanism for the nervous system to register precisely the timing of sensory events.
Big brown bats, Eptesicus fuscus, collected in the state of Maryland under a permit issued by the Department of Natural Resources were used as subjects (permit number 55440). The Johns Hopkins University’s Institutional Animal Care and Use Committee approved all the procedures used for this study (protocol number BA14A111). The protocols are in compliance with the Animal Welfare Act regulations and Public Health Service Policy. The university maintains accreditation by the Association for the Assessment and Accreditation of Laboratory Animal Care International.
Extracellular recordings were made from auditory neurons in the IC of seven adult big brown bats (one male, six females) in an acoustic booth. In the big brown bat, the IC sits on the dorsal surface of the brain and is up to approximately 2 mm in length both anterior-posteriorly and dorsoventrally. On the day of neural recording, the bat was placed into a custom-made bat holder, its head was immobilized via a headpost, and a craniotomy of <1-mm2 size above the central IC was made with a scalpel under a microscope. Multiple penetrations were made to the IC of the same bat in different recording days and bats were recorded between a few days and more than a month. Within a single penetration, the probe was systematically advanced dorsoventrally from the surface down to approximately 1,800 μm using a hydraulic Microdrive (FHC). Extracellular potentials were recorded by a silicon probe from Neuronexus that had the 1 × 16 arrangement of recording sites, with an intersite separation of 50 μm. The probe thickness was 25 μm and the site area of the probe was either 177 μm2 or 703 μm2. On the day of the headpost surgery, under isoflurane inhalation anesthesia, part of the skin and the temporal muscles overlying the IC were removed, and a custom headpost was attached to the bone at the midline using cyanoacrylate gel.
During neural recording, the awake bat was passively listening to the broadcasts of simulated echolocation calls or call–echo pairs from a custom electrostatic loudspeaker (1-cm diaphragm) placed 60 cm away from the contralateral ear of the recording IC at an angle of approximately 30° from the middle line. Digital echolocation calls were generated from customized LabVIEW scripts and played at a sampling rate of 1 MHz using the data acquisition card from National Instrument (PXIe 6358). We achieved a flat frequency response of the playback system for the frequency range of 20–100 kHz (±1 dB) by digitally compensating for the uneven frequency response with its compensatory impulse response [77,78]. The compensatory impulse response was computed using the Maximum Length Sequence method based on 5 seconds white noise recordings with a one-fourth–inch measurement microphone (Model 7016, ACO, United States). Acoustic stimuli were played at an amplitude of 70-dB SPL (root mean square) unless otherwise stated. Each acoustic stimulus was repeated at least 20 times. Specifically, the standard 3-ms echolocation call (Fig 1A, first harmonic sweeps down from 55 kHz to 25 kHz, and the second harmonic from 110 kHz to 50 kHz) was repeated 80 times in the first experiment, and all other stimuli, such as white noise or time-inversed echolocation calls shown in Fig 1D, were repeated 20 times. In the second experiment that tested the functional significance of the EFPs, each stimulus was repeated 30 times. The call–echo pairs in the third experiment were repeated 20 times. The time interval between stimulus presentations was 300 ms, and the order of stimulus presentation was randomized for each experiment. The neural recording was digitized at a sampling rate of 40 kHz using a Plexon system of 64 analog channels. The original neural signal was amplified by a factor of 20 times before digitization but later was restored to the correct scale during data analysis. Neural recording and sound broadcasting were synchronized via a transistor–transistor logic (TTL) signal outputted from a second analog output channel of the National Instrument card each time when an acoustic stimulus was broadcast, and the TTL signal was directly recorded by an analog input channel of the Plexon system. More details on the animal surgical preparation and neural recording can be found in Macias and colleagues (2018) [79].
All neural recordings were batch processed in Matlab (R2015a, Mathworks). The general steps of data processing include (1) band-pass filtering the original recording with Elliptic filters from the Wave_clus algorithms [42]; (2) an adaptive threshold, six times of the background noise level, based on the same equation from Quiroga and colleagues (2004), was used to identify spikes and the peaks of EFPs; and (3) the onset time of the acoustic stimuli was identified based on the TTL signal, then the response latency of the MUA or the EFP was computed from the first negative peak. Negative spikes were the dominant observations across our recording. We quantified the precision or variability of the peaks of the EFPs with the SD. Data points were considered as outliers and excluded from analysis if their values were greater than q3 + w × (q3 − q1) or less than q1 − w × (q3 − q1), where w was set to 1.2 and q1 and q3 were the 25th and 75th percentiles of the sample data, respectively. We applied the outlier exclusion criterion, as we are particularly interested in how precise in timing the EFPs can potentially achieve. Moreover, outliers can greatly bias the estimations of the mean and the SD, which are the principal measures in the current study. Although outlier exclusion results in less variation in the measured EFP latencies and the corresponding SD of the EFPs, it is important to note that the reported EFPs were based on a 90% detection reliability criterion (see Results above), which limited the outliers in <10% of the data points. Note that although a sampling rate of 40 kHz results in a temporal resolution of 25 μs, the SD of a 25-μs resolution sampling system can be as small as approximately 5 μs [80]. Thus, the reported best temporal precision of 17 μs is well within the resolution of our data sampling system.
A cell model, L4_SS_cADpyr230_1, was downloaded from the Human Brain Project's Neocortical Microcircuitry (https://bbp.epfl.ch/nmc-portal/welcome) [44]. We constructed cell populations consisting of 10–1,000 cells, by randomly rotating and distributing instances of the same cell model with a uniform cell density of 200,000 neurons per mm3, within a disc of 100-μm thickness. Because of the fixed cell density, the radius of the disc was dependent on the population size.
Each cell in the population received 100 conductance-based excitatory synaptic inputs randomly distributed across the cell membrane with a uniform area weighted density. The synapses were implemented using a double exponential function (Exp2Syn in NEURON) with the rise and decay time constants of 1 ms and 3 ms, respectively, and a reversal potential of 0 mV. The synaptic weights were normally distributed around 0.2 nS with a standard deviation of 0.04 nS. The arrival of the 100 synaptic inputs was normally distributed with a mean value of T = 230 ms after stimulation onset, with a standard deviation of 0.25 ms. This evoked an action potential in the cell model, and after the extracellular action potential was calculated, a time window of ±15 ms around the maximum value of the somatic membrane potential was extracted from the extracellular action potential and saved to file. The time resolution of the simulation was 1/32 ms.
Extracellular potentials were calculated using LFPy2.0 (http://EFPy.github.io/) [81], which runs on the NEURON simulator [82]. For the calculation, each cell compartment was treated as a line source, except for the somatic compartments, which were treated as a point source. The extracellular conductivity was set to 0.3 S/m [83]. The extracellular potential was calculated in the center of the cell population.
Each of the calculated extracellular action potentials was randomly jittered in time, with different SDs for the jitter, and summed to produce the population signal. For each of the different tested SDs for the jitter, 100 trials of the population signal were calculated. The signals were filtered with elliptic filters (scipy.signal.ellip) to obtain the EFP and MUA. The first peak of the filtered signal was identified as the first negative crossing below a threshold of six times the SD of the signal.
All code used for this project is available from https://github.com/torbjone/sharp_wave_ripples/.
Details of the mathematical model shown in Fig 3, spike sorting, and the information analysis were presented in the S1 Text.
|
10.1371/journal.pbio.2005130 | High-dimensional single-cell phenotyping reveals extensive haploinsufficiency | Haploinsufficiency, a dominant phenotype caused by a heterozygous loss-of-function mutation, has been rarely observed. However, high-dimensional single-cell phenotyping of yeast morphological characteristics revealed haploinsufficiency phenotypes for more than half of 1,112 essential genes under optimal growth conditions. Additionally, 40% of the essential genes with no obvious phenotype under optimal growth conditions displayed haploinsufficiency under severe growth conditions. Haploinsufficiency was detected more frequently in essential genes than in nonessential genes. Similar haploinsufficiency phenotypes were observed mostly in mutants with heterozygous deletion of functionally related genes, suggesting that haploinsufficiency phenotypes were caused by functional defects of the genes. A global view of the gene network was presented based on the similarities of the haploinsufficiency phenotypes. Our dataset contains rich information regarding essential gene functions, providing evidence that single-cell phenotyping is a powerful approach, even in the heterozygous condition, for analyzing complex biological systems.
| Diploid organisms harboring a wild-type gene and a loss-of-function mutation are called heterozygotes. They are expected to have weak or no individual phenotypes because the mutation is compensated for by the intact allele. The dominant inheritance of phenotypes in heterozygotes is an exceptional phenomenon called haploinsufficiency. Haploinsufficiency was thought to be a rare occurrence; however, a sensitive technique called high-dimensional single-cell phenotyping challenges this perspective. Investigations of single-cell phenotypes revealed that a large extent of the essential genes in yeast exhibit haploinsufficiency. Our analyses also provided crucial information on gene functional networks based on haploinsufficiency phenotypes. This work shows that high-dimensional single-cell phenotyping is a useful tool that can be used to better understand complex biological systems.
| The concepts of dominance and recessiveness were originally formulated by Gregor Mendel [1] and are still fundamental to modern genetics. Loss-of-function mutations are mostly recessive and rarely dominant in diploid organisms. Haploinsufficiency is a rare manifestation of the dominant phenotype arising from a copy of a loss-of-function mutation in the heterozygous state and was initially studied in Drosophila [2]. There is great interest in haploinsufficient genes because the loss of 1 functional allele is linked to human diseases including cancer and tumorigenesis, developmental and neurological disorders, and mental retardation [3]. Therefore, it is challenging to determine the number of genes in the genome that are sensitive to 1-copy gene loss [4,5].
Two models have been developed to explain the occurrence of haploinsufficiency. As can be seen in dosage-dependent sex determination in Drosophila [6], a reduction in the gene copy number affects regulatory genes working at a threshold level. Some proteins are likely produced at the lowest level possible for proper function. Therefore, haploinsufficiency may simply be due to a reduction in protein level in the heterozygous state, which is referred to as the insufficient amount hypothesis. A second theory, referred to as the balance hypothesis, predicts that the stoichiometry of various protein components is important for maintaining the integrity of a protein complex [7]. In yeast, representative haploinsufficient genes include cytoskeletal components such as actin (Act1) [8] and tubulin (Tub1) [9] as well as components of protein complexes such as spindle pole body component (Ndc1) [10] and myosin (Mlc1) [11]. In these circumstances, gene overexpression also results in an imbalance of the components and shows similar phenotypic consequences of 1-copy gene loss.
Genome-wide studies have been performed to investigate haploinsufficient growth phenotypes in the budding yeast Saccharomyces cerevisiae. Among 5,900 yeast genes analyzed, approximately 3% (184 mutants) exhibited haploinsufficient growth in rich media [12]. Many of the yeast haploinsufficient genes were functionally related and related to ribosomal function [12], suggesting a significant contribution of ribosomal function to rapid growth. By further investigating the growth phenotypes under limited nutrient conditions, up to 20% of the genome was found to display a haploinsufficient abnormality [13]. A recent systematic screen of another budding yeast, Candida albicans, revealed that 10% of the genes in the genome influenced cell size under optimal growth conditions [14]. However, the extent of haploinsufficiency was still restrictive, and little is known about the functional relationships between these genes.
One approach to identify haploinsufficiency is to monitor the phenotypes from different perspectives. Cell morphology is an attractive target for intensive analyses because it reflects a wide variety of cellular events, and hundreds of traits can be analyzed [15]. In this study, we investigated the haploinsufficiency of 1,112 heterozygotes of yeast essential genes using high-dimensional phenotyping with 501 morphological traits. We found that more than half of the essential genes displayed haploinsufficiency under optimal growth conditions, indicative of extensive haploinsufficiency. Similar haploinsufficiency phenotypes were caused by heterozygous deletion of functionally related essential genes. Correlation networks of haploinsufficient genes provided a global view of their functional relationships. Our dataset offers useful resources for the study of essential gene functions in S. cerevisiae.
We employed yeast heterozygous diploid strains with deletions in each of the essential genes and examined haploinsufficiency in terms of its effects on morphology (morphological haploinsufficiency) by performing single-cell high-dimensional phenotyping. To minimize variation due to inconsistencies in data acquisition, we collected the cultures after growth to a precise point in early log-phase in rich medium, used the automated image processing system CalMorph [15], and analyzed more than 200 cells for each strain. To exclude technical artefacts due to staining procedures and cell segmentation, automatic discriminators and classifiers built into CalMorph made it possible to obtain high-quality multivariate information on single cells [16]. In addition to 220 mean and 61 ratio morphological parameters, 220 variance parameters—which represent variance of the single-cell distribution in morphology—were extracted. To detect phenotypic abnormalities, a generalized linear model (GLM) was applied (S1 Table). As expected, haploinsufficient morphological phenotypes were rarely observed. Of all the combinations between 501 traits and 1,112 heterozygous diploids, only 0.764% (4,258 assays) were significantly different from the wild-type diploid based on a 1-sample 2-sided test (false discovery rate [FDR] = 0.01; P < 7.64 × 10−5; S1 Fig). However, an analysis of morphological phenotypes in each strain revealed a large number of haploinsufficient genes. A total of 59.1% (657 heterozygous diploids, S2A Table) of the heterozygous deletion mutants exhibited differences compared with the wild-type diploid in at least 1 of the morphological traits examined (FDR = 0.01; P < 7.64 × 10−5; red area in Fig 1A and S2B Fig). The number of abnormal mutants detected for each trait was relatively small, mostly within the IQR between 2 and 12. We estimated that the rate of false positive (FP) abnormal mutants detected by chance in our analysis was 6% (Fig 1B, black line), which was almost the same as the number of abnormal replicates in the wild type (Fig 1B, orange line). This confirmed that our statistical estimation of the number of haploinsufficiency phenotypes was not overestimated. We used an alternative approach to estimate the number of haploinsufficient mutants following dimensional reduction. A large number of heterozygotes (40% of 1,112) still displayed haploinsufficiency in at least 1 of the 20 principal components (PCs) covering 60% of variance of the morphological phenotypes (FDR = 0.05, S3C Fig).
We found that the cumulative number of haploinsufficient mutants increased with an increase in the number of morphological traits examined (Fig 1B, red line). Mean parameters—and, more effectively, variance parameters—contributed to haploinsufficiency detection (S4A Fig), highlighting the importance of single-cell phenotyping. Ratio parameters were less important because the cumulative number of haploinsufficient mutants reached 98% without the ratio parameters (S4B Fig, light blue line). We next investigated whether the differences between the morphologies of haploinsufficient mutants increased or decreased phenotypic variance and found significantly more phenotypic variance in the 657 haploinsufficient strains than in the other strains (S5 Fig; P < 0.01 after Bonferroni correction and Mann–Whitney U test). This observation is consistent with the previous finding that decreasing dosage with the use of conditional alleles often results in increased morphological variation within populations of isogenic cells [17]. Therefore, one widespread function of essential genes is to stabilize morphological phenotypes.
We counted the frequency of haploinsufficiency in nonessential genes by examining 100 randomly selected heterozygous gene-deletion mutants. For the 501 traits, 33% of the heterozygous diploids showed haploinsufficiency at the same threshold (P < 7.64 × 10−5; Fig 1B, gray line). Therefore, the frequency of haploinsufficiency in essential genes (Fig 1B, red line) was approximately 2-fold greater than that in nonessential genes (Fig 1B, gray line). We noted previously that 65% of the haploid mutants with nonessential deletions were morphologically distinct [15], indicating that the morphological phenotypes in heterozygous diploids were less commonly observed than those in haploid deletion mutants (S2B Table). These analyses indicated that essential genes have a large impact on haploinsufficient morphological phenotypes.
We tested the morphological haploinsufficiency of heterozygous diploids under nutrient-limited growth conditions in 50 randomly selected heterozygous deletion mutants, which exhibited no haploinsufficiency in rich media. After growth in poor synthetic medium, 40% (16.4% out of 40.9%) of heterozygous diploids that exhibited no obvious morphological phenotypes in rich media exhibited haploinsufficiency in at least 1 of the morphological traits (P < 7.64 × 10−5; Fig 1A, pink area). This indicated that up to 75.5% (59.1% + 16.4%) of the heterozygous diploids exhibited phenotypes either in rich or poor synthetic medium.
We examined the morphological haploinsufficiency to see whether it could be explained by functional defects of the genes. To investigate the relationship between gene function and a particular haploinsufficiency phenotype, we performed dimensional reduction by principal component analysis (PCA) and canonical correlation analysis (CCA) [18], which is used to explore the relationship between 2 multivariate sets of variables. PCA and CCA successfully compressed all combinations of 444 morphological traits and 830 gene ontology (GO) terms into linear combinations of phenotypic (21 phenotype canonical variables [pCVs]) and gene-function features (21 GO term canonical variables [gCVs]) (S6 Fig). In fact, analysis of the canonical correlation coefficient revealed a significant correlation between phenotype (pCVs) and gene function (gCVs) (P < 0.05, Bartlett’s chi-squared test). At a given canonical correlation coefficient in each pair of 21 CVs, no FPs were found by chance after 10,000 iterations of the randomization, indicating that randomized phenotypic data yielded no pairs of CVs. The phenotypic space composed of pCVs was suitable for understanding phenotypic features of haploinsufficient mutants with the same functional defects. For example, exploring the phenotypic space of pCV1 and pCV3 revealed that heterozygotes for RNA polymerase II (RNA pol II) core complex (green) and for subunits of the cytosolic chaperonin containing TCP-1 (CCT) complex (red) were plotted in different directions (Fig 2B). This graphically demonstrated that the heterozygous mutations in RNA pol II and chaperonin CCT caused specific morphologies, namely, large/elongated cell shape and large actin region/nonelliptical cell shape, respectively (Fig 2A, S3A Table).
The logistic regression analysis can be used to identify the best combinations of pCVs for each GO term, yielding the maximum likelihood prediction of the gene functions with haploinsufficiency phenotypes (e.g., cytosolic large ribosomal subunit [ribosomal protein of the large subunit (RPL)] in Fig 2C). We applied this approach to every GO term and identified 306 GO terms corresponding to 553 genes with a significant correlation between gene function and haploinsufficiency phenotype (P < 0.05, likelihood ratio test after Bonferroni correction) (Fig 3 and S3B Table). Therefore, haploinsufficiency phenotypes were associated with gene function in 90% of the haploinsufficient genes, suggesting that the observed phenotypes were mostly explained by functional defects of the genes.
To better understand morphological haploinsufficiency, we examined the overlap between haploinsufficient genes for growth [12] and haploinsufficient genes for morphology. A contingency table test showed significant correlations between these 2 datasets (S4 Table; P < 0.01 according to Fisher’s exact test), suggesting a common integrant. A previous study revealed that ribosomal function was specifically enriched in haploinsufficiency based on cell growth [12]. Although many ribosomal genes were also morphologically haploinsufficient, specific gene functions were not enriched among the 657 morphologically identified haploinsufficient genes (FDR = 0.1); instead, genes encoding most of the essential cellular processes—such as replication, transcription, translation, protein degradation, membrane trafficking, transporter, cell cycle progression, morphogenesis, and macromolecular synthesis—were represented (S3B Table). We also noted that specific gene functions were not enriched in genes that were not morphologically haploinsufficient (FDR = 0.1). Therefore, careful high-dimensional and single-cell phenotyping detected numerous haploinsufficient genes with functions in diverse cellular processes.
A previous study indicated that genes involved in protein complexes were enriched among haploinsufficient genes related to growth [12]. The genes involved in protein complexes were also significantly enriched among haploinsufficient genes related to morphology (S5 Table, P < 0.01 by Fisher’s exact test for 1 side). This suggested that specific gene functions were enriched in both haploinsufficient morphological genes and genes involved in protein complexes. In fact, some gene functions (such as nuclear polyadenylation-dependent mRNA catabolic process, cytosolic large ribosomal subunit, etc.) were enriched with high degrees of protein–protein interaction (S7A Fig, PPI). Similar but distinct gene functions were significantly enriched with high degrees of genetic interaction (S7A Fig; genetic interaction) [19]. By comparing Fig 3 with S7A Fig, a Venn diagram was constructed (S7B Fig), which indicated that among 124 GOs of protein complexes, 70 GOs were enriched in morphologically identified haploinsufficient genes. Therefore, our analysis suggested that numerous haploinsufficient genes are involved in protein complexes with diverse cellular functions.
Haploinsufficient genes are the genes that are sensitive to 1-copy gene loss. Therefore, we next analyzed the correlation of haploinsufficient genes for morphology with overexpression-sensitive genes [20] and with highly expressed genes [21]. We revealed a significant correlation with overexpression-sensitive genes (S8A Fig; Spearman rank correlation coefficient, P < 0.01 by t test) but failed to detect any correlation with highly expressed genes (S8B Fig, Spearman rank correlation coefficient, P = 0.38 by t test). However, we detected a significant correlation when we selected genes annotated with a specific GO (S9 Fig, Wald-test, FDR = 0.05). This implies that the correlation between morphologically identified haploinsufficient genes and highly expressed genes is GO specific. Based on these results, we discussed the feasible models for the mechanism of haploinsufficiency (see Discussion).
A previous study of heterozygous diploids showed that the essential genes involved in ribosome biogenesis cause coupling of the growth rate to cell size [22]. Analysis of our dataset confirmed a significant correlation between growth rate and cell size in 198 heterozygous ribosome biogenesis mutants (Fig 4A). Aside from cell size, we revealed that other morphological features were correlated with growth rate in these ribosome biogenesis mutants (S6 Table, likelihood ratio test, FDR = 0.05). Of 163 correlated morphological features, we extracted the independent features (S7 Table and S10 Fig) and summarized them with a schematic representation (Fig 4B). Therefore, our results provide a deeper understanding of a mechanism that may link cell growth with cell morphogenesis, including growth in size, cell cycle progression, actin morphogenesis, and nuclear morphogenesis.
Because the haploinsufficiency phenotypes were due to functional defects of the genes, we further assessed the degree of similarity between the phenotypic profiles of individual haploinsufficient mutants. To do this, a full matrix of gene–gene pairwise similarities was calculated based on the haploinsufficiency phenotypes. Although phenotypic correlation coefficients between all pairs of the heterozygous diploids were distributed largely from –0.23 to +0.23 (mean ± 1 SD), the mean values of those sharing the same GO categories were typically positive (S11 Fig). There were only a few (0.98%) highly correlated (>0.5) cases. We analyzed the interactions with correlations above 0.5 and found many cases of interactions within the protein complex GO (S12 Fig). Therefore, the similar haploinsufficiency phenotypes were associated with the deletion mutants in the same GO categories. After dimensional reduction by CCA, a high level of precision and recall curve for GO terms was achieved (S13 Fig), indicating that the positive correlation coefficient had substantial predictive power for gene function. We compared the precision-recall characteristics of our phenotypic data to the results from other high-throughput studies (S14 Fig) and found that our data (red) were almost as precise and sensitive as protein interaction [23] (green) and microarray co-expression data [24] (purple) and were more predictable than phosphoprotein (orange) [25] and genetic interaction data [26] (blue). We then tested pairs of correlation coefficients between representative functional gene groups (S8A Table) and observed both positive and negative correlations. For example, the mean value between “cytoplasmic translation” and “ribosomal large subunit assembly” (both involved in protein synthesis) was positive, while that between “ribosomal large subunit assembly” and “proteasome regulatory particle” was negative (Fig 5). The negative correlations reflected the opposing nature of the cellular processes, namely, protein synthesis and degradation. Our results strongly suggest that positive and negative correlations of the haploinsufficiency phenotypes reflect functional relationships in cellular processes.
We used correlations between haploinsufficiency phenotypes to construct global functional maps among the yeast essential genes. Based on the patterns of the relationships, we systematically mapped 513 essential genes belonging to 46 GO terms (Fig 6A and S8B Table). We observed 15 core gene groups containing 285 haploinsufficient genes with functions in DNA replication, transcription, nuclear transport, translation, phospholipid metabolism, and protein degradation that served as a hub: these genes were related directly and/or indirectly to all of the other genes. Pairwise testing did not detect significant phenotypic correlations between the core gene groups (Fig 6B), indicative of the different and diverse functions of the hub genes. These phenotypic relationships provide a global view of the functional relationships between large numbers of haploinsufficient genes.
Comprehensive single-cell phenotyping of heterozygous diploids in budding yeast revealed that more than half of the essential gene mutants are haploinsufficient in morphology. Up to 76% of the heterozygous diploids showed distinct morphological phenotypes either in rich or minimal media. High-dimensional phenotyping with many points of view yielded an even larger number of haploinsufficient mutants. This suggests that future high-dimensional assays will identify more haploinsufficient genes that are linked to human diseases. Among phenotypic values acquired from hundreds of individual cells, the variance value of the traits was found to be more effective than others, demonstrating the importance of single-cell phenotyping. The morphological phenotypes of the haploinsufficient heterozygotes could be mainly explained by gene function. There was morphological similarity within the deletion mutants of functionally related genes, as evidenced by dense gene clusters with rich functional information, and functional networks based on morphological similarity.
Phenotypes can be perturbed by environmental changes, epigenomic effects, and/or experimental artefacts [27]. To demonstrate that the observed haploinsufficiency phenotypes were due to chromosomal heterozygous deletions, we determined whether the haploinsufficiency phenotypes could be explained by gene-functional defects. We found that 90% of genes with functional defects (553 of 610 haploinsufficient genes with reliable GO annotations) were associated with the phenotypes of heterozygous diploids. The strong correlation between gene function and the haploinsufficiency phenotype provides concrete evidence that a decrease in the gene dosage could result in malfunctioning in a large proportion of essential genes.
Given the results from previous comprehensive studies of haploinsufficient genes, it was quite surprising that such a large proportion of essential genes displayed haploinsufficiency. Studies in budding yeast revealed that approximately 9% of essential genes in the genome are haploinsufficient for growth in rich medium [12]. A careful survey of the Drosophila genome showed that only 56 loci were associated with an altered phenotype when present as a single copy [28]. Compared with results from a previous study, we found that most of the genes involved in essential cellular processes were haploinsufficient in terms of morphology.
Genes encoding components of protein complexes were significantly enriched among the haploinsufficient genes, which supports the balance hypothesis. In addition, the significant correlation between overexpression-sensitive and haploinsufficient genes supports the balance hypothesis discussed previously [7,12]. On the other hand, many genes encoding noncomplex enzymes were also haploinsufficient, which supports the insufficient amount hypothesis. Although we failed to detect a significant correlation between highly expressed and haploinsufficient genes on the whole, we detected a significant correlation when we selected haploinsufficient genes annotated with specific GO terms, including carbohydrate-derivative biosynthetic process (GO:1901137 in alcohol metabolic process group; S9 Fig), RNA methyl transferase activity (GO: 0008173 in tRNA processing group; S9 Fig), and mitotic cohesin complex (GO: 0030892). The correlation between highly expressed and haploinsufficient genes supports the insufficient amount hypothesis, and haploinsufficiency of these genes can be easily explained by this hypothesis. Therefore, according to our analysis, it is conceivable that both the insufficient amount and balance hypotheses are correct. Further study will be necessary to determine which hypotheses are applicable for each haploinsufficient gene.
Our dataset will provide researchers with a tool for gaining insights into the functions of yeast essential genes. First, haploinsufficiency phenotypes can be used to understand the function of essential genes. Compared with the various pleiotropic phenotypes frequently observed in conditional lethal mutants [29,30], haploinsufficiency phenotyping is equally reliable. Second, phenotypic similarities between heterozygous diploids can be used either to identify previously known functional connections or propose previously unknown functional connections. It should be noted that phenotypic similarities between the nonessential deletion mutants were used to predict gene function [15]. We observed both positive and negative correlations between haploinsufficiency phenotypes, suggesting that high-dimensional single-cell phenotypes reflect functional relationships in the cellular network. Third, it would also be interesting to compare haploinsufficient genes observed under different conditions. Because more than 1,000 chemical genetic assays revealed a growth defect for all deletion mutants [31], phenotyping in multiple environments is a promising strategy. Therefore, as is the case for growth phenotyping [13], morphological phenotyping under different growth conditions will reveal important aspects of gene function. Finally, comparisons between haploinsufficient and chemical-induced morphological profiles [32] will be used to explore intracellular drug targets. We will be able to make more precise predictions by integrating haploinsufficient morphological profiles with chemical-genetic interaction profiles [33] or other gene features. These will give us additional tools for drug target prediction.
A collection of heterozygous gene-deletion mutants was purchased from EUROSCARF (http://www.euroscarf.de). Essential genes were defined previously [34]. The yeast diploid strain BY4743 was used as the wild type. Strains heterozygous for 1,112 essential genes and 100 randomly selected nonessential genes and the wild-type strain were cultured under optimal growth conditions at 25°C in nutrient-rich yeast extract peptone dextrose (YPD) medium containing 1% (w/v) Bacto yeast extract (BD Biosciences, San Jose, CA), 2% (w/v) Bacto peptone (BD Biosciences), and 2% (w/v) glucose, which was prepared as described previously [15]. Strains heterozygous for 50 essential randomly selected genes and the wild-type strain were cultured under severe growth conditions at 37°C in nutrient-poor synthetic minimal dextrose (SD) medium, which was prepared as described previously [35].
To minimize variation due to inconsistencies in data acquisition, we used a precise protocol to prepare yeast cells growing in early log-phase. Strains were activated from the freezer stock by streaking onto YPD agar plates and incubating for 3 d at 25°C. Three colonies from each strain were inoculated into 2 mL of YPD liquid medium in a 20-mL glass test tube (Iwaki, Shizuoka, Japan), and the liquid culture was incubated on a rotator (30 rpm with RT-50; TITEC, Saitama, Japan) at 25°C for 20 h. Then, the cells were transferred into 20 mL of fresh liquid medium in a 100-mL conical flask (Iwaki). The cells were further incubated in a shaking water bath (110 rpm with LT10-F; TITEC) at 25°C at least for 16 h. A total of 5.0 × 106 cells at log-phase were harvested and used for fixation and fluorescence staining.
Yeast cells were fixed for 30 min in growth medium supplemented with formaldehyde (final concentration, 3.7%) and potassium phosphate buffer (100 mM [pH 6.5]) at 25°C as described in [36]. Yeast cells were then collected by centrifugation at room temperature and further incubated in potassium phosphate buffer containing 4% formaldehyde for 45 min. Next, actin staining was performed by overnight treatment with 15 U/mL rhodamine-phalloidin (Invitrogen, Carlsbad, CA) and 1% Triton-X in phosphate-buffered saline (PBS). Staining of cell-surface mannoproteins was performed by 10-min treatment with 1 mg/mL fluorescein isothiocyanate (FITC)-conjugated concanavalin A (Sigma-Aldrich, St. Louis, MO) in P buffer (10 mM sodium phosphate and 150 mM NaCl [pH 7.2]). After washing twice with P buffer, the yeast cells were mixed with mounting buffer (1 mg/mL p-phenylenediamine, 25 mM NaOH, 10% PBS, and 90% glycerol) containing 20 mg/mL 4’,6-diamidino-2-phenylindole (DAPI; Sigma-Aldrich) to stain DNA. Finally, the specimens were observed using an Axio Imager microscope equipped with a 6100 ECplan-Neofluar lens (Carl Zeiss, Oberkochen, Germany), a CoolSNAP HQ cooled charged coupled device (CCD) camera (Roper Scientific Photometrics, Tucson, AZ), and AxioVision software (Carl Zeiss).
Image processing was performed using CalMorph (version 1.3) software designed for diploid yeast strains [37]. CalMorph can collect a large amount of data regarding many morphological parameters of individual cells such as cell cycle phase and cell form from a set of photographs of cell walls, nuclei, and actin cytoskeletons. The CalMorph user manual is available at the Saccharomyces cerevisiae Morphological Database (SCMD; http://yeast.gi.k.u-tokyo.ac.jp/datamine/) [38]. Descriptions for each trait were presented previously [15].
To assess haploinsufficiency cell morphology phenotypes statistically, we used the GLM as described previously [39] with minor modifications. The haploinsufficiency phenotypes of heterozygotes were detected using the 1-sample 2-sided test with a null distribution estimated from 114 replicated wild-type strains.
The null distribution for each trait was estimated using 1 of 4 probability density functions (PDFs), as described previously [39]. To minimize the effects of confounding factors affecting microscopic output, we applied the linear model using dummy variables (S1 Table, and S1 Text). The maximum likelihood estimation for each PDF was performed using R function “gamlss” contained in the “gamlss” package [40]. The validity of the null distributions estimated by the wild-type phenotype was assessed using the R-squared value of a quantile–quantile plot. A theoretical distribution for each trait was estimated using the “qqplot” function of the default package using random values (n = 11,400) generated from the PDF estimated as a null distribution. To calculate the R-squared value, the theoretical distribution was compared to the distribution of the wild type (n = 114). The median of R-squared values among 501 traits was 0.966 (IQR 0.964–0.976), indicating that the selected model and its estimated parameters approximated the distributions of the wild type.
P values for each mutant were calculated based on the estimated PDF at 2 sides (low and high tails), such that twice the minimum P values were used for statistical tests (1-sample 2-sided test). The FDR was estimated using the R function “qvalue” in the “qvalue” package [41]. Similarly, the number of deletion mutants for nonessential genes was estimated based on the 1-sample 2-sided test with 122 replicated wild-type and 4,718 nonessential gene-deletion mutant strains [15].
The number of mutants detected for at least 1 trait was counted for each threshold (S2 Fig). To estimate the number of samples detected by chance for at least 1 trait, we performed parametric bootstrap resampling using PDFs with maximum likelihood estimations. Random values of 114 samples were generated from each PDF for each parameter. The number of trials (n = 3,000) with at least 1 falsely detected trait among 501 traits was counted at each threshold and averaged. In S2 Fig, the confidence intervals from the FPs were estimated by assuming binomial distribution.
The purpose of this analysis was to reduce the dimensions from 501 traits and identify biologically important morphological features. We used Z values of 501 traits as a morphometric profile and a Boolean matrix of GO terms as a gene function for each heterozygote. First, we obtained Z values using test statistics of the Wald test using the R function “coeftest” in the “lmtest” package [42] and selected 657 heterozygotes (59%) with significant haploinsufficiency phenotypes at an FDR of 0.01 (Fig 1). We further discarded 47 genes that were annotated by GO terms with fewer than 3 genes. We then selected 830 GO terms that annotated more than 2 genes in the remaining 610 haploinsufficient genes and fewer than 200 genes in the genome with no identical sets of annotated genes. Finally, we used Z values of 444 morphological traits calculated from 610 of the 657 heterozygotes (S2B Fig), such that the 444 traits were detected in at least 1 of the 610 heterozygotes.
To reduce dimensionality, we subjected the morphometric profiles to PCA and the first 17, 29, 50, 91, and 130 PCs (phenotype principal components [pPCs]) contributed more than 0.6, 0.7, 0.8, 0.9, and 0.95, respectively, to the cumulative contribution ratio (CCR). Next, to estimate functional relationships among the 610 genes, we used the structure of 830 GO terms. Dimensionality of GO terms can be reduced by PCA on a Boolean matrix (if a gene was annotated by GO, then its value was 1; otherwise, it was 0), as described previously [43]. The 830 GO terms for the 610 genes were then subjected to PCA, and the first 59, 84, 120, 181, and 346 GO term principal components (gPCs) contributed 0.6, 0.7, 0.8, 0.9, and 0.99, respectively, to the CCR indicating that approximately 346 gene functions were related to the 610 genes.
After projection of Z values on pPCs and a zero matrix on gPCs for 114 replicates of the wild type, we applied CCA to the 130 pPCs and the 346 gPCs, for which the CCRs were 0.95 and 0.99, respectively (S6 Fig). Significance of the canonical correlation coefficient was tested at P < 0.05 based on Bartlett’s chi-squared test [44] to obtain 21 morphological features (pCVs) and 21 gene function features (gCVs). To characterize each pCV based on morphological features, linear regression analysis was performed based on the Z value of each trait on pCV and detected at P < 0.05 after Bonferroni correction using the F test. Morphological features for each pCV are summarized in S3A Table by successive PCA, as described previously [45].
To detect correlation between GO terms and haploinsufficiency phenotypes, we applied multiple logistic regression analysis to each of the 830 GO terms with combinational optimization techniques for pCVs as explanatory variables. Logistic regressions were performed using the R function “brglm” in the “brglm” package, which was designed to determine a solution to the problem of separation [46]. Combinational optimization was performed using the R function step, in the default package after adaptation of the “brglm” function. A best linear model consisting of 1 of 21 pCVs as an explanatory variable was selected by optimization of algorithms based on Akaike’s Information Criterion (AIC) [47]. The selected model was tested at P < 0.05 after Bonferroni correction by the likelihood ratio test using the R function “lrtest” in the “lmtest” package [42].
The hierarchical cluster analysis (HCA) in Fig 3 was performed using the R function “hca.” Dissimilarity was calculated based on the ratio of shared genes to the union of genes annotated with 2 arbitrary genome-wide GO terms. The GO terms were divided into 20 groups as listed in S3B Table at a height value less than 0.99, such that height was the minimum ratio of the different genes between clusters (complete linkage).
Precision and recall were calculated as described in Baryshnikova and colleagues (2010) [48] with minor modifications. Correlation coefficients of all (185,745) pairs of 610 genes were calculated using 130 pPC scores or 21 pCV scores. The gene pairs were sorted in ascending order of correlation coefficient and were ranked by the correlation coefficient. The number of gene pairs for which 2 genes were co-annotated by at least 1 of the 830 GO terms listed in S3B Table were counted as true positives (TPs) for each nth (n = 1, 2, …. 185,745) rank of gene pairs from the first to nth rank of the gene pairs. TP was used for the recall. The precision of each TP was calculated by dividing TP by each rank of pairs.
We first divided the genes into functional gene groups with no common term. The 553 haploinsufficient genes with significant high probabilities of correlation to the gene functions were classified into disjunctional functional gene groups using GO annotations in common. The binary distance between each pair of genes was calculated based on a Boolean matrix of the selected 306 GO terms and used for clustering by the complete linkage method using static branch cutting with a height value less than 1; 62 gene groups were identified, each of which contained from 1 to 33 genes (Fig 6B). To assign the most appropriate GO terms to each gene group, enrichment of GO terms was analyzed using Fisher’s exact test (P < 0.05 after Bonferroni correction; S8A Table). In 49 of the groups, more than 1 GO term was enriched. The remaining groups were therefore identified as functional gene groups with no GO terms in common.
Next, we calculated pairwise correlation coefficients between the functional gene groups. To detect significant relationships between the gene groups, we performed pairwise CCA between arbitrary pairs of the 62 gene groups (62C2 = 1,891) using 21 pCV scores. To eliminate possible detection bias, we used a smaller number of genes than the number of pCVs by reducing dimensionality of genes after applying PCA to the data of heterozygous genes. For pairwise CCA, we applied CCA to pCV scores using the genes and/or the selected PCs as variables, and extracted heterozygote canonical variables (hCVs) as independent components that correlated between the gene groups. We then tested the significance of the canonical correlation coefficient of the first hCV at P < 0.05 after Bonferroni correction using Bartlett’s chi-squared test [44]. Among 1,891 pairs of the 62 gene groups, 136 pairs were detected with significant relationships between the gene groups (Fig 6B).
A good way to show a global view of functional relationships based on phenotypic correlation is through graphical representation of gene networks. Similarity of phenotypes between the pairs of 513 heterozygotes was calculated using 21 pCV scores and expressed as a correlation coefficient. To visualize the network of the 46 GO term-enriched gene groups (513 genes, S8B Table) with significant relationships to other groups (Fig 6B), we used the R function “qgraph” [49], with which a correlation matrix can be represented as a network. We fed the matrix of the pCV-score–based correlation coefficient after zero filling cells into the “qgraph” of R function when at least 1 of 2 genes in the combination was not significantly related to the first hCV at P < 0.05 by t test for correlation coefficient (see Pairwise CCA section).
|
10.1371/journal.pcbi.1003475 | Escherichia coli Peptidoglycan Structure and Mechanics as Predicted by Atomic-Scale Simulations | Bacteria face the challenging requirement to maintain their shape and avoid rupture due to the high internal turgor pressure, but simultaneously permit the import and export of nutrients, chemical signals, and virulence factors. The bacterial cell wall, a mesh-like structure composed of cross-linked strands of peptidoglycan, fulfills both needs by being semi-rigid, yet sufficiently porous to allow diffusion through it. How the mechanical properties of the cell wall are determined by the molecular features and the spatial arrangement of the relatively thin strands in the larger cellular-scale structure is not known. To examine this issue, we have developed and simulated atomic-scale models of Escherichia coli cell walls in a disordered circumferential arrangement. The cell-wall models are found to possess an anisotropic elasticity, as known experimentally, arising from the orthogonal orientation of the glycan strands and of the peptide cross-links. Other features such as thickness, pore size, and disorder are also found to generally agree with experiments, further supporting the disordered circumferential model of peptidoglycan. The validated constructs illustrate how mesoscopic structure and behavior emerge naturally from the underlying atomic-scale properties and, furthermore, demonstrate the ability of all-atom simulations to reproduce a range of macroscopic observables for extended polymer meshes.
| The structure of the bacterial cell wall has been a point of controversy and contention since it was first discovered. Although the basic chemical composition of peptidoglycan, the key constituent of the cell wall, is now well established, its long-range organization is not. This dearth of information at the mesoscopic scale is a result of the inability of experimental imaging techniques to simultaneously visualize both the atomic-level detail of the peptidoglycan network and its macroscopic arrangement around the bacterium. Now, using molecular dynamics (MD) simulations, we have carefully constructed and validated models of sections of the Escherichia coli cell wall in full atomic detail. By comparing various properties of these models, including elasticity, pore size, and thickness with experiments, we can discriminate between them, resolving which best represents the native wall structure. In doing so, our study provides approaches for connecting measurements made in atomic-scale MD simulations with large-scale and even macroscopic properties.
| The cell wall rests outside the cytoplasmic membrane and provides bacteria with shape, rigidity, and protection from lysis due to the significant turgor pressure emanating from within [1]. It is primarily composed of a porous, mesh-like network of polymerized peptidoglycan, a repeating disaccharide/oligopeptide molecule. Because it is covalently connected, the cell wall is also the largest macromolecule in nature [2]. The chemical composition of peptidoglycan is largely conserved: relatively long glycan strands are cross-linked by short oligopeptides (see Fig. 1) [3]. In Gram-negative bacteria the cell wall presents as a relatively thin network (2–7 nm) between the inner and outer membranes, while in Gram-positive bacteria, it is much thicker, between 20 and 35 nm [1], [4].
Multiple theoretical models for the architecture of peptidoglycan at the mesoscopic scale have been conceived [1]. The models fall into two primary classes, a horizontal layer in which the glycan strands run parallel to the cell surface in a circumferential direction [5]–[7] and a scaffold in which the strands are oriented perpendicular to the surface [8], [9]. While some experiments have been interpreted as support for the scaffold model, e.g., the NMR structure of a peptidoglycan subunit [10], more recent electron cryo-tomography (ECT) on purified Gram-negative sacculi revealed circumferential glycan strands [11]. Even more complex models have been put forth, such as cables of coiled peptidoglycan encircling Gram-positive bacteria based on atomic force microscopy (AFM) measurements [12], although ECT on these bacteria failed to find distinct cable-like structures [4]. Using biochemical experiments and atomic-scale simulations, it was demonstrated that only layers composed of circumferential glycan strands could fully account for the ECT observations on Gram-positive bacteria, namely a distinct curling and thickening behavior of the sacculus, i.e., the part of the cell wall remaining after cell lysis, upon shearing [4].
One limitation of many of the previously developed models is that they are typically constructed with an idealized geometric arrangement, which is then deformed according to a set of mathematical rules. Not surprisingly, this procedure tends to generate cell walls with an unnatural degree of order [3]; indeed, regular patterns of quadrilateral or hexagonal shapes are often depicted [7], [8]. Furthermore, while apparent disagreement with a chosen set of experimental data has been used to indict some models over others [9], an alternative explanation is that the specific rules used to construct the model as well as the presumed experimental constraints were too strict [13].
In an attempt to circumvent some of the limitations of previous models, we have constructed and simulated patches of Gram-negative cell wall in their full atomic detail. By modeling the cell wall as a intricate composite of its individual components, we restrict the number of assumptions necessary for its construction. A single-layered model was chosen based on ECT of Gram-negative sacculi and on recent simulations of Gram-positive cell-wall patches [4]. Each patch of peptidoglycan was built from the level of individual residues on up, quantifying the behavior at each level, and connecting it with experimental measurements of various structural and mechanical properties. These comparisons are used both to validate the constructions and to illustrate the robustness of the cell wall to variations in average glycan strand length within the range of those observed in vivo. The widespread agreement with experimentally measured properties favors the disordered circumferential model of peptidoglycan in Gram-negative bacteria over other models.
While the glycan strand composition is uniform across all bacteria, composed of alternating -1,4-linked GlcNAc and MurNAc saccharides, the peptide stem, connected to the lactyl moiety of the MurNAc residue, is quite diverse [3], [6]. In E. coli the full, five-residue sequence is L-Ala (1) D-isoGlu (2) - (3) D-Ala (4) D-Ala (5), where -pm is -diaminopimelic acid, a lysine derivative [3]. Also of note is that the D-isoglutamate is connected through a -carboxy linkage to the pm residue (see Fig. 1A). While alanine is already present in the CHARMM force field, the remaining four constituents were originally absent. Therefore, we developed new CHARMM-compatible topologies and parameters for these constituents, as well as for the connections between them (see Methods along with topology and parameter files provided in the Supporting Information). The novel force fields have already been successfully utilized for simulation of Gram-positive peptidoglycan [4].
After parameterization, a single glycan strand 320 residues long was constructed without peptides. A useful property to quantitatively characterize the flexibility of a polymer is the persistence length . It is defined as(1)where is the position along the strand, is the angle between the tangent vectors at positions and , and the average is taken over all starting positions [14]. Effectively, is a measure of the stiffness of the strand. Two 5-ns simulations of the 320-mer strand were carried out, and was determined by the initial decay of the correlation in Fig. 2 [15], [16]. The two simulations provided values of and 13.6 nm; extending the latter simulation to 10 ns changed this value only marginally (). This persistence length is of the same order of magnitude as that found for other simple polysaccharides from experiments and/or modeling, which can span a large range, e.g., 4.5–13.5 nm for pectin [17] and 14.5 nm for cellulose [18]. Although not measured here, the persistence length of peptide cross-links is at least an order of magnitude less, being no more than 3–4 Å, making them significantly more stretchable than the relatively rigid glycan strands.
The peptides project outward from the glycan strand, presumably in a helical fashion (see Fig. 2B) [1]. The periodicity of these peptides is intimately connected to the orientation and degree with which neighboring glycan strands can form cross-links with one another. An angle of between successive peptide side chains was assumed in the classical layered model, thus placing every other one in the plane of the cell wall [5], [6]. An NMR structure of a peptidoglycan fragment, however, displayed an angle of , in line with that in the scaffold model [10]. To determine the equilibrium angle for an isolated peptidoglycan strand, two 60-residue-long strands were constructed and simulated for 10 ns, one with an initial peptide-peptide angle of and one with an angle of . For the strand initially at 120, the average angle relaxed to by the end of the 10-ns simulation, while the one initially at was (see Fig. 2B). Based on these results, we conclude the native periodicity of the peptides is approximately four per turn. However, the significant variability in the angle in simulations, even within a single strand, indicates that this periodicity is not strictly maintained and could be easily modified by external forces.
In order to construct the full peptidoglycan network, individual glycan strands need to be covalently linked through their peptides. Although this linkage takes a variety of forms depending on species, in E. coli the most common link is a peptide bond made between the amino group of the pm residue (position 3) and the carbonyl group of the penultimate D-Ala (position 4), shown in Fig. 1B [6]. In the course of transpeptidation, the terminal D-Ala (position 5) is also cleaved, both processes being carried out by penicillin-binding proteins [19]. The degree of cross-linking varies between species and even growth states within a single species [1]; for E. coli it is typically around 50% on average, i.e., about half of the peptides are linked and half are free [20]. Although alterations to the cross-linking fraction likely affects the mechanical and structural properties of the peptidoglycan network, this variable is not explored in the current study.
Two-dimensional periodic patches of peptidoglycan were constructed following a specific set of procedures designed to minimize user bias (see Methods); an example of a resulting system is shown in Fig. 3A. Despite being initially constructed as an organized, patterned network, the final organization of the peptidoglycan resembles the “disordered circumferential layered” model observed in cryo-tomography images of purified sacculi [11]. However, because no tension was applied, the possibility remains that the peptidoglycan becomes more ordered under native cellular conditions, which is explored below [1].
Much like the fraction of peptides cross-linked, in the bacterial cell wall the average glycan strand length takes on a large range of values, as low as six disaccharides in the stationary growth phase of Helicobacter pylori [21] and more than 50 in Proteus morganii [22]. Even in E. coli, a range of values spanning from 9 to 60 disaccharides has been measured by different experimental techniques for different stages of growth [1]. To examine the dependence of mesoscale properties of peptidoglycan on the average length of the glycan strand, multiple models with a specific average, but non-uniform, number of disaccharides were constructed, including 83.2, 175.8, and 262.4 disaccharides, denoted avg8, avg17, and avg26, respectively (see Fig. 3). Additionally, as an extreme case for comparison, two patches of cell wall with unbroken, periodic (and therefore effectively infinite) glycan strands with unit-cell lengths of 15 and 30 disaccharides, denoted Inf1 and Inf2, were modeled. Because of the small number of strands used (12 for avg17 and avg26, for example) it's not possible to reproduce distributions, although the limited range of lengths in avg17 (8–26 disaccharides) does agree with where the majority of the strand lengths in CG models falls [23].
A defining property of the peptidoglycan layer is its tensile elasticity, i.e., its response to applied strain coming from the turgor pressure inside the bacterial cell, also referred to as Young's modulus. Elasticity also serves as a key metric for comparing the constructed models to experimental measurements. Because peptidoglycan is orthotropic, the elasticities along its two symmetry axes are not identical [24]. Based on the theory of mechanical deformation of a two-dimensional sheet (see SI for a full derivation starting from the material's constitutive relations), the Young's moduli in each orthogonal direction, for the glycan strands and for the peptide cross-links, are given by(2)(3)where and are the applied strains in each direction, defined as , and and are the resulting stresses, measured in units of force/area. The dimensionless Poisson's ratios, and , relate the spontaneous strain arising in one direction given an applied strain in the other. In order to calculate the elasticity from simulation, varying strains were applied in the plane of the peptidoglycan by altering its dimensions, with one dimension stretched and the other held fixed at its equilibrium value, calculated from a minimum 20-ns constant-pressure simulation (see Methods). Because only one of or is allowed to be non-zero in each simulation, Eqs. 2 and 3 can be simplified to(4)(5)While the stresses, and , are formally the derivatives of the free energy with respect to strain in the corresponding dimension, by virtue of the reversible work theorem they can be directly related to the thermodynamic pressure in that dimension, i.e., a mean force (see Text S1) [25], [26].
Determination of each model's elasticity was based on six or more 2-ns-minimum simulations in which the peptidoglycan was stretched in the direction parallel to the glycan strands between 1.25% and 17.5% () relative to the relaxed state and six more simulations between 5% and 45% () parallel to the peptide cross-links. Each simulation was repeated to ensure consistency of the results, giving at least 24 simulations per cell-wall patch (see Fig. S3 in SI). The resulting elasticities and Poisson's ratios are presented in Table 1, along with measurements and calculations from other studies [24], [27]–[30].
All simulated models reproduce the expected anisotropy of the elastic moduli in the two orthogonal directions [27]. The glycan strands are found to be much stiffer, with values of ranging from approximately 11 MPa to 66 MPa, compared to 4–18 MPa for (both ranges for finite average strand lengths only). These values are similar to the ranges found in AFM measurements on E. coli, i.e., = 35–60 MPa perpendicular to the cell axis and 15–30 MPa parallel [27], as well as other theoretically derived elasticities [24], [29] (see Fig. 4). The glycan elasticity increases with average strand length, and for infinite strand lengths, it grows to as much as 200 MPa (see Inf1 and Inf2 in Table 1). , on the other hand, has no apparent correlation with average strand length.
The relationship between the Poisson's ratios, specifically that for all models, indicates that strain in the direction of the glycan strands induces a significant deformation in the peptide direction, but that the reverse is not true. Given that it is known that the glycan strands are aligned with the circumference of the cell and the peptide links with the long axis, the result is that stress applied to the cell wall will be primarily absorbed in the axial direction, leading to a lengthening of the cell, but not an increase in radius, just as observed experimentally [4], [31]. As average strand length increases, both Poisson's ratios decrease, effectively decoupling the two components of the peptidoglycan layer. One result of this decoupling is that the ratio of the two elasticities, , increases monotonically with average strand length.
Besides elasticity, other distinguishing physical characteristics of the cell wall as a whole include its thickness, the size of pores within it, the ordering of its strands, and the area per disaccharide. Using the five model patches developed, all of these characteristics were determined under different applied strains and compared to experimental measurements (see Table 2).
Based on different techniques, the thickness of the E. coli peptidoglycan layer has been assigned a range of values, including 2.5 nm (small-angle neutron scattering [32]), 6.0 nm (AFM [27]), and 6.4 nm (cryo-electron microscopy [33]). More recent ECT experiments estimated the thickness to be 4 nm at most [20] and AFM experiments measured ∼2 nm [34]. In contrast, cell walls from Pseudomonas aeruginosa appear even thinner, ranging from 2.4 nm (cryo-EM [33]) to 3.0 nm (AFM [27]), while that from Caulobacter crescentus is up to 7-nm thick [11].
The thickness in the simulated constructs was determined in two different ways. First, the mass density as a function of , the coordinate orthogonal to the plane of the cell wall, was measured. This density was calculated for all the heavy atoms in the peptidoglycan, averaged over each trajectory. Values for the thickness were taken as the width of the density profile at 10% of its peak (see Fig. 5E). For the relaxed cell walls, the thickness ranged from 3.4–3.9 nm, in agreement with the most recent ECT measurements [11]. Under strain, this thickness decreased by up to 20%.
As a complementary measure of thickness, pressure profiles as a function of were measured for the patch in each simulation. The thickness was then taken to be the stress-bearing part of the wall, i.e., that fraction of the simulation system with a significantly increased pressure compared to bulk water (see Fig. S1 in SI). This thickness was as much as 1–1.5 nm less than that derived from the mass density. Such a result is unsurprising, as the free peptide chains, which project outward from the cell wall, will contribute to the mass density but do not bear any stress (see SI for a detailed discussion of the pressure profile calculations).
The maximum pore size in the cell wall in different states has been measured indirectly by determining the largest objects that can pass through it. For example, fluorescently labeled dextran molecules were used to estimate the pore radius in the E. coli wall as 2.06 nm in the relaxed state [5]. In another study, proteins up to 100 kDa in size were released from osmotically shocked cells, giving an estimated pore radius of 3.1 nm for stretched peptidoglycan [35]. For the cell walls with finite glycan strand lengths studied here, i.e., avg8, avg17, and avg26, the maximum pore radius averaged over time was 2.05 to 2.44 nm in the relaxed state (see Fig. S4). This radius almost uniformly increased under strain, with the maximum observed being 3.43 nm, in agreement with the large pore size observed in the osmotic-shock experiments [35]. These pore sizes fall in the same range as those observed in CG simulations [23], although neither captures the very large pores (10-nm diameter) observed in recent AFM experiments [34], most likely explained by cell wall-spanning macromolecular machinery [36].
ECT images of frozen E. coli sacculi have revealed a lack of significant ordering of the glycan strands in the relaxed state [11], although some question remains as to whether this disorder persists when the cell wall is under tension as in a living cell [1]. For the simulated cell walls, the ordering was quantified by measuring the angle between segments of each strand and the circumferential axis (see Fig. S7). The temporally and spatially averaged angle was typically close to as expected, although persistently off-axis. The standard deviation in the angle was found to be significant at around 20– (see Table 2). For avg8, avg17, and avg26, the difference in this deviation was minimal under strain when compared to the relaxed state, suggesting that the wall does not become more ordered under tension. In contrast, for constructs Inf1 and Inf2, tension in the direction of the strands notably decreases their off-axis fluctuations. The sensitivity of these fluctuations to tension is due to the strands' inability to redirect the applied stress into the peptide cross-links, reflected also in their reduced Poisson's ratios (see Table 1).
The average surface area per disaccharide has been estimated at based on the number of pm molecules in a given bacterium [37]. We note that this area can depend on a variety of factors, however, including ion concentration, pH, and the presence of denaturants [2], [4], although in the current study, only neutralizing ions are used (see Methods). For the different patches examined here, the unit area for the relaxed cell wall ranges from 2.6 to , and rises to 3–4 under strain (see Table 2). While at first, the discrepancy between experiment and modeling appears large, it should be noted that the cell wall may not be uniformly single layered, with up to three layers in some regions predicted [32]. Peptidoglycan in these additional layers would serve to lower the effective unit area for a single-layered cell wall. Considering a range of possible cell walls from completely single-layered to completely double-layered gives a range of possible unit areas of 2.5–5.0 . If one assumes that the initial strand spacing used during modeling is linearly related to the resulting unit area, this implies that the average strand spacing can be no less than 2 nm (unit area of 2–2.67 ) and no more than 4 nm (unit area of 4–5.33 ).
The native architecture and organization of the bacterial cell wall are largely inaccessible to direct imaging techniques, though at the very edge of the resolution of ECT, glycan strands could be discerned in a Gram-negative sacculus. These strands were, nonetheless, fragmented, and the cross-links were indiscernible [11]. Furthermore, the imaged samples are no longer part of living cells. For these reasons, modeling fills a critical gap between biochemical data on the cell wall's constituents and biophysical data on its macroscopic properties. In this paper, patches of an E. coli cell wall were made using a circumferential layered model, supported by ECT imaging of both Gram-negative and Gram-positive sacculi [4], [11], plausibility arguments based on the thickness and glycan strand length [1], and the average peptide-peptide angle measured above (see Fig. 2B). The patches were constructed using only a few initial parameters, including the initial strand spacing (roughly 3 nm), degree of cross-linking (50%), and average glycan strand length (between 8 and 26 disaccharides).
In the simulated cell-wall patches, peptidoglycan was found to be relatively inelastic in the direction of the glycan strands, while very elastic in the direction of the peptide cross-links. The calculated Young's moduli for the two directions, and , respectively, were found to be in good agreement with multiple AFM measurements [27], [29], with the best agreement being found for the constructs avg17 and avg26 (see Table 1). The average glycan strand lengths for these two constructs also match those measured experimentally for E. coli cells in the stationary (17.8) and exponential growth (25.8) phases [20]. determined for live bacteria grown in aragose gel (5–15 MPa) was notably higher than our calculations, although other factors such as the outer membrane stiffness or incomplete gel polymerization may have inflated the number [30].
Further evidence of the relative stretchability of the peptide cross-links compared to the glycan strands comes from the decrease in Poisson's ratios as average glycan strand length increases. At short lengths, there is a significant coupling between the peptide cross-links and the glycan strands, allowing the former to absorb stress from the latter. At longer lengths, however, strain applied to the glycan strands is primarily absorbed by the strands alone, which, due to their inability to stretch much beyond their initial lengths, induces a large stress in the cell wall in their direction. This resistance to expansion, thus, does not depend on the peptide cross-links but is intrinsic to the glycan strands. Indeed, an intriguing suggestion is that longer glycan strands can compensate for a decrease in cross-linking percentage to maintain cell integrity [38]. While the fraction of peptides in cross-links was fixed near 50% for all models here, is directly related to the glycan strand length, whereas is independent. Although it remains to be shown, we hypothesize that, conversely, will be more sensitive than to the degree of cross-linking.
Beyond elasticity several other quantifiable properties were measured from the simulations, including the cell-wall thickness, maximum pore radius, and unit area per disaccharide. Excellent agreement with experimentally determined thicknesses [11] and pore sizes [5], [35] was found. The unit area measured in simulations (2.6–4 ) implies a cell wall that is more sparse than that estimated from experiment (2.5 ). However, those experimental estimates are based on quantifying the total number of pm molecules per cell, irrespective of their place in the cell wall [37]. Neutron-scattering experiments have led to the suggestion that the Gram-negative cell wall is primarily a single layer, but includes regions of up to three layers over 25% of the surface [32]. The excess peptidoglycan in these additional, but limited and incomplete, layers would raise the experimental unit area for a single layer to 3.75 , in significantly better agreement with that from the models examined here, also supporting the choice of initial strand spacing of 3 nm.
The average angle of the glycan strands with respect to the circumferential axis was found to be near , although the standard deviation was typically 20–, even under tension (see Table 2). The lack of alignment amongst the strands argues in favor of a disordered circumferential model, as previously indicated by ECT [11]. A chiral patterning of peptidoglycan has been suggested based on recent experiments, and was attributed to a helical movement of MreB, a proposed cytoskeletal protein [39], [40]. Recent total internal reflection fluorescence and ECT experiments have indicated, however, that MreB moves circumferentially around the cell and does not form long filaments [41]–[44]. The glycan-strand angle measured here was often negative (range of − to ), which hints at a slight intrinsic chirality in stressed peptidoglycan networks, irrespective of their assembly. Whether this could explain the experimental results remains unclear.
The widespread agreement between simulation and experiment for all of the aforementioned properties, including elasticity, thickness, pore size, and unit area, serves to validate the connection made between the modeled atomic-scale properties of peptidoglycan and the macro-scale properties probed experimentally. The present molecular models support a cell wall composed predominantly of a single layer of peptidoglycan with glycan strands running circumferentially around the cell in a disordered fashion. Furthermore, assuming our model is correct, we predict that the disorder, which is primarily due to the random orientation of the peptide cross-links relative to the strands, persists under native cellular conditions. While we do not consider possible growth mechanisms here in detail, the insertion of new peptidoglycan strands has been predicted to be a function of such disorder, as well as mechanical tension and MreB [45], [46]. To examine tension-dependent insertion, we also created a peptidoglycan patch in which one strand was deleted, tension applied, and then the strand was added back to the gap that formed. Because some cross-links between the re-added strand and the rest of the patch formed in alternate locations, a slight decrease in the degree of connectivity resulted and one larger pore was observed (radius of 3.6 nm vs. 2.9 nm; see Text S1 and Fig. S5 for more details). Because this pore may serve as a site for addition of the next peptidoglycan strand, it cannot be assumed that larger pores are an inevitable product of tension-dependent insertion. However, over repeated growth cycles, the insertion mechanism used is likely to become increasingly relevant to the large-scale structure that develops.
While a number of other simulations of bacterial cell walls have been carried out in recent years [23],[39],[47], they are all highly coarse grained (CG), a necessary approach for modeling complete sacculi. Coarse graining the system requires, however, that one make a number of assumptions about the properties of individual “beads” in the CG model, including what underlying atoms they represent, how they are connected and interact with each other, and how they are affected by the surrounding environment, e.g., solvent. Where possible such assumptions are rooted in experimental data, although the reliability of those data and their conversion to model parameters is not always straightforward. On the other hand, models built starting from the atomic scale, in which the parameters are not specialized for each application, can utilize the same experimental data for validation, as done here. The atomic-scale model is limited in size compared to the CG models, however, and therefore cannot fully reproduce distributions in strand length [23] nor capture structural features beyond the modeled scale, e.g., pore sizes up to 10-nm in diameter [34]; additionally, a visual comparison of the previous CG models with the atomic-scale models here suggests that the latter models are still too ordered, likely a remnant of the initial construction [23]. Thus, future iterations will be used to probe more realistic growth models, the dependence of cellular-scale properties on the cross-linking fraction and strand spacing, and also the interactions of the network with various growth and remodeling enzymes and embedded proteins.
Force-field parameters for GlcNAc were developed by linking glucose and acetamide, with those charges and parameters near the interface determined. Similarly, MurNAc parameters were developed by linking GlcNAc with lactic acid. Charges of interfacial atoms, namely C2 on the sugar ring and the the NH group on the acetamide side chain in both residues along with C3 on the ring, the in the lactic acid side chain, and the bridging O3 oxygen in MurNAc, were modified. These charges were determined from ab initio quantum chemical calculations using a pre-release version of the Force Field Toolkit (ffTK) plugin for VMD, following the CHARMM parametrization procedures [48], [49]. Bond, angle, and dihedral parameters involving the interfacial atoms were similarly determined. Because D-isoglutamate and pm are nearly identical to their standard amino-acid counterparts, glutamate and lysine, their parameters were developed solely by analogy. The complete topology and parameter set used for subsequent simulations is provided in Text S2 and S3.
Because simulating the actual transpeptidase reactions is prohibited by both current knowledge of the order of events and available computational resources, a procedure was developed to build the peptidoglycan network with a statistical view of the general organization. In the first step, a set of E. coli peptidoglycan strands with the number of disaccharides chosen according to a random Gaussian distribution of specified mean are placed parallel to one another separated by a given distance (typically 2–3 nm, with a 0.5 nm random deviation). Each system is fully solvated in explicit water and sufficient ions were added to the solution to neutralize the high negative charge in the peptidoglycan. The final atom count ranged from 100,000 to 545,000 atoms. Initially, the glycan strands are held fixed for a 2-ns simulation while the peptides are left free to move. Next, the trajectory is analyzed to find when each available pm -nitrogen first comes near an available D-Ala carbonyl oxygen, and for what fraction of time they are within this distance. Finally, the list of possible links is ordered according to the first contact using a more stringent distance criterion along with a minimum time within range. Links are then added, in order, such that when a given pm or D-Ala residue is linked, its entire peptide is removed from further consideration. The time and distance criteria are chosen to target roughly 50% cross-linking overall, as typically observed for E. coli [1], [20].
The cross-linked peptidoglycan network is first relaxed using energy minimization, and then allowed to equilibrate during MD simulations with no applied restraints. It should be noted that the network is periodic, with glycan strands as well as peptides covalently linked across the simulation system's periodic boundaries, thus mimicking a much larger patch of cell wall (see Fig. 3). The resulting network is simulated for at least 20 ns under constant pressure conditions, which allows its dimensions to fluctuate. The relaxed in-plane dimensions of each patch were taken as the average over the last 10 ns. These dimensions are: 9.30.218.10.5 (avg8), 18.70.233.40.25 (avg17), 17.20.451.20.4 (avg26), 18.60.313.60.1 (Inf1), and 16.80.327.40.2 (Inf2).
All simulations were run with the molecular dynamics package NAMD 2.9 [50] and the CHARMM force field [51]–[53]. A constant temperature of 310 K was held using Langevin dynamics; a pressure of 1 atm in the direction normal to peptidoglycan layer was maintained with a Langevin piston [54]. A 2-fs time step was utilized, with short-range non-bonded interactions (12-Å cutoff) evaluated every time step and long-range electrostatics every two time steps using the particle-mesh Ewald method [55]. All figures were made using VMD [56].
|
10.1371/journal.pgen.1004280 | Isl1 Directly Controls a Cholinergic Neuronal Identity in the Developing Forebrain and Spinal Cord by Forming Cell Type-Specific Complexes | The establishment of correct neurotransmitter characteristics is an essential step of neuronal fate specification in CNS development. However, very little is known about how a battery of genes involved in the determination of a specific type of chemical-driven neurotransmission is coordinately regulated during vertebrate development. Here, we investigated the gene regulatory networks that specify the cholinergic neuronal fates in the spinal cord and forebrain, specifically, spinal motor neurons (MNs) and forebrain cholinergic neurons (FCNs). Conditional inactivation of Isl1, a LIM homeodomain factor expressed in both differentiating MNs and FCNs, led to a drastic loss of cholinergic neurons in the developing spinal cord and forebrain. We found that Isl1 forms two related, but distinct types of complexes, the Isl1-Lhx3-hexamer in MNs and the Isl1-Lhx8-hexamer in FCNs. Interestingly, our genome-wide ChIP-seq analysis revealed that the Isl1-Lhx3-hexamer binds to a suite of cholinergic pathway genes encoding the core constituents of the cholinergic neurotransmission system, such as acetylcholine synthesizing enzymes and transporters. Consistently, the Isl1-Lhx3-hexamer directly coordinated upregulation of cholinergic pathways genes in embryonic spinal cord. Similarly, in the developing forebrain, the Isl1-Lhx8-hexamer was recruited to the cholinergic gene battery and promoted cholinergic gene expression. Furthermore, the expression of the Isl1-Lhx8-complex enabled the acquisition of cholinergic fate in embryonic stem cell-derived neurons. Together, our studies show a shared molecular mechanism that determines the cholinergic neuronal fate in the spinal cord and forebrain, and uncover an important gene regulatory mechanism that directs a specific neurotransmitter identity in vertebrate CNS development.
| Neurons utilize various chemicals to transmit signals to a target cell. Distinct types of neurons in the spinal cord and forebrain, collectively termed cholinergic neurons, utilize the same chemical, acetylcholine, for signal transmission. These neurons play critical roles in controlling locomotion and cognition. In this study, we have found that the Isl1 gene orchestrates the process to generate cholinergic neurons in the spinal cord and forebrain. Isl1 forms two different types of multi-protein complexes in the spinal cord and forebrain. Both complexes bind the same genomic regions in a group of genes critical for cholinergic signal transmission, and promote their simultaneous expression. These cholinergic genes include enzymes that synthesize acetylcholine and proteins required to package acetylcholine into vesicles. The Isl1-containing multi-protein complexes were able to trigger the generation of cholinergic neurons in embryonic stem cells and neural stem cells. Our study reveals crucial mechanisms to coordinate the expression of genes in the same biological pathway in different cell types. Furthermore, it suggests a new strategy to produce cholinergic neurons from stem cells.
| The choice of neurotransmitter is one of the most fundamental aspects of neuronal fate decision. Cholinergic neurons are located in diverse regions of the CNS, which do not share the developmental origin, and regulate complex behaviors. In the spinal cord, cholinergic motor neurons (MNs) control locomotion, whereas in the forebrain, cholinergic neurons regulate cognitive processes [1], [2]. Defects in function or survival of cholinergic neurons result in severe human pathologies, including spinal cord injuries, diseases associated with impaired motor function and cognitive disorders resulting from the loss of forebrain cholinergic neurons (FCNs) [3]. Despite the crucial roles of cholinergic neurons in human physiology and pathology, the mechanisms that specify cholinergic neuronal cell fate throughout the CNS during vertebrate development remain largely unknown.
The cholinergic neurotransmission system requires the function of several key factors that are highly expressed in all cholinergic neurons, termed cholinergic pathway genes (Fig. 1A) [4], [5]. Understanding the gene regulatory mechanisms that control the expression of cholinergic pathway genes in different groups of cholinergic neurons will provide crucial insights into the process of cholinergic fate specification in CNS development. Given that each of the cholinergic pathway genes is essential for efficient cholinergic neurotransmission, it is probable that they are up-regulated in a coordinated fashion as neurons acquire cholinergic neuronal identity during vertebrate development. Supporting this possibility, the vesicular acetylcholine transporter (VAChT, also known as Slc18a3) gene is encoded within an intron of the choline acetyltransferase (ChAT) gene in all metazoans examined thus far, including C.elegans, Drosophila and mammals [6]. This unique genomic arrangement suggests that the ChAT and VAChT genes are co-regulated by a single set of transcription factors. Furthermore, in a subset of cholinergic MNs of C. elegans, an Ebf-type transcription factor UNC-3 regulates a battery of cholinergic genes via a shared UNC-3-response motif [7].
Two critical questions remain to be answered. First, is a battery of cholinergic pathway genes coordinately regulated by a common transcription factor in vertebrate CNS, similar to UNC-3-directed control of cholinergic genes in C.elegans? Second, could there be a transcription factor(s) that determines cholinergic fate across different types of cholinergic cells in the vertebrate CNS? While very limited information is available for the first question, it is interesting to note, for the latter question, that a LIM homeodomain (LIM-HD) transcription factor Isl1 is expressed in several cholinergic neurons in the spinal cord, hindbrain, forebrain and retina, such as spinal MNs, hindbrain MNs, some FCNs, and starburst amacrine cells [8], [9], [10], [11]. Deletion of Isl1 gene results in a loss of MNs in the spinal cord and hindbrain [12]. Conditional deletion of Isl1 gene using a Six3-Cre transgene led to a reduction of restricted FCNs in the brain and cholinergic amacrine cells in the retina [13]. These findings point to the possibility that Isl1 may function as a cholinergic fate determinant in vertebrate CNS. However, it remains unknown whether Isl1 directly control the cholinergic phenotype and, if so, how Isl1 controls the fate of distinct cholinergic cell types whose gene expression patterns and functions are vastly different despite the shared property of cholinergic neurotransmission.
In the developing spinal cord, Isl1 directs motor neuron fate specification by cooperating with another LIM-HD factor Lhx3 [12], [14], [15], [16]. In differentiating MNs, Isl1 binds to Lhx3 and a LIM-interactor NLI (also known as Ldb), thereby forming the Isl1-Lhx3-hexamer complex, also termed MN-hexamer (Fig. S1A) [14], [17]. The combinatorial expression of Lhx3 and Isl1, resulting in the formation of the Isl1-Lhx3-hexamer, is capable of triggering MN specification in chick spinal cord, embryonic stem cells (ESCs), and induced pluripotent stem cells [14], [17], [18], [19], [20]. However, it is unclear whether the Isl1-Lhx3-hexamer directly controls cholinergic neuronal identity, an essential characteristic of MNs. In the developing forebrain, FCNs are derived from the medial ganglionic eminence (MGE) in the ventral telencephalon [21], [22]. A LIM-HD protein Lhx8 is highly expressed in the MGE [21], [23]. The formation of FCNs is severely disrupted in Lhx8-deficient mice [24], [25], [26]. Lhx8 appears to function in combination with Isl1 in driving the differentiation of cholinergic striatal interneurons [27], but the mechanisms by which Lhx8 and/or Isl1 control cholinergic fates in the developing forebrain remain unclear.
In this study, we found that the Isl1-Lhx3-hexamer directly activates the expression of a suite of cholinergic genes by binding to cholinergic gene enhancers that were discovered via ChIP-seq experiments. We also found that Isl1 is co-expressed with Lhx8 and NLI in the embryonic ventral forebrain and forms a hexamer complex with Lhx8 and NLI, named Isl1-Lhx8-hexamer. Interestingly, like the Isl1-Lhx3-hexamer in the spinal cord, the Isl1-Lhx8-hexamer directly controls cholinergic pathway gene expression via the same cholinergic gene enhancer in the forebrain. These findings imply that, despite distinct developmental histories and locations within the nervous system, MNs and some FCNs employ a common molecular mechanism that determines their cholinergic neuronal identity.
Given that MNs acquire cholinergic neuronal characteristics as they become specified, we considered the possibility that the Isl1-Lhx3-hexamer, a determinant of the MN fate, regulates expression of a battery of cholinergic genes by directly binding to the enhancer of each cholinergic gene. Intriguingly, our ChIP-seq analysis, which mapped the genomic binding sites of the Isl1-Lhx3-hexamer in mouse embryonic stem cells [20], revealed Isl1-Lhx3-bound peaks in the key cholinergic pathway genes; ChAT, VAChT, high affinity choline transporter (CHT, also known as Slc5a7), a transporter that regulates the uptake of choline from the synaptic cleft into cholinergic neurons, and ATP-citrate lyase (Acly), an enzyme that synthesizes acetyl-CoA (Fig. 1A–C). ChAT has a strong peak within an intronic region that lies downstream of the VAChT gene, which is itself encoded within the intron of the ChAT gene. The Acly gene has two strong peaks in intronic regions and one upstream peak, while the CHT gene has a peak ∼100 kb downstream of its coding region. All the peaks have at least one hexamer response element (HxRE) (Fig. 1C, Fig. S1B) [20].
To test whether the Isl1-Lhx3-hexamer is recruited to Isl1-Lhx3-bound peak regions of the cholinergic genes in vivo, we purified genomic DNA bound by the Isl1-Lhx3-hexamer from E12.5 embryonic spinal cords using ChIP assays with α-NLI, α-Isl1, and α-Lhx3 antibodies. All three components of the Isl-Lhx3-hexamer bound to the peaks in the cholinergic genes, while they did not bind to the genomic regions without the peaks (Fig. 1D), indicating that the endogenous Isl1-Lhx3-hexamer is recruited to the cholinergic pathway genes in the developing spinal cord. Together, our unbiased, genome-wide ChIP-seq data, along with in vivo ChIP results, strongly suggest that the cholinergic pathway genes are directly activated by the Isl1-Lhx3-hexamer during MN fate specification.
To test whether the Isl1-Lhx3-hexamer is capable of inducing the expression of multiple cholinergic genes in embryonic spinal cord, we misexpressed Isl1 and/or Lhx3 in the chick neural tube and monitored the expression of cholinergic genes. Co-electroporation of Isl1 and Lhx3 triggered the ectopic expression of a panel of cholinergic genes, including ChAT, VAChT, Acly and CHT, in the dorsal neural tube, while electroporation of Isl1 or Lhx3 alone did not (Fig. 2A, Fig. S2, data not shown). These data indicate that the Isl1-Lhx3-hexamer is capable of upregulating the cholinergic pathway genes in the developing spinal cord.
To test whether Isl1 is needed for the cholinergic neuronal differentiation in the developing CNS, we deleted the Isl1 gene in neural progenitors using nestin-Cre [28], [29]. In E12.5 Isl1f/f;nestin-Cre mice, Isl1 expression in MNs in the ventral spinal cord was greatly reduced (Fig. 2B). In this condition, expression of cholinergic genes, such as Acly, ChAT, VAChT and CHT, is drastically downregulated (Fig. 2B). The weak signal of VAChT was detected only in the remaining Isl1-expressing cells of Isl1f/f;nestin-Cre mice (Fig. 2B). These results support a role of Isl1 in controlling cholinergic fate decision in the spinal cord.
To test whether the Isl1-Lhx3-hexamer binding sites in the cholinergic genes act as enhancers to activate the cholinergic pathway genes in the embryonic spinal cord, we first examined whether the Isl1-Lhx3-hexamer activates the transcription of a reporter gene linked to each cholinergic gene peak, referred to here as ChAT-enh, Acly-enh1 and CHT-enh (Fig. 3A–E), using luciferase reporter assays in mouse embryonic P19 cells. As NLI is expressed endogenously in P19 cells, co-expression of exogenous Isl1 and Lhx3 leads to the formation of the Isl1-Lhx3-hexamer [17], [30]. The co-expression of Isl1 and Lhx3 strongly activated the Acly:LUC, ChAT:LUC and CHT:LUC reporters, but not LUC vector alone, in P19 cells, whereas the expression of Isl1 or Lhx3 alone did not (Fig. 3A–C). The Acly:LUC reporter with point mutations in the HxRE was not activated by the co-expression of Isl1 and Lhx3 (Fig. 3A). The DNA-binding defective forms of Lhx3 or Isl1 failed to synergize to activate Acly:LUC (Fig. 3A), indicating that DNA-binding activity of both Isl1 and Lhx3 is needed to activate the reporter. To further test the role of the HxRE in each enhancer for the potent transcription response to the combination of Isl1 and Lhx3, we generated luciferase reporters that are linked to multiple copies of the HxRE found within Acly-enh1, ChAT-enh, and CHT-enh, respectively. These minimal HxRE reporters were also highly activated by co-expression of Isl1 and Lhx3 (Fig. 3D, E, data not shown), establishing that the HxRE motif mediates activation of the cholinergic enhancers by the Isl1-Lhx3-hexamer. These data establish that the Isl1-Lhx3-hexamer is capable of activating each cholinergic enhancer in the Acly, ChAT/VAChT and CHT genes in heterologous cell types.
To identify in vivo cell types in which the cholinergic enhancers activate gene expression in the developing spinal cord, we electroporated the neural tube of chick embryos with the GFP reporters linked to each cholinergic enhancer in ovo at a time when MNs are being specified. Interestingly, the Acly-enh1 drove strong GFP expression in MNs within the developing spinal cord (Fig. 3F, S3A). In contrast, GFP was not expressed in non-MN cell types, despite efficient transfection of those cells following in ovo electroporation (Fig. 3F, Fig. S3A), suggesting that only MNs have the transcriptional machinery that allows activation of Acly-enh1. Additionally, Acly-enh1 with point mutations in the HxRE motif failed to activate target gene expression in MNs (Fig. 3G, S3B), demonstrating that the HxRE motif is responsible for the MN-specific enhancer activity of the Acly-enh1. Furthermore, the multimerized HxRE motifs from the Acly-enh1 were sufficient to drive GFP reporter expression in MNs (Fig. 3H, Fig. S3C). Thus, the HxRE motif is necessary and sufficient for the MN-specific enhancer activity of the Acly-enh1, suggesting that the endogenous Isl1-Lhx3-hexamer is responsible for Acly enhancer activity. Consistent with this idea, co-expression of Isl1 and Lhx3, which assembles the Isl1-Lhx3-hexamer with endogenous NLI and triggers the formation of ectopic MNs in the dorsal spinal cord [14], ectopically activated both Acly-enh1:GFP and Acly-HxRE:GFP reporters, but it failed to activate Acly-enh1 with mutations in HxRE motif (Fig. 3F–H). These data indicate that the Isl1-Lhx3-hexamer is able to activate the Acly enhancer in the dorsal neural tube. The expression of Isl1 or Lhx3 alone failed to increase the transcriptional activity of the Acly-enh1 or Acly-HxRE (Fig. S3D, E). Similarly to the Acly enhancer, the ChAT enhancer also directed gene expression specifically to MNs and became ectopically activated by the co-expression of Isl1 and Lhx3 (data not shown). Together, these data demonstrate that the endogenous Isl1-Lhx3-hexamer binds to the cholinergic enhancers via the HxRE motifs and triggers the transcription of their target cholinergic genes as MNs are specified in embryonic spinal cords, establishing the Isl1-Lhx3-hexamer as a critical determinant of cholinergic neuronal identity in MNs.
The coordinated upregulation of cholinergic pathway genes by the Isl1-Lhx3-hexamer in differentiating MNs, along with the previous loss-of-function studies suggesting that Isl1 and Lhx8 play important roles in the generation of FCNs [13], [24], raises the possibility that Isl1 and Lhx8 might form a complex similar to the Isl1-Lhx3-hexamer that drives cholinergic neuronal fate in the developing forebrain. To understand the role of Isl1 and Lhx8 in cholinergic gene expression in the developing forebrain, we determined the expression pattern of Isl1, Lhx3, and and NLI, which might form an Isl1-Lhx3-hexamer-like complex, using double immunohistochemistry analyses. All FCN precursors arise from the Nkx2.1-expressing MGE and preoptic area (POA) in the ventral telencephalon, and take two distinct migratory pathways; tangential migration to form striatal interneurons in the caudate-putamen (CPu) and radial migration to generate projection neurons in the basal forebrain (Fig. 4A) [21], [31]. In E12.5 forebrain, Lhx8 expression is largely confined within the subventricular zone (SVZ) and mantle zone (MZ) of the MGE, but a few Lhx8+ cells were found in the MZ of the lateral ganglionic eminence (LGE), which are likely the cells tangentially migrating from the MGE (Fig. 4B, C). In contrast, Isl1 is more abundantly expressed in the SVZ and MZ of the LGE, but is also expressed in the SVZ and MZ of the MGE (Fig. 4C, D). NLI is highly expressed in both MGE and LGE (Fig. 4D, E). Thus, Isl1, Lhx8 and NLI are co-expressed in a substantial fraction of cells in the MGE and LGE at E12.5. A similar expression pattern for Isl1, Lhx8, and NLI was observed in E13.5 forebrain (data not shown).
By E16.5, VAChT+ cholinergic neurons were readily detectable in the CPu and basal meganocellular complex (BMC) (Fig. 5A, 6). Although a majority of CPu cells are derived from the LGE, Nkx2.1+ progenitors in the MGE produce distinct subtypes of striatal interneurons in the CPu, including cholinergic interneurons [21], [31]. Most cells in BMC, where a subset of cholinergic projection neurons is located, are generated from the Nkx2.1+ MGE [21], [31]. In E16.5 brains, Isl1+ cells were much more abundant in the CPu than in the BMC, whereas Nkx2.1+ and Lhx8+ cells were more abundant in the BMC than in the CPu (Fig. 5B, C, F, G), correlated with their expression at earlier developmental time points (Fig. 4). Despite this distinct pattern of gross expression, a number of Isl1+ cells co-expressed Nkx2.1 and Lhx8 in both the CPu and BMC, as shown by double immunohistochemistry analyses (Fig. 5B, C, F, G). Given that Nkx2.1+ and Lhx8+ striatal interneurons are originated in the MGE [22], [31], a subset of Isl1+ cells in the CPu, which co-express Nkx2.1 and Lhx8, is likely interneurons that are produced from the MGE.
Isl1f/f;nestin-Cre mice die soon after E12.5, precluding us from observing cholinergic neuronal differentiation in the forebrain. To understand the role of Isl1 in FCN specification in the forebrain, we generated Isl1f/f;Nkx2.1-Cre mice, in which Isl1 gene is deleted in cells derived from the Nkx2.1-expressing MGE [31]. As expected, in E16.5 Isl1f/f;Nkx2.1-Cre mice, the number of Isl1+ cells was greatly reduced in the BMC, but not in the CPu where most of Isl1+ cells were derived from LGE and thus did not express Nkx2.1-Cre (Fig. 5B–I). In the CPu and BMC of the Isl1-conditional mutants, neither Isl1/Nkx2.1-double positive cells nor Isl1/Lhx8-co-expressing cells were found (Fig. 5B–I), indicating that Isl1 is deleted in cells produced from Nkx2.1+ MGE and that Isl1/Lhx8-co-expressing cells in the CPu and BMC are derived from the MGE. Interestingly, in the CPu of Isl1f/f;Nkx2.1-Cre embryos, Lhx8+ and Nkx2.1+ interneurons were significantly reduced by ∼62% and ∼43%, respectively (Fig. 5J, K), suggesting that Isl1 is required for specification of a subset of Nkx2.1+/Lhx8+ striatal interneurons.
To monitor cholinergic neuronal differentiation, we performed immunostaining assays with VAChT antibodies. At E16.5, cholinergic neurons were detected in the CPu and BMC, and co-expressed Isl1 and Lhx8 in both areas (Fig. 6A–D). In E16.5 Isl1f/f;Nkx2.1-Cre mice, however, cholinergic neurons were almost eliminated in the CPu (Fig. 6A, B). While ∼44% Lhx8+ cells and ∼35% Nkx2.1+ cells were cholinergic in the CPu of control embryos, almost all of Lhx8+ and Nkx2.1+ neurons did not express VAChT in the CPu of Isl1f/f;Nkx2.1-Cre mice (Fig. 6E, F). The number of cholinergic neurons in the BMC area of Isl1f/f;Nkx2.1-Cre embryos appeared to be reduced compared to that in the littermate controls, but the heterogeneity of cholinergic neurons in the BMC made the quantification very challenging (Fig. 6C, D). The remaining cholinergic neurons in the BMC expressed Lhx8 (Fig. 6D). Similar to E16.5, the number of cholinergic neurons remained markedly decreased in the CPu of E17.5 and P2 mice (Fig. S4).
Together, our data indicate that Isl1 and Lhx8 are co-expressed in at least two different populations of FCNs in the CPu and BMC, and that Isl1 function in the MGE-derived cells is required for the specification of cholinergic interneurons in the CPu during forebrain development.
The co-expression of Isl1 and Lhx8 in FCN precursors and FCNs and requirement of Isl1 and Lhx8 for the specification of a subset of FCNs (Fig. 5, 6, S4) [13], [24] support the possibility that Isl1 and Lhx8 cooperate for the FCN specification by forming a hexamer complex (Fig. 7A), similar to the Isl1-Lhx3-hexamer (Fig. S1). The Isl1-Lhx3-hexamer assembly is dependent on the ability of Lhx3 to interact with Isl1 [14]. Another LIM-HD factor Lhx1 interacts with NLI, a common cofactor of the LIM-HD transcription factors, but does not bind to Isl1, thus forming only a typical LIM tetramer complex consisting of 2NLI:2Lhx1 [14]. Thus, we investigated whether Lhx8 can interact with Isl1, like Lhx3, using in vitro GST-pull down assays (Fig. 7B). As previously shown, Lhx3 interacted with Isl1 as well as NLI, whereas Lhx1 bound only to NLI. Interestingly, Lhx8 strongly associated with both Isl1 and NLI in vitro. Lhx8 also interacted with both Isl1 and NLI in HEK293 cells (Fig. 7C). Combined with the notion that NLI strongly self-dimerizes [32], our data supports a model by which Lhx8, Isl1 and NLI can form a hexameric complex consisting of two NLIs, two Isl1s and two Lhx8 molecules (Figure 7A). We refer to this complex as the Isl1-Lhx8-hexamer.
To further test the formation of Isl1-Lhx8-hexamer complex, we examined whether Lhx8 interacts with NLIDD-Isl1ΔLIM, in which the dimerization domain (DD) of NLI is fused to LIM domains-deleted Isl1 (Figure 7D). As NLIDD-Isl1ΔLIM lacks the LIM-interaction domain (LID) of NLI, it cannot bind to LIM-HD factors via typical interaction interfaces between NLI-LID and LIM-domains of LIM-HD factors, which lead to tetramer formation. Co-immunoprecipitation assays revealed that NLIDD-Isl1ΔLIM associated with Lhx8 in cells despite the lack of NLI-LID in this fusion (Fig. 7D), further supporting the formation of the Isl1-Lhx8-hexamer in cells. Together, along with the fact that Lhx8, Isl1 and NLI are co-expressed in FCN precursors in the ventral telencephalon, these results suggest that Lhx8, Isl1, and NLI form the Isl1-Lhx8-hexamer complex in the ventral telencephalon during development.
To investigate whether the Isl1:Lhx8 dimer, the DNA-binding unit of the Isl1-Lhx8-hexamer, recognizes specific DNA sequences, we performed the unbiased screening method SELEX (for systematic evolution of ligands by exponential enrichment) assay with Isl1, Lhx8, or an Isl1-Lhx8 fusion, in which full-length Isl1 and Lhx8 proteins were linked by a flexible short linker (Fig. 7E). Isl1-Lhx8 highly enriched a 15 nucleotide-long Isl1:Lhx8-binding motif after the third round of SELEX reaction, while Isl1 or Lhx8 failed to enrich any specific DNA sequences. The same motif was also isolated by SELEX with the mixture of Isl1 and Lhx8, which were translated from Isl1-T2A-Lhx8 construct in vitro (Fig. 7F), indicating that the Isl1-Lhx8-binding motif is not an artifact caused by use of the Isl1-Lhx8 fusion protein. These data indicate that Isl1:Lhx8 dimer in the Isl1-Lhx8-hexamer has high affinity to the specific DNA motif. Notably, Isl1-Lhx8-binding motif has a resemblance to the previously identified Isl1:Lhx3-site [17], [20], such as TAAT sequences, but also has unique features (Fig. S5).
Considering the shared function of Isl-Lhx3-hexamer and Isl1-Lhx8-hexamer in inducing cholinergic genes and the similar features of their binding motifs, it is possible that they bind to the same enhancer regions of cholinergic pathway genes. To test whether, in the developing forebrain, the endogenous FCN-hexamer is recruited to the same cholinergic enhancers identified as targets of the Isl1-Lhx3-hexamer in our ChIP-seq analysis, we performed ChIP assays for the Isl1-Lhx8-hexamer using the dissected E15.5 embryonic forebrains and found that Isl1, Lhx8, and NLI, all components of the Isl1-Lhx8-hexamer, bound to the cholinergic enhancers (Fig. 8A). These results suggest that the cholinergic enhancers recruit the Isl1-Lhx8-hexamer in the embryonic forebrain.
To test the effect of the Isl1-Lhx8-hexamer on the transcriptional activity of cholinergic enhancers, we performed luciferase reporter assays in P19 cells using the Acly-HxRE:LUC and ChAT-HxRE:LUC reporter constructs. The co-transfection of Isl1 and Lhx8 activated each cholinergic enhancer, while expression of Isl1 or Lhx8 alone had minimal effect (Fig. 8B, C). These results suggest that the Isl1-Lhx8-hexamer complex triggers the transcriptional activity of cholinergic enhancers.
To investigate the activity of the Acly enhancer in the ventral forebrain, we injected the Acly-HxRE:GFP reporter along with the expression vectors encoding LacZ, Isl1 or Lhx8 into the ventral regions of E15.5 brain slices. The brain slices were then electroporated, cultured in vitro for four days, and examined for GFP expression (Fig. 8D). Among many LacZ+ electroporated cells, only a small number of basal forebrain cells expressed GFP (data not shown). The co-electroporation of Isl1 and Lhx8 along with the Acly-HxRE:GFP reporter drastically increased the number of GFP+ cells and the levels of GFP expression in the ventral forebrain, whereas expression of Lhx8 or Isl1 alone did not exhibit potent effects on Acly-HxRE:GFP. Likewise, the transfection of cortical progenitors using in utero electroporation revealed that the expression of Isl1-Lhx8, but not Isl1 or Lhx8 alone, strongly activates Acly-HxRE in the developing cortex (Fig. S6A). These results indicate that the combinatorial expression of Lhx8 and Isl1 promotes Acly-HxRE enhancer activity in the developing forebrain.
Together, these data suggest that the Isl1-Lhx8-hexamer is sufficient to activate the cholinergic enhancers in heterologous cells and the developing forebrain.
The binding and activation of cholinergic enhancers by Isl1-Lhx3 and Isl1-Lhx8 in the spinal cord and forebrain, respectively, prompted us to ask whether both complexes are capable of inducing the cholinergic gene battery irrespective of rostro-caudal positions within the CNS. To address this question, we misexpressed LacZ, Isl1, Lhx8, Lhx3, Isl1-Lhx8 or Isl1-Lhx3, along with EF1 promoter driven-GFP vector to mark the electroporated cells, in the E13.5 mouse cortex using in utero electroporation, and compared the expression levels of cholinergic genes between electroporated and control cerebral hemispheres at E18.5 using quantitative RT-PCR (Fig. 9A). The expression level of transgenes was higher in electroporated sides than in control sides, as expected (Fig. S7A). The expression of Isl1-Lhx8 substantially induced expression of ChAT, VAChT, and CHT in the cortex, compared to expression of Isl1 or Lhx8 alone (Fig. 9B), indicating that Isl1 and Lhx8 function in combination to induce expression of cholinergic genes in the developing forebrain. Interestingly, Isl1-Lhx3 did not trigger cholinergic gene expression in the forebrain (Fig. 9B), despite its potent activity to induce cholinergic pathway genes in the developing spinal cord (Fig. 2A). Moreover, unlike the spinal cord, Isl1-Lhx3 failed to upregulate MN genes, Isl2, Hb9, and chodl [33], in the forebrain (Fig. S7B, data not shown), suggesting that the Isl1-Lhx3-hexamer is unable to turn on the MN gene program in the forebrain. Together, our results strongly support a model whereby the Isl1-Lhx8-hexamer orchestrates upregulation of a battery of cholinergic pathway genes in the developing forebrain.
Our results that Isl1-Lhx3 failed to upregulate MN genes and cholinergic genes raise the question of whether Isl1-Lhx8-hexamer is functional in the spinal cord. To address this question, we expressed Lhx8, Isl1, Isl1-Lhx8, or Isl1 plus Lhx8 in chick spinal cord using in ovo electroporation, and monitored cell differentiation and cholinergic gene expression three days post-electroporation. Lhx8 triggered ectopic generation of Chx10+ V2a interneurons in the dorsal spinal cord (Fig. 9C, S7C) like Lhx3 [14], underlining the similarity between Lhx8 and Lhx3. Co-expression of Isl1 with Lhx8 blocked Lhx8 from inducing V2a interneurons, suggesting that Isl1 binds to Lhx8 and changes the target gene specificity of Lhx8 as it does with Lhx3 [14] (Fig. 9C, D, S7C). Interestingly, however, co-expression of Isl1 and Lhx8 induced neither ectopic Hb9+ MNs nor cholinergic genes in the dorsal spinal cord (Fig. 9C, S7C). Likewise, Isl1-Lhx8 rarely triggered MN formation or cholinergic gene expression in the dorsal spinal cord (Fig. 9C, S7D), indicating that the Isl1-Lhx8-hexamer is ineffective in activating cholinergic gene expression in the spinal cord. Together, our data highlight that the proper cellular context is critical for the Isl1-Lhx3-hexamer and Isl1-Lhx8-hexamer complexes to function in target gene regulation.
Given that the Isl1-Lhx8-hexamer directly regulates the expression of cholinergic gene battery in the developing forebrain, it is possible that the Isl1-Lhx8-hexamer triggers cholinergic neuronal fate in stem cells. To test this possibility, we generated ESCs, in which the expression of Isl1-Lhx8 is induced by doxycycline (Dox), namely Isl1-Lhx8-ESCs (Fig. 10A, B). Isl1-Lhx8 forms the Isl1-Lhx8-hexamer with endogenous NLI in Dox-treated Isl1-Lhx8-ESCs (data not shown). The expression of cholinergic pathway genes, ChAT, VAChT, CHT and Acly, but not a MN gene Hb9, were readily induced by Isl1-Lhx8 under monolayer culture condition (Fig. 10C, D), suggesting that the Isl1-Lhx8-hexamer controls the expression of cholinergic pathway genes in ESCs. We also monitored the cholinergic gene expression in floating culture of embryoid bodies (EBs), which acquire the characteristics of forebrain neural precursors [34]. In the absence of Dox, many TuJ1+ neurons were observed in EBs, but VAChT+ neurons were hardly detected (Fig. 10E, F). Dox treatment markedly induced VAChT+TuJ1+ cholinergic neurons in EBs (Fig. 10E, F), suggesting that Isl1-Lhx8 triggers the cholinergic neuronal fate in stem cells. Likewise, RT-PCR also revealed that Isl1-Lhx8 significantly induced the expression of ChAT, VAChT and CHT in EB culture conditions (Fig. 10G). In the same conditions, Isl1-Lhx8 did not induce the expression of MN genes, such as Hb9, Isl2, and Chodl (Fig. 10G, data not shown). In contrast, Isl1-Lhx3 induced Hb9 as well as the cholinergic genes in both monolayer culture and floating embryoid bodies treated with retinoic acid and sonic hedgehog agonist (Fig. 10H, data not shown) [19]. Together, these results indicate that the Isl1-Lhx8-hexamer is capable of triggering the cholinergic neuronal fate, but not MN fate, in stem cells.
Establishment of correct neurotransmitter characteristics is an essential step of neuronal fate specification, but very little is known about how a battery of genes involved in a specific chemical-driven neurotransmission is coordinately regulated during vertebrate development. In this study, we report that Isl1 directly regulates a battery of genes establishing a cholinergic neurotransmitter characteristic in two developmentally unrelated cell types in vertebrate CNS (Fig. 11). Furthermore, we show that Isl1 does not do this alone, but performs its actions by forming two distinct cell type-specific transcription complexes, the Isl1-Lhx3-hexamer in the spinal cord and the Isl1-Lhx8-hexamer in the forebrain, both of which target common enhancer regions in each of the cholinergic pathway genes.
In C. elegans, a set of dopamine pathway genes, which encode dopamine synthesizing enzymes and dopamine transporters, are co-regulated through a specific cis-regulatory element that is activated by the ETS transcription factor AST-1 [35]. Likewise, cholinergic pathway genes are co-regulated by a single transcription factor UNC-3 via UNC-3-binding motif in cholinergic MNs of C.elegans [7]. Does the vertebrate CNS with much more complex circuits utilize a similar strategy in establishing a particular neurotransmitter identity in multiple types of neurons sharing a neurotransmission system? In vertebrate genome, gene regulatory motifs could occur far away from each gene transcription unit. Thus, identification of a common motif in a battery of neurotransmission-involved genes in vertebrates is much more difficult than in the nematode genome in which the regulatory sequences typically reside in proximity to the transcription start sites. While genome-wide unbiased ChIP-seq approaches could provide a solution to this challenging task, transcription factor(s) controlling a suite of neurotransmission genes need to be identified first to permit ChIP-seq analyses. Expression of Isl1 in multiple cholinergic cell types throughout CNS [8], [9], [10], [11] suggests Isl1 as a good candidate factor to control cholinergic pathway genes. Loss-of-function studies established that Isl1 is required for cholinergic fate specification in spinal MNs, a subset of FCNs, and retinal amacrine cells (this study) [13]. Our study suggests that, to trigger cholinergic neuronal fate, Isl1 functions in combination with other proteins by forming cell type-specific transcription complexes; the Isl1-Lhx3-hexamer in the spinal cord and the Isl1-Lhx8-hexamer in the forebrain. Our ChIP-seq and subsequent analyses revealed that the core set of cholinergic pathway genes shares the binding motif, which recruits the Isl1-Lhx3-hexamer and the Isl1-Lhx8-hexamer in the embryonic spinal cord and forebrain, respectively, and is activated by these complexes. In addition to ChAT, VAChT, CHT and Acly, our ChIP-seq also uncovered the hexamer-binding peaks in other cholinergic pathway genes, such as acetylcholine esterase and a cluster of nicotinic acetylcholine receptors Chrna5/a3/b4 (data not shown). An important area for future study is whether similar Isl1-containing complexes exist to control cholinergic fate decision in other areas of the CNS, such as retina and hindbrain. Isl1 is co-expressed with Phox2a, a paired-like homeodomain transcription factor, in the cranial MNs of the hindbrain [36]. Interestingly, a recent report shows that Isl1 associates with Phox2a and binds to the same cholinergic enhancer in the ChAT gene, which we identified in this study, when co-expressed with Phox2a [37], raising a possibility that Isl1 forms a complex with Phox2a in the hindbrain MNs to control cholinergic gene expression. Together, these results strongly support a model in which the cholinergic pathway genes are concomitantly activated by cell type-specific Isl1-containing complexes during cholinergic neuronal differentiation in the developing CNS (Fig. 11).
While the concept that a defined transcription factor controls the cholinergic gene battery is shared between nematodes and vertebrates, a clear difference is also noteworthy. In C. elegans MNs, a single transcription factor UNC-3 serves as a key regulator of the cholinergic pathway genes, whereas, in vertebrate CNS, Isl1-containing cell type-specific transcription complexes control the cholinergic gene battery. The combinatorial utilization of transcription factors is beneficial to generate massively divergent cell types in development. It is possible that regulation of the cholinergic genes by a single transcription factor in ancestral species has been diversified to a transcription complex in vertebrates, as the CNS circuitry becomes more complex. Another possibility is that the hexamer complexes and Ebf transcription factors, vertebrate UNC-3 orthologs, function cooperatively and/or redundantly to control cholinergic genes in the vertebrate CNS. Several findings support a possibility that the Isl1-containing hexamers act together with Ebf. Ebf proteins are expressed in differentiating MNs and ventral forebrain during embryonic development [38], [39], [40]. We found that Ebf1 associates with both types of hexamers in cells (data not shown). Finally, our de novo motif analysis of the Isl1-Lhx3-hexamer-bound ChIP-seq peaks uncovered that the Ebf-binding site is enriched in a subset of the peaks (data not shown). Thus, in the future, it will be interesting to investigate if Ebf factors collaborate with the hexamers in regulating cholinergic genes and other hexamer-targets. Our study demonstrated co-expression of Isl1 and Lhx8 in cholinergic neurons in the embryonic CPu and BMC. Interestingly, while Isl1 is required for cholinergic neuronal differentiation in the CPu, it is dispensable for differentiation of at least a subset of cholinergic neurons in the BMC. Lhx8 alone might be sufficient for the acquisition of the cholinergic phenotype in the remaining FCNs in the BMC. In addition, considering that some cholinergic neurons are still formed in Lhx8-deficient mice [24], the Lhx8-independent pathway may be present to trigger cholinergic gene expression in the basal forebrain.
Our finding that both Isl1-Lhx3-hexamer in spinal MNs and Isl1-Lhx8-hexamer in the forebrain bind to the HxRE motif in the cholinergic genes prompts the question of whether these two complexes share other target genes. Given the differences between MNs and FCNs in their functions, synaptic partners, and patterns of cell migration and axon trajectory, it is highly probable that the two complexes have largely separate sets of target genes, which establish MN- or FCN-specific characteristics, while sharing the cholinergic pathway genes as common targets. The Isl1-Lhx3-hexamer and Isl1-Lhx8-hexamer likely bind to similar but distinct sequences, and the HxREs in the cholinergic genes might have characteristics to be recognized by both Isl1-Lhx3-hexamer and Isl1-Lhx8-hexamer. In this respect, it is notable that the most optimal binding motifs for Isl1:Lhx3 or Isl1:Lhx8 identified by the SELEX methods exhibit unique features as well as shared sequences (Fig. S5) [17]. The HxRE motifs in the cholinergic genes show variations from both Isl1:Lhx3- and Isl1:Lhx8-binding sequences (Fig. S1B) [17], [20]. Isl1-Lhx8 failed to bind to the MN-specific enhancer of Hb9, which recruits the Isl1-Lhx3-hexamer [30], [41] (data not shown), further suggesting that the Isl1-Lhx8-hexamer and the Isl1-Lhx3-hexamer have unique genomic binding sites. This idea is consistent with the recent finding that the genome occupancy of Isl1 substantially changes depending on whether Isl1 is expressed alone, or co-expressed with Lhx3 or Phox2a, each of which binds Isl1 [37]. Future studies to identify the genome-wide binding sites for the Isl1-Lhx8-hexamer and to compare the target genes and motifs among Isl1-containing cell type-specific complexes will provide important insights into one of fundamental questions of developmental biology; how a single transcription factor directs fates of multiple neuronal types with a common trait. Additional mechanisms likely operate for the Isl1-Lhx3 and Isl1-Lhx8 complexes to choose distinct sets of targets, given that the ability of Isl1-Lhx3-hexamer and Isl1-Lhx8-hexamer to activate target genes is highly dependent on the cellular context. Cholinergic genes were induced only by Isl1-Lhx3 in the spinal cord and only by Isl1-Lhx8 in the forebrain. Moreover, Isl1-Lhx3 readily activated MN genes in the developing spinal cord, but not in the forebrain. First, collaborating transcription factors or cofactors could contribute to the cell context-specific activation of the target genes for each hexamer complex (Fig. 11). The Isl1-Lhx3-hexamer has been shown to cooperate with Neurog2 (Ngn2) and Stat3 in MN gene regulation [20], [30]. It will be interesting to test whether the Isl1-Lhx8-hexamer interacts with other transcription factors, such as Mash1, Olig2, Dbx1/2 or Gbx1/2, to control FCN differentiation in the ventral forebrain. Second, the in vivo chromatin context may play a role in cell type-specific gene expression. For instance, MN genes, such as Hb9 and Isl2, may possess transcription-permissive chromatin environment in the spinal cord and transcriptionally inactive chromatin in the forebrain, thus allowing the gene activation by the Isl1-Lhx3-hexamer only in the developing spinal cord, but not in the forebrain. In this regard, it is noteworthy that the activation of Acly-HxRE:GFP reporter gene, which is free from chromain-mediated regulation, is cell context-independent. Both Isl1-Lhx3 and Isl1-Lhx8 was capable of activating the Acly-HxRE:GFP reporter in both the developing spinal cord and forebrain (Fig. 3, S6). Together, our study provides key insights into the gene regulatory logic of cholinergic neuronal differentiation, which would be useful to generate cholinergic neurons for therapeutic or drug screening purposes.
All animal procedures were conducted in accordance with the Guidelines for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee (IACUC) at OHSU.
Rat Isl1, Isl1-N230S, and mouse Lhx3, Lhx3-N211S, Lhx8, Lhx1, Isl1-T2A-Lhx3, Isl1-Lhx3 fusion, Isl1-Lhx8 fusion, NLI, and LacZ genes were cloned into pCS2, pcDNA3 (Invitrogen) containing a HA, Flag or myc-epitope tag, or pCIG for expression in mammalian cells and chick embryos and for in vitro transcription and translation reactions. All of these vectors except Lhx8 were previously described [14], [30], [41]. NLIDD-Isl1ΔLIM is a fusion of 1-298aa of NLI containing the self-dimerization domain of NLI and 111-349aa of Isl1, which is a C-terminal region of Isl1 that does not include the LIM domains. Isl1-N230S and Lhx3-N211S are missense mutatnts, which are deficient in their ability to directly bind DNA [14]. Isl1, Lhx3, Lhx8, Isl1-Lhx3 and Isl1-Lhx8 were also cloned into the pCIG-2 vector for electroporation of mouse brains. Isl1 and NLI were cloned into the bacterial expression vector pGEX4T-1 (Amersham) for in vitro GST-pull down experiments. Lhx8 and NLI were cloned into the mammalian GST expression vector pEBG for GST-pull down experiments in cell lines. Isl1-Lhx8 was generated by fusing Isl1 full-length and Lhx8 full-length via flexible linker GGSGGSGGSGG. Isl1-T2A-Lhx8 was generated by inserting T2A sequences between full-length Isl1- and full-length Lhx8-coding sequences.
The location of Isl1-Lhx3-bound ChIP-seq peaks for cholinergic genes in mouse genome (mm9) is the following; ChAT/VAChT, chr14:33256618–33257117; CHT, chr17:54298028–54298480; Acly, chr11:100381966–100382465, chr11:100379377–100379876, and chr11:100395147–100395645. Mouse genomic regions covering the Acly-enhancer, ChAT-enhancer and CHT-enhancer were amplified using PCR, and two or three copies of these enhancers were cloned into TK-LUC or synthetic TATA-GFP reporter vectors. Primers to amplify these genomic enhancers are Acly-enahncer1, forward 5′- GA AGA TCT TGA TAG CAC ACT ACT TTG CTC TGG, reverse 5′- CG GGA TCC CAG TGA CGC ACG GCG AGC GGG AAG; ChAT-enhancer, forward 5′- GA AGA TCT TAC TAA TTG GAT TAA TTG ATT TGC, reverse 5′- CG GGA TCC GGG AAT TAA TAA CTT AGA ATT TGA; CHT-enhancer, forward 5′- GA AGA TCT TGA GCA GCC TAT GCC ACA AGG ACA, reverse 5′-CG GGA TCC AGG AAT CCA TCA CAA AGC TAA GAC. AAGCTGATTA sequences in Acly-enh1 were mutated to CCGCGCGGCC to generate the Acly-enh1-HxRE-mt reporter. Acly-HxRE:LUC, Acly-HxRE:GFP, ChAT-HxRE:LUC and ChAT-HxRE:GFP reporters were created by cloning multiple copies of the following duplex oligonucleotides into synthetic TATA-GFP or TK-LUC vectors. Acly-HxRE, 5′- CAG AGC TAAT CAG CTTG AGTG GGT-3′; ChAT-HxRE 5′- TGG TAC TAAT TGG ATTA ATTG ATT-3′.
The generation of Isl1f/f, Nestin-Cre, and Nkx2.1-Cre mice has been described previously [28], [29], [31]. Isl1f/f mice were crossed with Isl1f/+;NesticCre mice or Isl1f/+;Nkx2.1Cre mice to generate Isl1f/f;NesticCre or Isl1f/f;Nkx2.1Cre embryos, respectively, for analyses. Mouse embryos were collected at the indicated developmental stages, and fixed in 4% paraformaldehyde, embedded in OCT and cryosectioned in 12 µm thickness for immunohistochemistry assays or 18 µm thickness for in situ hybridization with digoxigenin-labeled probes.
These assays were performed as described [14], [42]. In chick electroporation assays, DNAs were injected into a ∼ Hamburger and Hamilton (HH) stage 13 chick neural tube. The embryos were harvested 3 days post-electroporation and fixed in 4% paraformaldehyde, embedded in OCT and cryosectioned in 12 µm thickness for immunohistochemistry assays or 18 µm thickness for in situ hybridization with digoxigenin-labeled probes. Each set of chick electroporation experiments was repeated independently three to six times with at least three embryos injected with the same combination of plasmids for each experimental set. Representative sets of images from reproducible results were presented.
For immunohistochemistry assays, the following antibodies were used; rabbit anti-Hb9 [43], mouse anti-Mnr2/Hb9 (5C10, DSHB), rabbit anti-Isl1/2 [9], guinea pig anti-Chx10 [43], rabbit anti-Lhx3 [15], guinea pig anti-VAChT (AB1588, Millipore), goat anti-ChAT (AB144P, Millipore), rabbit anti-GFP (A6455, Molecular Probes), rabbit α-Nkx2.1 (Santa Cruz), guinea pig α-Lhx8 (generated using mouse Lhx8 211–367aa region as antigen), rabbit anti-NLI [44], TuJ1 (Covance) and mouse anti-Flag (Sigma).
For in situ hybridization analyses, cDNA for mouse ChAT, Acly and CHT and chick CHT, Acly, VAChT and ChAT were cloned to pBluescript vector and these vectors were used to generate digoxigenin-labeled riboprobes.
HEK293T cells were seeded onto 10 cm tissue cultures dishes, cultured in DMEM media supplemented with 10% fetal bovine serum, and transfected using Superfect (Qiagen). 48 hours after transfection, cells were harvested and lysed in IP buffer (20 mM Tris-HCl, pH 8.0, 0.5% NP-40, 1 mM EDTA, 150 mM NaCl, 2 mM PMSF, 10% Glycerol, 4 mM Na3VO4, 200 mM NaF, 20 mM Na-pyroPO4, and protease inhibitor cocktail). In these studies, precipitations were performed with either α-Flag antibody (Sigma) or glutathione sepharose beads (GE-Healthcare). The interactions were monitored by western blotting assays using α-Flag (Sigma) and α-HA (Babco) antibodies. Following western blotting with fluorescence-labeled secondary antibodies, the bound fractions of proteins were scanned by the Odyssey imaging system (Li-Cor) following western blotting with fluorescence-labeled secondary antibodies.
In vitro GST-pull down assays were performed as described [45]. BL21 E. coli were transformed with pGEX vector alone, pGEX-Isl1, or pGEX-NLI to express the GST-fusion proteins and lysed by sonication. The GST-fusion proteins were purified by incubating the lysates with glutathione sepharose beads (GE-Healthcare). The beads were then washed and incubated with the putative interacting partners Lhx8, Lhx3 and Lhx1, which were generated in vitro by using the TnT T7 Quick Coupled transcription/translation system (Promega). Bound proteins were eluted by boiling, and were monitored by western blotting assays using α-HA (Babco) antibodies and Odyssey imaging system (Li-Cor).
SELEX was performed as described [46] with proteins in vitro transcribed and translated from the following vectors; Flag- tagged Isl1-Lhx8 fusion, Flag-Isl1, Flag-Lhx8, and Isl1-T2A-Lhx8 which produce both Flag-Isl1 and HA-Lhx8 proteins. The proteins, which were generated by using the TnT T7 Quick Coupled transcription/translation system (Promega), were incubated with a pool of double-stranded oligonucleotides containing a central core region of 22 random nucleotides with identical 5′- and 3′-flanking regions. For each SELEX reaction, ∼30 clones were randomly selected and sequenced. The motif analysis was conducted using Multiple Em for Motif Elicitation (MEME) [47].
P19 embryonic carcinoma cells were cultured in α-minimal essential media supplemented with 2.5% fetal bovine serum (FBS) and 7.5% bovine calf serum. For luciferase assays, P19 cells were seeded and incubated for 24 hours, and transient transfections were performed using Lipofectamine 2000 (Invitrogen). An actin promoter-β-galactosidase plasmid was cotransfected for normalization of transfection efficiency, and empty vectors were used to equalize the total amount of transfected DNA. Cells were harvested 36–40 hours after transfection. Cell extracts were assayed for luciferase activity and the values were normalized with β-galactosidase activity. Data are presented as means of triplicate values obtained from representative experiments. All transfections were repeated independently at least four times. Luciferase reporter data are shown in relative activation fold (mean +/− standard deviation).
The overall procedures for ventral forebrain electroporation and organotypic slice culture were previously described [48]. E15.5 mouse embryos were harvested and brains were dissected and embedded in 3% low melting point agarose dissolved in complete Hanks Balanced Salt solution (cHBSS). 250 µm thick slices of the brains were generated using a Leica VT1200 vibratome. Slices containing the appropriate regions of the ventral forebrain were focally injected with combinations of plasmids. The slices were then mounted on the anode above a 1 mm agarose slice and cHBSS was used to gap the cathode, and electroporated using ECM 830 electroporator (BTX) under the following condition; 60 mV, 5 ms interval pulse, 500 ms delay, and 5 pulses. Immediately after the electroporation, the slices were transferred to transwell inserts (0.4 µm pore size) and cultured for three to five days in vitro with slice media containing 5% heat inactivated horse serum added below the insert at 37°C with 5% CO2. Slices were fixed in 4% paraformaldehyde, washed in PBS and analyzed post-fix using immunofluorescence histochemistry.
The overall procedures for ex vivo brain electroporation and organotypic slice culture were previously described [49]. E15.5 mouse embryos were harvested and then the heads were removed and placed in cHBSS. Each combination of DNA constructs mixed with 0.5% Fast Green (Sigma) were injected into the lateral ventricles of isolated E15.5 mouse heads using a Picospritzer III microinjector. The electroporation was carried out on whole heads using ECM 830 electroporator (BTX) under the following condition; 30 mV, 100 ms intervals, 4 pulses, and 100 ms delay. For organotypic slice culture, brains were dissected immediately following electroporation, and embedded in 3% low melting point agarose dissolved in cHBSS. 250 µm thick slices of the brains were generated using a Leica VT1200 vibratome and transferred to transwell inserts (0.4 µm pore size). The slices were then cultured for three to five days in vitro with slice media containing 5% heat inactivated horse serum added below the insert at 37°C with 5% CO2. Slices were fixed in 4% paraformaldehyde and analyzed for GFP expression.
Each set of mouse brain electroporation experiments was repeated independently three to six times. For each set of mouse brain electroporation, three to four brain slices were electroporated per condition. Reproducible results were presented in the figures. Confocal images were acquired using a Nikon Eclipse Ti inverted microscope with perfect focus and a motorized stage coupled to a 4 laser line A1 scanning confocal system. Representative sets of images were presented.
For in utero electroporation, timed-pregnant C57BL/6N females were anesthetized at stage E13.5 with isoflurane (4% during induction, 2.5% during surgery), and the uterine horns were exposed by way of laparotomy. 1 µℓ of the expression vector in PBS containing 0.05% fast green (Sigma-Aldrich, St Louis, MO, USA) was injected into the lateral ventricle of the embryo using a 100 mm glass capillary (1B100-4, World Precision Instruments, Inc., USA). Electroporation was performed using Tweezertrodes (diameter, 5 mm; BTX, Holliston, MA, USA) with 5 pulses of 45 V for 50 millisecond duration and 950 millisecond intervals using a square-wave pulse generator (ECM 830; BTX). The uterine horns were then returned to the abdominal cavity, the wall and skin were sutured, and embryos were allowed to continue their normal development and collected for the further analyses at indicated stages.
Total RNAs were extracted using the Trizol (Invitrogen) and reverse-transcribed using the SuperScript III First-Strand Synthesis System (Invitrogen). For quantitative PCR of ChAT, VAChT, Acly and CHT, the following probes and primers predesigned by the TaqMan Gene Expression Assay (Applied Biosystems) for each gene were used with TaqMan Universial Master MixII and 7500 ABI qPCR machine (Applied Biosystems); ChAT (Assay ID-Mm01221882_m1), VAChT (Assay ID-Mm00491465_s1), Acly (Assay ID-Mm01302282_m1), CHT (Assay ID- Mm00452075_m1) and Eukaryotic 18S rRNA Endogenous Control (FAM Dye/MGB Probe, Non-Primer Limited). In addition, the following primers were used with the SYBR green kit (11762-500, Invitrogen) and Mx3000P (Stratagene). Hb9, 5′-GTT GGA GCT GGA ACA CCA GT, 5′-CTT TTT GCT GCG TTT CCA TT; ACLY, 5′-GAA GCT GAC CTT GCT GAA CC, 5′-CTG CCT CCA ATG ATG AGG AT; ChAT, 5′-CCT GCC AGT CAA CTC TAG CC, 5′-GGA AGC CTT TAT GAT GAG AA; CHT, 5′-GTG GTC TAG CTT GGG CTC AG, 5′-GGC AAT GAG TGC AGA GAC AA; VAChT, 5′-TTG ATC GCA TGA GCT ACG AC, 5′-CCA CTA GGC TTC CAA AGC TG; Hb9, 5′-GTT GGA GCT GGA ACA CCA GT, 5′-CTT TTT GCT GCG TTT CCA TT; Isl2, 5′-GCA AAC TCG CTG AGT GCT TTC, 5′-ACC ATA CTG TTG GGG GTG TC; Chodl, 5′-CAG TGG AAT GAC GAC AGG TG, 5′-GGT TCC CAA AGC AAC CAG TA; Isl1, 5′-GAC ATG ATG GTG GTT TAC AGG C, 5′- GCT GTT GGG TGT ATC TGG GAG; Lhx3, 5′-AGA GCG CCT ACA ACA CTT CG, 5′-GGC CAG CGT CTT TCT TCA GT; Lhx8, 5′-CAG TTC GCT CAG GAC AAC AA, 5′-AGC CAT TTC TTC CAA CAT GG; GAPDH, 5′-ACC ACA GTC CAT GCC ATC AC, 5′-TCC ACC ACC CTG TTG CTG TA; and Cyclophilin A, 5′-GTC TCC TTC GAG CTG TTT GC, 5′-GAT GCC AGG ACC TGT ATG CT. RT-PCR experiments were performed with three or four independent sets of samples. Data are represented as the mean of duplicate or triplicate values obtained from representative experiments. Error bars represent standard deviation.
The ChIP-seq data used for the analysis in this paper has been deposited in the Gene Expression Omnibus (GEO) database (assession no. GSE50993) [20].
To perform the ChIP assays with mouse embryonic tissues, we dissected E12.5 spinal cords or E15.5 forebrains. The microdissected spinal cords from five E12.5 embryos or the forebrains of three E15.5 embryos were combined together for each ChIP reaction with a specific antibody. The tissues were dissociated completely, fixed by 1% formaldehyde for 10 min at room temperature, and quenched by 125 mM glycine. Next, cells were washed with Buffer I (0.25% Triton X-100, 10 mM EDTA, 0.5 mM EGTA, 10 mM Hepes, pH 6.5) and Buffer II (200 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 10 mM Hepes, pH 6.5) sequentially. Then, cells were lysed with lysis buffer (0.5% SDS, 5 mM EDTA, 50 mM Tris-HCl, pH 8.0, Protease inhibitor cocktail) and were subjected to sonication for DNA shearing. Next, cell lysates were diluted 1∶10 in ChIP buffer (0.5% Triton X-100, 2 mM EDTA, 100 mM NaCl, 50 mM Tris-HCl, pH 8.0, Protease inhibitor cocktail) and, for immunoclearing, were incubated with IgG and protein A agarose beads for one hour at 4°C. Supernatant was collected after quick spin and incubated with appropriate antibodies and protein A agarose beads to precipitate the hexamer/chromatin complex for overnight at 4°C. After pull-down of the hexamer/chromatin complex/antibody complex with protein A agarose beads, the beads were washed with TSE I (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl, pH 8.0, 150 mM NaCl), TSE II (same components as in TSE I except 500 mM NaCl) and Buffer III (0.25M LiCl, 1% NP-40, 1% deoxycholate, 1 mM EDTA, 10 mM Tris-HCl, pH 8.0) sequentially for 10 minutes at each step. Then the beads were washed with TE buffer three times. The hexamer/chromatin complexes were eluted in elution buffer (1% SDS, 1 mM EDTA, 0.1M NaHCO3, 50 mM Tris-HCl, pH 8.0) and decross-linked by incubating at 65°C overnight. Eluate was incubated at 50°C for more than two hours with Proteinase K. Next, DNA was purified with Phenol/chloroform and DNA pellet was precipitated by ethanol and resolved in water. The purified final DNA samples were subjected to quantitative PCR reactions using the SYBR green kit (11762-500, Invitrogen) and Mx3000P (Stratagene). The total input was used for normalization. All ChIP experiments were repeated independently at least three times. Data are represented as the mean of duplicate or triplicate values obtained from representative experiments, and error bars represent standard deviation.
The following primers were used for ChIP-PCR.
ChAT-enhancer
forward 5′-TAC TAA TTG GAT TAA TTG ATT TGC
reverse 5′-GGG AAT TAA TAA CTT AGA ATT TGA
ChAT-negative
forward 5′- CTG TGG CTC ATA ACG CTC ATT TTG
reverse 5′- AGT TTG TGG TGG GCC GAG ATG GCA
Acly-enh1
forward 5′- TGA TAG CAC ACT ACT TTG CTC TGG
reverse 5′-CAG TGA CGC ACG GCG AGC GGG AAG
CHT-enhancer
forward 5′-TGA GCA GCC TAT GCC ACA AGG ACA
reverse 5′- CAT TAG GAG AGC TTG TTC CAG TGA
The following antibodies were used for ChIP-PCR; mouse/rabbit IgG (Santa Cruz), rabbit anti-Isl1 [9], rabbit anti-Lhx3 [15], rabbit anti-NLI [44], and goat anti-Lhx8 (sc-22216, Santa Cruz).
The generation of Isl1-Lhx3-ESCs was described previously [19]. To generate Isl1-Lhx8-ESCs, the A172LoxP ES cell line [50] was maintained in an undifferentiated state on 0.1% gelatin-coated dishes in the ESC growth medium that consisted of Knockout DMEM, 10% FBS, 0.1 mM non-essential amino acids, 2 mM L-glutamine, 0.1 mM β-mercaptoethanol and recombinant leukemia inhibitory factor (LIF, 1000 units/ml, Chemicon). Flag-tagged Isl1-Lhx8 fusion was inserted into Tet-inducible plasmid p2Lox. The Isl1-Lhx8 vector was co-transfected with pSALK-Cre into the A172LoxP ES cell line using Lipofectamine 2000 (Invitrogen). Stable transfectants were isolated by selection with neomycin (G418, 400 µg/ml) for seven days. Dox-dependent induction of Flag-Isl1-Lhx8 expression was monitored by western blotting and immunohistochemical analyses using α-Isl1, α-Lhx8 and α-Flag antibodies.
To induce cell differentiation, Embryoid bodies (EBs) were formed and cultured for 2 days using the hanging drop method (1×103 ESCs per 20 µℓ drop). Hanging drops were transferred to suspension culture in 6 well low attachment dishes and cultured. EBs were cultured without or with doxycycline (2 µg/ml) for 2–5 days in the ESC medium without LIF or in the differentiation medium that contains KnockOut serum replacement (Life technologies). Then, EBs were collected for either RT-PCR or immunohistochemical analyses.
|
10.1371/journal.pcbi.1000238 | Formation and Growth of Oligomers: A Monte Carlo Study of an Amyloid Tau Fragment | Small oligomers formed early in the process of amyloid fibril formation may be the major toxic species in Alzheimer's disease. We investigate the early stages of amyloid aggregation for the tau fragment AcPHF6 (Ac-VQIVYK-NH2) using an implicit solvent all-atom model and extensive Monte Carlo simulations of 12, 24, and 36 chains. A variety of small metastable aggregates form and dissolve until an aggregate of a critical size and conformation arises. However, the stable oligomers, which are β-sheet-rich and feature many hydrophobic contacts, are not always growth-ready. The simulations indicate instead that these supercritical oligomers spend a lengthy period in equilibrium in which considerable reorganization takes place accompanied by exchange of chains with the solution. Growth competence of the stable oligomers correlates with the alignment of the strands in the β-sheets. The larger aggregates seen in our simulations are all composed of two twisted β-sheets, packed against each other with hydrophobic side chains at the sheet–sheet interface. These β-sandwiches show similarities with the proposed steric zipper structure for PHF6 fibrils but have a mixed parallel/antiparallel β-strand organization as opposed to the parallel organization found in experiments on fibrils. Interestingly, we find that the fraction of parallel β-sheet structure increases with aggregate size. We speculate that the reorganization of the β-sheets into parallel ones is an important rate-limiting step in the formation of PHF6 fibrils.
| It is believed that the self association of certain protein molecules into aggregated structures, known as amyloid fibrils, plays an important role in a variety of human diseases, such as Alzheimer's disease and Parkinson's disease. Although the ability to form such amyloid fibrils is a common property for proteins, the process leading to these fibrils is incompletely understood. The early stages of the process involve small transient heterogeneous structures made of a few protein chains and are especially difficult to characterize. Here we use atomic-level simulations to explore the early part of the aggregation process for a fibril-forming fragment of the protein tau associated with Alzheimer's disease. We find that a multitude of small aggregates, rich in sheetlike structures, form through a nucleation process. Interestingly, a statistically preferred type of aggregate, consisting of two tightly packed sheets, emerges with increasing aggregate size. Growth of these larger aggregates seems to be a slow process that correlates with the emergence of more uniformly ordered sheets. We speculate that reorganization of the protein chains leading to that ordered arrangement is an important bottleneck to amyloid fibril formation for this peptide.
| A century ago, Alois Alzheimer reported dense extracellular deposits and intracellular neuronal aggregates in the brain of a patient suffering from memory loss, focal symptoms, delusions, and hallucinations [1]. The extracellular deposits have been subsequently identified as amyloid plaques composed of an accumulation of β-amyloid peptides, while the intracellular neuronal aggregates are neurofibrillary tangles (NFTs) formed by the microtubule-associated protein tau. Tau filaments adopt multiple morphologies, among which paired helical filaments (PHFs) are the principal constituent of NFTs in the Alzheimer's disease (AD) brain, while straight filaments are a minor variant [2]. In electron micrographs, the PHF appears as a twisted double-helical ribbon of subunits that alternate in width between 10–20 nm and has a half-period of 80 nm [3]. The β-amyloid filaments were known a long time ago to exhibit the characteristic “cross-β” structure, a β-sheet rich structure in which the β-strands are aligned perpendicular to the fibril direction and the interstrand hydrogen bonds are parallel to the fibril axis [4]. However, it is only recently that the “cross-β” characteristics of tau filaments from AD brain and from full-length recombinant protein have been conclusively demonstrated [5]–[7].
Protein tau is primarily expressed in neurons, and is involved in microtubule assembly and stabilization [8],[9]. It is highly soluble and flexible in aqueous solution [10], belonging to the “intrinsically disordered” proteins. Even when it is bound to the surface of microtubules, tau retains most of its disordered character [11]. In adult human brains, there are six isoforms of tau. Depending on the isoform, three or four repeats constitute the core of the microtubule-binding domain. Coincidently, the second and third repeats in the microtubule-binding domain are also the core of PHFs with the cross-β structure, while the rest of the protein forms the fuzzy coat of PHFs [5].
It has been suggested that the motifs VQIINK (PHF6*) in the second repeat and VQIVYK (PHF6) in the third repeat of tau play a key role in the formation of PHF [12],[13]. By transmission electron microscopy (TEM), it was found that AcPHF6 (Ac-PHF6-NH2) peptides aggregate into straight filaments [14]. Further, X-ray diffraction patterns and electron micrographs were reported for assemblies of some PHF/tau-related peptides, including AcPHF6 and a longer peptide containing both PHF6* and PHF6 [15]. Assemblies of the latter peptide were found to have a twisted fibrillar structure, whereas the data for AcPHF6 were found to be consistent with a tubular assembly with double walls [15]. An X-ray study of PHF6 microcrystals, on the other hand, found a cross-β spine consisting of β-sheet pairs with a “dry steric zipper” organization at the sheet-sheet interface [16].
Recent findings using transgenic mice models have suggested that soluble aggregated tau rather than NFTs might induce neurodegeneration [17]–[19]. The demonstration of toxicity of soluble aggregates has brought up the possibility of using the oligomeric forms as drug targets. Therefore, it is of great importance to understand the initial nucleation and growth process of tau aggregation. While X-ray diffraction, electron micrography, and microcrystallography have provided information on the structural organization of tau filaments, the initial oligomerization process of full-length protein tau or its peptide fragments remains far from being well understood.
Computational studies have complemented the experiments to provide insights into amyloid formation. Although a wide range of models [20]–[34] has been employed to simulate amyloid aggregation (for a recent review, see [35]), due to the limitations of currently available computer power, most computational studies were limited to small oligomers, short time-scales, or restrained simulations. An alternative approach, to test the stability of preformed structures, has also been explored [36]–[39].
In this work, we study the aggregation of AcPHF6 by an all-atom protein model with a simplified interaction potential using Monte Carlo (MC) simulations. The runs, with up to 36 chains, capture many well known properties of oligomerization. They lead to a multitude of small oligomers rich in β-strand content. We also observe two distinct processes during oligomerization: formation of stable oligomers and emergence of growth capable stable oligomers. Surprisingly, we find that stability of oligomers is not synonymous with their ability to grow. For system sizes permitting the formation of more than one stable oligomer, we find that this type of conformation is more probable than having one large aggregate. Stable oligomers undergo considerable structural reorganization through reptation motion and exchanges of chains with the environment. Growth of stable oligomers is facilitated by a particular kind of ordered structure. New chains do not necessarily attach to a growing oligomer in an ordered manner, so that at every size of the oligomer, there is a slight “barrier” corresponding to a required structural reorganization, before an incremental growth occurs.
Both experimental [40] and computational [21] works suggest that amyloid formation proceeds via a nucleation process. According to the nucleated conformational conversion model [40], this process shows a two-step behavior: an initial chance association of a sufficient number of monomers to form stable but disordered oligomers, followed by the emergence, through a reorganization process, of ordered oligomers and fibrils. In this article, we begin by studying the first step, i.e., the formation of stable oligomers.
We started with twelve chains of AcPHF6 randomly positioned in the cell, at 308 K. To investigate the concentration dependence, we performed simulations in a number of cubic cells with side lengths of 65 Å, 70 Å, 75 Å, and 80 Å (see Table 1). These concentrations range from 58 mg/ml (73 mM) to 31 mg/ml (40 mM), which are typical values in simulations [30] but higher than the experimental concentrations (0.1–1.0 mg/ml [14]).
To identify aggregates in the simulations, we have used a criterion based on contacts between residues belonging to different chains. Two residues were defined to be in contact if the distance between any pair of heavy atoms of these two residues was less than 4.5 Å. Two chains were considered to have a link if they had at least four inter-chain contacts. A set of chains was considered to form a single aggregate, if the graph with those chains as nodes and inter-chain links as edges, was connected. In Figure 1a, we show how the size (number of chains) of the biggest aggregate evolved with MC time in representative runs at the highest (side length 65 Å) and lowest (side length 85 Å) concentrations, respectively (Figure S1 shows the same for six runs with side length 70 Å). In the run at high concentration, aggregation is fast. In contrast, the run at low concentration exhibits a long apparent waiting phase before a large aggregate appears for the first time. In this phase, many meta-stable aggregates with 2–8 chains form and dissolve, without growing into mature stable aggregates. At step 67 (equivalent to 67×5×107 MC steps), a stable aggregate forms for the first time (see below for its conformation), which does not dissolve into smaller aggregates, and the system enters a new, aggregated phase. This behavior is suggestive of a nucleation process, with the nucleation event occurring at step 67. It is worth stressing that the event observed here is nucleation of oligomer formation which is not the same as nucleation of fibril formation. The formation of a critical nucleus for fibrillization generally involves a reorganization process, which might be the rate-limiting step.
Figure 1b shows the evolution of the hydrophobicity energy and the hydrogen bond energy along this (low concentration) trajectory. Both these energies (anti-) correlate with the size of the largest aggregate. The hydrophobic interaction seems to be the most important driving force in the aggregation process, whereas hydrogen bonding also plays a significant role in defining the geometry of the aggregated structures. The aggregation process optimizes both these interactions. Also shown in Figure 1b is the β-strand content, which is strongly correlated with the hydrogen bond energy.
Figure 2 depicts four examples of meta-stable states from the pre-nucleation phase in the run at low concentration in Figure 1. The first example (Figure 2a) is a six-stranded, mixed parallel/antiparallel β-sheet with a clear twist, in contact with a random coil. Other β-sheets containing 2–6 strands were also observed. The second example (Figure 2b) is a relatively irregular aggregate composed of two small β-sheets with two and three strands, respectively, which are packed against each other. Completely irregular aggregates, without any β-sheet structure, were rare. The third example (Figure 2c) is a small β-sandwich consisting of one two-stranded and one four-stranded β-sheet. Finally, the fourth example (Figure 2d) is a four-stranded β-sheet with four random coils attached to it. All these four aggregates dissolved later and the system remained in the pre-nucleation stage.
According to classical nucleation theory [41], the critical nucleus is in “unstable equilibrium”. The aggregates containing fewer chains than the critical nucleus dissolve spontaneously, while those larger than the critical nucleus grow spontaneously. The system must overcome a free energy barrier and form a critical nucleus before stable aggregates form. This free energy barrier is low when the concentration is high. Indeed, in our simulations, the length of the pre-nucleation phase showed a strong concentration dependence. This is illustrated by Figure 1a, in which nucleation takes place after about 8 steps for side length 65 Å and after about 67 steps for side length 80 Å.
We now illustrate the behavior of the 12-chain system around and after the nucleation step, using the same run as in Figure 2 (side length 80 Å). Figure 3 shows six snapshots from the later part of this run. Once the spontaneous fluctuations result in the formation of a critical nucleus, as at step 67 in this run, the system has reached a point where a stable aggregate may form. The aggregate no longer disperses into smaller pieces or completely dissolves.
At the nucleation event at step 67 (Figure 3a), the aggregate consists of two random coils attached to a twisted six-stranded β-sheet. One step later (Figure 3b), a new chain has joined this aggregate by forming a two-stranded β-sheet with one of the two random coils, and the β-sheet has grown to become seven-stranded. Subsequently, the aggregate undergoes reorganization. At step 71 (Figure 3c), there is one large β-sheet composed of eight chains in contact with a small two-stranded β-sheet. The large sheet is concave, which maximizes contacts with the small sheet. The side chains of V1, I3, and Y5 are buried at the sheet-sheet interface. From this point on, there is a dynamic equilibrium between the aggregate and individual monomers in “solution”. In other words, an individual chain associates with the aggregate or dissociates from the aggregate from time to time, but the total number of peptides within the aggregate does not change significantly with time. Most of the time, the aggregate contains 10–11 chains (Figure 1), but not necessarily the same 10–11 chains at different moments. Individual chains attach and detach at the edges of the β-sheets. Occasionally, the size of the aggregate decreases to nine or increases to twelve chains (Figure 1).
During the dynamic equilibrium, conformational reorganization occurs within the aggregate. Typically, the aggregate consists of two twisted β-sheets wrapped around each other. But the number of chains in each sheet is not constant. We even observed, at step 86 (Figure 3d), a single chain in 310-helix conformation in contact with the concave face of a nine-stranded β-sheet. The most common type of aggregate seen in this run is composed of one five-stranded and one six-stranded β-sheet, as at step 91 (Figure 3e). In sandwich structures like this, large changes in the relative orientation of the two β-sheets were observed. For example, in this run, the angle between the two β-sheets changes from 10° to 60° between step 91 (Figure 3e) and step 96 (Figure 3f). Smaller β-sheets adjust their relative orientation more easily than larger β-sheets. The local bending and the alignment pattern also vary with time. We observed that the alignment of the edge strands of a β-sheet can change without their detachment from the β-sheet. However, the alignment pattern in the central part of a β-sheet were not seen to change, which is understandable as there are constraints from neighboring strands.
Structures resembling those in Figure 3 have been observed in previous simulations of smaller systems, including 6-chain simulations with explicit water for other peptides of the same length as PHF6 [28]. The curved β-sheets seen in Figure 3b–d are reminiscent of the open β-barrels reported by Derreumaux and coworkers [42],[43].
The structures shown in Figures 2 and 3, all from a single run, illustrate some general features seen in all our simulations. For example, a vast majority of our observed aggregates are β-sheet rich, and both curved sheets and sandwich-like structures are frequently occurring motifs. However, the details of the structures in Figures 2 and 3, like the exact alignment of the β-strands, are not statistically representative.
To statistically characterize aggregated structures, we carried out an additional set of fifty 12-chain runs, starting from different random initial configurations. For computational convenience, the side length was here set to 70 Å instead of 80 Å. In these runs, the same two phases were seen as in the 80 Å run described above, but the aggregation process was faster. In seven of the fifty runs, the aggregate converted to a stable β-barrel containing 8–12 chains. This type of conformation was not further investigated, because the main focus of the present study is aggregate growth, and a β-barrel is unlikely to grow into a larger aggregate.
For sandwich structures, we made a size analysis based on these fifty runs. Here we counted the total number of chains in the aggregate and the difference in number of chains between the two β-sheets after a stable two-sheet aggregate had formed. Figure 4a shows the observed distribution of aggregate size. The peak is at 11, whereas aggregates with ≤7 chains are quite rare. Note that the distribution depends on the concentration. Figure 4b illustrates the difference in size between the two β-sheets in two-sheet aggregates. Small size differences of 0 or 1 are most common. We note that whenever any collection of objects is randomly divided into two groups, there are more ways of constructing the groups with nearly equal size than of constructing them with a large size difference. Therefore large size differences would be expected to be suppressed entirely due to combinatorial considerations, irrespective of the specific properties of the system. In Figure 4b, the largest size differences involve one very small aggregate, and the observed probabilities are nearly consistent with the random estimate. But the probability of the smallest size differences is enhanced at the expense of the medium size differences of 2 and 3 in Figure 4b, suggesting that size symmetry of sheets in an oligomer is further favored due to interactions.
Inspection of the aggregates in Figure 3 shows that the β-strands tend to be arranged so that there are many hydrophobic intra-sheet contacts between them, involving the V1, I3 and Y5 side chains. A vast majority of the larger aggregates seen in our simulations share this property. In order to maximize the hydrophobic interaction between two adjacent strands in a sheet, the V1, I3, and Y5 side chains of both strands must point to the same side of the sheet (and Q2, V4, and K6 to the other). This is achieved if the strands are parallel and either in-register or off-register by two residues, or if they are antiparallel and off-register by one or three residues. Obviously, this property depends on sequence. For example, if the peptide has an alternating hydrophobic/hydrophilic pattern but an odd number of residues, both parallel and antiparallel in-register arrangements will maximize the hydrophobic interactions. Figure 5a illustrates the strand organization and the orientation of the side chains for the largest of the two β-sheets in Figure 3e. The strand organization is such that all V1, I3, and Y5 side chains point to the same direction.
In our simulated aggregates, there were two dominating β-strand arrangements, illustrated by Figure 5a, namely parallel in-register and antiparallel out-of-register by one residue. These arrangements lead to different hydrogen bond patterns. With a parallel in-register arrangement, each pair of adjacent strands is connected by five hydrogen bonds (see Figure 5a). If two strands are antiparallel and off-register by one residue, there will be either four or six hydrogen bonds connecting them. The arrangement can be repeated so that there are six hydrogen bonds between all pairs of adjacent strands (see Figure 5a), with successive strands shifting in the same direction. Another possibility is that the third chain is in the same relation to the second as the second is to the first, which leads to a zig-zag pattern with successive strands shifting in opposite direction, as in Figure 5b. The drawback of this organization is that there are only four hydrogen bonds between the second and the third strand. We observed this pattern in our simulations, but with a relatively low frequency. The organization to the right in Figure 5a, which maximizes the number of hydrogen bonds, was more common. Unlike both these organizations, the parallel organization, to the left in Figure 5a, is in-register, which may be advantageous in large sheets. In fact, we found that the fraction of parallel over antiparallel β-sheet structure increased with aggregate size, as will be discussed below.
Having observed the formation of stable oligomers in the 12-chain runs, we increased the system size to 24 chains to study oligomer growth. For this system size, we performed a set of 72 unseeded and 35 seeded runs (see Table 1), at the same temperature as before (308 K). The side length was 95 Å, corresponding to a concentration of 36.7 mg/ml.
In our unseeded 24-chain runs, the same two phases were observed as in the 12-chain runs: an initial waiting phase with only small aggregates, followed by a phase with large aggregates in dynamic equilibrium with small aggregates and free monomers. The waiting phase was often short, due to the relatively high concentration. In the aggregated phase, there was in some runs one dominating aggregate, typically with 16–20 chains. These big aggregates were invariably composed of two stacked β-sheets, as in Figure 6a. In a majority of the runs, there were two, rather than one, major aggregates, each with typically 8–10 chains. These aggregates were similar to those observed in our 12-chain runs. Some of them were of the barrel type. In total, we saw 15 stable β-barrels in our 72 unseeded 24-chain runs. As in the 12-chain case, these aggregates, which are unlikely to grow into larger aggregates, will not be further analyzed in this work.
To get a quantitative picture of the aggregation behavior of the 24-chain system, we determined the sizes of the largest and next-largest aggregates for each conformation in our unseeded runs, which we denote by x and y, respectively. Figure 7 shows minus the logarithm of the histogram of x and y. The most populated region corresponds to two aggregates of similar size, 8–10 chains. In addition, there is a weak local minimum at large x and low y, corresponding to configurations with one dominating aggregate. However, the minimum corresponding to two distinct similar size aggregates is much deeper. If the growth of an arbitrary stable oligomer had been easy, upon the formation of the first such oligomer, the remaining free monomers would have rapidly associated with that, leading to one big aggregate. But in Figure 7 we see that the probability of two similar-size aggregates is significantly larger. This indicates that further growth is not fast compared to the formation of a new nucleus. Instead, at a size of ∼10 chains, it seems that the aggregation process reaches a stage at which conformational reorganization is required before further growth.
Our 35 seeded 24-chain simulations further elucidate this point. These runs were started from initial conformations prepared by taking aggregates from the 12-chain runs and adding random coils. Despite the presence of a template, a large aggregate (with >15 chains) appeared in only 15 of these runs. In 17 of the remaining runs, the free monomers instead assembled into a second aggregate, thus leading to a state with two distinct aggregates of similar size. In the remaining 3 runs, the newly added chains stayed in the pre-nucleation phase. Even when a stable seed is present, an independent new aggregate thus forms in about half of our runs, which suggests that the time scale for the conformational reorganization required for further growth is comparable to that of the formation of a new stable aggregate of about 10 chains.
Figure 6b gives an example of a large aggregate from a seeded simulation. After the first large aggregate had appeared in this run, the system spent the rest of the run in a phase of dynamic equilibrium, with only small fluctuations in aggregate size. Figure 6b shows a randomly chosen snapshot from this phase. This aggregate shares several common features with that from the unseeded simulation shown in Figure 6a. Both aggregates are composed of two large β-sheets that are twisted and wrap around each other. The overall twist is 8.8° and 6.3° for the two sheets in Figure 6a, and 12.0° and 13.9° for those in Figure 6b (see Methods). In both structures, parallel strand pairs are either in-register or out-of-register by two residues, while antiparallel pairs are off-register by one residue, as discussed in connection with Figure 5. In addition to being twisted, the strands are also bent, which helps to make better side-chain contacts. Twisting, bending, strand alignment and side-chain packing are all important factors influencing the final conformation.
Based on our 72 unseeded 24-chain runs, we performed a statistical analysis of some important properties of the aggregated structures. Of particular interest is the β-sheet organization, which in most of our simulated aggregates is mixed parallel/antiparallel (see Figures 2, 3, and 6) but is known to be parallel in AcPHF6 fibrils [14]. In our 72 runs, we counted parallel and antiparallel pairs of adjacent strands for all aggregates of a given size. Figure 8 shows the fractions of parallel and antiparallel β-sheet structure, as obtained this way, against aggregate size. For small sizes, there is a clear statistical preference for the antiparallel organization. However, the fraction of parallel structure increases steadily with aggregate size. In aggregates with more than 18 chains, the parallel organization is more common than the antiparallel one.
Using the same runs, we also analyzed the relative orientation of the two β-sheets in sandwich structures. For all two-sheet structures with 5–7 chains in each sheet, the angle between the two sheets was determined (see Methods). Figure 9a shows the calculated distribution of this angle. The distribution exhibits a broad peak in the region 5°–35°. Relative orientations in the range 45°–90° occur but are rare.
Finally, we also calculated the overall twist for all β-sheets in two-sheet aggregates with at least five strands in each (see Methods). The two edge strands of a sheet were not included in the analysis, because we found that those strands often were more twisted than the rest of the sheet. Figure 9b shows the observed distribution of the average twist angle. Its maximum is near 11°. The shape of the distribution is asymmetric, with a shoulder near 0°.
It is worth pointing out that the runs presented in this article are of limited length. In any given run, it is most likely that some important free-energy minima were not sampled, due to high free-energy barriers. Our findings are, however, based on a set of many independent simulations. We feel confident that major trends seen, like the increase in the fraction of parallel β-sheet structure with aggregate size, are statistically robust.
We also did seventy-two 36-chain simulations (see Table 1), seeded with minimum energy conformations from our 72 unseeded 24-chain runs, to which 12 random coils were added. In many of these runs, the formation of a new independent aggregate with ∼10 chains was observed, whereas the one or two major aggregates present initially grew by only one or a few chains. In a few runs, significant growth was observed for one of the pre-existing seeds. The behavior of the 36-chain system thus supports the picture emerging from the 24-chain runs. New independent aggregates large enough to be long-lived form relatively easily in our simulations, but that an aggregate reaches this size does not mean that further growth is fast. Most aggregates seem to require conformational reorganization before they can grow, which prevents rapid growth. A detailed analysis of our 36-chain runs is beyond the scope of the present work and will be presented in a forthcoming publication.
As of now, the nucleation event in the oligomerization process is difficult or impossible to examine using experimental approaches. Light and neutron scattering techniques are capable of revealing the shape of micelles of Aβ, but these aggregates are larger than the critical nucleus [44]. Computer simulations offer unique opportunities to observe and analyze early aggregation events at the molecular level.
In our 12-chain simulations of AcPHF6 peptides from random initial conformations, two distinct phases can be identified: an early phase with only smaller aggregates, followed by a phase characterized by the presence of a large aggregate. This behavior suggests that the formation of stable oligomers occurs through a nucleation process. In a simple nucleation process, an embryo of the new phase will grow spontaneously once it has reached a certain critical size. By contrast, when proteins/peptides aggregate, the size of the embryo is not the only relevant parameter; the ability of the critical embryo to grow depends also upon its conformation. Our run at low concentration in Figure 1 illustrates this. Here the nucleation event occurred after 67 steps. Some aggregates of comparable size appeared before step 67, but these aggregates dissolved into smaller aggregates and/or monomers. The pre-nucleation behavior of our system resembles what has been seen in several previous studies of small systems with 2–9 chains. These studies found many different meta-stable aggregates with various forms of β-sheet structure, and suggest that the barrier for converting from one of these aggregates to another is low [26], [37], [45]–[48].
The aggregate observed at the nucleation stage in the run discussed above (step 67) is composed of a twisted six-stranded β-sheet and two attached random coils (Figure 3a). A striking feature of this β-sheet is that the V1, I3, and Y5 side chains point in the same direction in all strands. The aggregates from the pre-nucleation phase (see Figure 2) do not have this property. In the proposed dry steric zipper model for PHF6 fibrils, based on X-ray microcrystallography, the β-sheet pair adopts a face-to-face stacking arrangement in which the side chains of V1, I3, and Y5 nestle between sheets [16]. The V1, I3, and Y5 side chains pointing to the same side of a β-sheet might very well be a prerequisite for its participation in a stable double-layered structure. Even a single strand with the V1, I3, and Y5 side chains pointing to the opposite side of the β-sheet could mean that the entropy loss upon formation of the aggregate cannot be compensated by an enthalpy reduction from intermolecular interactions.
The second component of this aggregate (Figure 3a) is the two random coils that are in contact with the β-sheet. One of the random coils is attached to the V1–I3–Y5 side of the sheet, solely by side-chain interactions, while the other is attached to an edge of the sheet. This kind of arrangement has been observed in previous simulations for Aβ16–22 [45], Aβ11–25 [49], and the GNNQQNY segment of yeast prion protein Sup35 [31],[50]. The attachment of random coils to a β-sheet might help to stabilize the sheet through side-chain interactions.
In the aggregated phase, after the nucleation event, we observed different kinds of β-sheet rich aggregates. A common type of conformation was the β-sandwich. In these aggregates, the two sheets tended to be of similar size (Figure 4), and V1, I3 and Y5 side chains were typically found at the sheet-sheet interface (see Figure 3c, 3e, and 3f), as in the dry steric zipper arrangement [16].
An interesting question is how ordered early oligomers are. Nguyen and Hall employed a coarse-grained protein representation and discontinuous molecular dynamics to simulate amyloid aggregation for a system of 96 polyalanine peptides (Ac-KA14K-NH2). As a first step, preceding fibrillization, the chains were found to form irregular aggregates, which then converted into small β-sheets [21]. In our simulations, very few completely disordered aggregates were found. In part, this may be due to the short length of our peptide, which leads to a high propensity for it to be in an extended state. Another factor influencing the aggregation into either ordered or amorphous species is the hydrophobicity of the sequence. Indeed, a recent study by Cheon et al. [30] found amorphous aggregates for the more hydrophobic Aβ16–22 peptide but only ordered aggregates for the less hydrophobic Aβ25–35 peptide. Our results suggest that AcPHF6 behaves like Aβ25–35 rather than like Aβ16–22 in this respect.
The large aggregates seen in our study also differ from the aggregates seen in the polyalanine simulations of Nguyen and Hall [21]. One difference is that the β-sheets lacked twist in the polyalanine study. Another, possibly related, difference is that Nguyen and Hall saw multiple-sheet stacking, whereas our aggregates were sandwich structures with only two stacked β-sheets. The relation between sheet stacking pattern and sheet twist was studied in recent simulations [37].
β-sheets in the Protein Data Bank (PDB) are generally twisted. A recent study analyzed β-sheet twist in PDB structures in terms of adjacent pairs of residues on neighboring β-strands [51]. The average twist angle was found to be 17°±7° for parallel β-sheets, 15°±10° for non-hydrogen bonded residue pairs in antiparallel β-sheets, and 8°±8° for hydrogen bonded residue pairs in antiparallel β-sheets [51]. The distribution of sheet twist angle presented in Figure 9b is in line with the above statistics. Sheet twisting has also been observed in explicit-water molecular dynamics studies of preformed β-sheets [31],[36],[37],[39]. For a pair of ten-stranded β-sheets of the peptide GNNQQNY, the average twist within each sheet was found to be about 11° after a 20 ns simulation [36] using GROMACS [52], which is in excellent agreement with our results shown in Figure 9b. That the sheet twist found in this study as well as in ours is comparable to that of native proteins is not surprising, because the aggregates were relatively small.
For native proteins, there is a tendency that β-sheets with few strands are associated with larger twist angles than those containing a large number of strands [53]. From this observation, one might expect the sheet twist to be smaller in amyloid fibrils than in native proteins. This expectation is supported by data from cryo-electron microscopy experiments on insulin fibrils (twist angle 1.5°–2.5°) [54] and TEM experiments on fibrils of rationally designed peptides (twist angle 1°–3°) [55]. Solid-state NMR data on TTR105–115 [56] and Aβ1–40 [57] fibrils were found [58] to lead to slightly larger twist angles (26°±24° for TTR105–115, 14°±37° and 17°±38° for two Aβ1–40 data sets), but with large statistical uncertainties.
One parameter in describing a two-sheet aggregate is the relative orientation of the sheets. For small aggregates, one might expect large variations in this parameter. In our simulations, we measured an angle describing the relative orientation of the sheets. The distribution of this angle was indeed found to be broad (Figure 9a). Further, we saw large rotations of sheets relative to each other during the course of our simulations. For example, a ∼50° rotation occurred between steps 91 and 96 in one of the runs (Figure 3e and 3f). This rearrangement of the aggregate is concurrent with a reduction of both the hydrophobic interaction energy and the hydrogen bond energy (Figure 1b). Large relative rotations of β-sheets have also been seen in explicit-water molecular dynamics simulations for Aβ16–22 [37], using the AMBER/parm99 force field [59]. After 20 ns, a pair of preformed β-sheets had rotated by ∼90° relative to each other, leading to a better packing and stronger hydrophobic interactions. These observations of large β-sheet rotations in simulations based on completely different models indicate that this kind of movement is common in small aggregates.
To study further growth after the formation of a stable oligomer, we performed a large set of unseeded 24-chain runs. In some of these runs, a large aggregate with ∼20 chains formed, However, in most runs, the chains formed two aggregates of similar size rather than one big one (Figure 7). Even when a stable seed was present, an independent new aggregate with ∼10 chains appeared in a about half of the runs. These results indicate that while oligomers with ∼10 chains easily form in this system, many of them are not growth-competent; further growth seems to require time-consuming conformational rearrangement.
Since virtually all oligomers seen in our simulations had a large fraction of β-sheet structure, we further conclude that for a given oligomer to be able to grow, it is not sufficient that it has a high β-sheet content; the organization of the β-strands is also important. Interestingly, we found a correlation between aggregate size and the ratio of parallel to antiparallel β-sheet structure (Figure 8). Antiparallel structure was most common for small aggregates, but the fraction of parallel structure increased steadily with aggregate size (Figure 8). The β-strand organization in AcPHF6 fibrils is known from experiments to be parallel. FTIR (Fourier Transform Infrared) spectra showed amide I bands with maxima at 1619 cm−1 and 1647 cm−1, characteristic of a parallel β-strand arrangement, while no high frequency component corresponding to an antiparallel arrangement was found [14].
Could the aggregates we observe be kinetic traps en route to fibrils, or are they more likely to be off-pathway states? This depends on the time scale for the reorganization process and cannot be answered based on our current data. Most β-sheets we observe contain some strands far from the edges whose orientation must change for the sheet to become parallel. This could, in principle, occur through breaking and joining of β-sheets [60], but whether that is viable mechanism for the system studied here is unclear. Another mechanism is repeated attachment/detachment of edge strands. The time scale for changing the orientation of a central strand by this mechanism is, however, unknown.
The conformational reorganization of soluble β-sheet aggregates, toward more ordered structure, has been investigated experimentally by Decatur and coworkers using isotope-edited FTIR spectroscopy for Aβ16–22 and H1, a 14-residue fragment of the prion protein [61]–[63]. Two competing mechanisms of reorganization were proposed: the detachment-reattachment of strand(s) from/onto existing β-sheets, which was found to dominate at low concentration, and the sliding or reptation of individual strands without detachment from the aggregate, which was found prevalent at high concentration. The reptation motion has been observed in simulations for Aβ16–22 [22],[64], TTR105–115 [32] and GNNQQNY [65]. Both the above mechanisms of β-sheet reorganization were seen in our simulations, along with large-scale motion of whole β-sheets relative to each other.
The time required for the conversion of early formed aggregates into growth-competent ones depends on the character of the early aggregates, and therefore on sequence. The process need not be faster if the early aggregates are β-sheet rich, because the system then has to escape from deep unwanted minima. Our results suggest that β-sheet rich aggregates form fast for AcPHF6, but the reorganization needed for further growth may be slow. This finding is consistent with the Aβ16–22 and H1 results of Decatur and coworkers [61]–[63], although the precise character of the reorganization process may have been different in their systems. For AcPHF6, we find that changes in the parallel/antiparallel organization are an important part of the reorganization process.
While our simulated aggregates show features reminiscent of the proposed dry steric zipper model [16] for AcPHF6 fibrils, we found no indication that a nanotubular structure [15] would emerge with increasing aggregate size. It must be kept in mind, however, that there is a huge gap in size between our simulated aggregates and full fibrils.
We have carried out extensive seeded and unseeded Monte Carlo simulations of the aggregation of the peptide AcPHF6, derived from the tau protein, using an all-atom protein model with a simplified interaction potential. Our results suggest that the formation of stable AcPHF6 oligomers occurs through a nucleation process. In the pre-nucleation phase, a variety of meta-stable aggregates formed and dissolved. At the nucleation stage, the aggregates had already acquired a large fraction of β-sheet structure; no completely disordered aggregates of significant size were seen in our simulations. The oligomers formed in this nucleation step are thus β-sheet rich, but they are not necessarily growth-competent. Our results indicate that further growth requires conformational reorganization. The reorganization process appears to be slow, and might be the main bottleneck to fibril formation for this peptide.
In some runs, large aggregates appeared, with ∼20 chains or more. All these aggregates were composed of two twisted β-sheets, packed against each other with V1, I3 and Y5 side chains forming the sheet-sheet interface. This kind of conformation bears a striking resemblance to the dry steric zipper structure that has been proposed for PHF6 fibrils [16]. Morover, while most aggregates we saw had a mixed parallel/antiparallel β-strand organization, there was a clear tendency that the fraction of parallel β-sheet structure increased with aggregate size. In the fibrils, the β-strand organization is known to be parallel [14]. From these observations, it is tempting to speculate that reorganization of the β-sheets into parallel ones is a key step in the formation of PHF6 fibrils.
The package PROFASI (PROtein Folding and Aggregation SImulator) [66] was employed in this work. The model is an implicit water all-atom description (including all hydrogen atoms) of the protein chains with only torsional degrees of freedom (without bond stretching and angle bending). In addition to these, each chain has three translational and three rotational degrees of freedom. The interaction energy of the model iswhere Eloc is a local electrostatic interaction between adjacent peptide units which influences the Ramachandran Φφ distribution, and the others are non-local terms. The excluded volume term Eev represents an r−12 repulsion between atom pairs. Ehb is an explicit hydrogen-bond term modeling backbone-backbone and charged side chain-backbone hydrogen bonds. Ehp describes an effective hydrophobic interaction between pairs of non-polar side chains which depends on the degree of contact of the two side chains. The details of each interaction term and the corresponding parameters can be found elsewhere [67],[68]. Whereas it is a minimalistic model, with the potential deliberately kept simple for the sake of clarity and computational efficiency, the model has successfully captured the folding thermodynamics and kinetics of peptides and small proteins, peptide aggregation, and the mechanical unfolding of a 76-residue protein [20], [45], [67]–[70].
All the simulations were carried out in a cubic cell with periodic boundary conditions at a constant temperature. The temperature was set to 0.46 in reduced units, corresponding to ca 308 K, for optimal computational efficiency. If the temperature is too high, the chains will not aggregate, while if the temperature is too low, the kinetic evolution of the system will be slow. The temperature studied is close to the experimental conditions [14].
The conformational MC updates included chain translations (6.65%), chain rotations (6.65%), single-variable updates of side-chain (51.0%) as well as backbone (26.6%) angles, and Biased Gaussian Steps that favor local deformations of the protein [71] (9.1%). The results did not change when we slightly modified the relative frequencies of the different moves, for example, to (6.1%, 6.1%, 46.7%, 24.4%, 16.7%) or (7.3%, 7.3%, 56.1%, 29.2%, 0.0%). Note that none of these updates move more than one chain at a time. The conformations were saved every 103 MC steps. Since the peptide is very short, a single backbone torsion angle change does not yield drastic changes in the global structure as it would for long peptide chains. For this reason, we believe that for this peptide, the MC dynamics mimic the random motions of the peptides and can be interpreted as a discrete form of Brownian dynamics [72], so that the events along the Markov chain in our MC simulation can be considered as a coarse-grained dynamic process. For conciseness, the unit of simulation time is 50 million MC steps in this article, so 1 step is equal to 5×107 MC steps unless noted otherwise. Note that one cannot compare the reaction rates of systems that have different numbers of degrees of freedom using the number of MC steps directly. When the number of degrees of freedom is doubled, to represent the same time scale, the required MC steps should also be doubled.
We performed both unseeded runs started from random initial conformations, and seeded runs, where the initial conformations contained aggregates from simulations of smaller systems. Our runs are summarized in Table 1. All statistical errors quoted were obtained by the jackknife method [73].
For the analysis of the simulation data, we defined the end-to-end vector of a chain as the one from the first backbone N atom to the last backbone C atom. When the acute angle between two normalized end-to-end unit vectors was <30° and the interstrand main-chain hydrogen bond energy was <−6.1 kcal/mol (corresponding to 2–3 hydrogen bonds), the two chains were considered to form a sheet. A pair of strands was defined as parallel (antiparallel) if the dot product of their normalized end-to-end vectors was between 0 and 1 (−1 and 0). To define the direction of a sheet, we used the average of all end-to-end vectors within the sheet, calculated after reversing antiparallel end-to-end vectors to make all vectors roughly point in the same direction. The relative orientation of two sheets was calculated as the acute angle between the direction vectors of the sheets. To describe the twist of a β-sheet, we defined the twist angle between pairs of adjacent strands in the sheet as the acute angle formed by the backbone direction vectors of two strands. The backbone direction vector was taken to start at the middle point of the peptide bond between Q2 and I3 and end at the middle point of the peptide bond between V4 and Y5. The first and last residues were omitted because they were often unstructured. The average twist angle within a sheet indicates the overall twisting of the sheet. A positive sign of the twist angle indicates a left-handed twist about the in-sheet axis perpendicular to the peptide chain. It corresponds to a right-handed twist about the axis running in the same direction as the peptide chain.
|
10.1371/journal.ppat.1000658 | Potential Benefits of Sequential Inhibitor-Mutagen Treatments of RNA Virus Infections | Lethal mutagenesis is an antiviral strategy consisting of virus extinction associated with enhanced mutagenesis. The use of non-mutagenic antiviral inhibitors has faced the problem of selection of inhibitor-resistant virus mutants. Quasispecies dynamics predicts, and clinical results have confirmed, that combination therapy has an advantage over monotherapy to delay or prevent selection of inhibitor-escape mutants. Using ribavirin-mediated mutagenesis of foot-and-mouth disease virus (FMDV), here we show that, contrary to expectations, sequential administration of the antiviral inhibitor guanidine (GU) first, followed by ribavirin, is more effective than combination therapy with the two drugs, or than either drug used individually. Coelectroporation experiments suggest that limited inhibition of replication of interfering mutants by GU may contribute to the benefits of the sequential treatment. In lethal mutagenesis, a sequential inhibitor-mutagen treatment can be more effective than the corresponding combination treatment to drive a virus towards extinction. Such an advantage is also supported by a theoretical model for the evolution of a viral population under the action of increased mutagenesis in the presence of an inhibitor of viral replication. The model suggests that benefits of the sequential treatment are due to the involvement of a mutagenic agent, and to competition for susceptible cells exerted by the mutant spectrum. The results may impact lethal mutagenesis-based protocols, as well as current antiviral therapies involving ribavirin.
| RNA viruses are associated with many important human and animal diseases such as AIDS, influenza, hemorrhagic fevers and several forms of hepatitis. RNA viruses mutate at very high rates and, therefore, can adapt easily to environmental changes. Viral mutants resistant to antiviral inhibitors are readily selected, resulting in treatment failure. The simultaneous administration of three or more inhibitors is a means to prevent or delay selection of resistant mutants. A new antiviral strategy termed lethal mutagenesis is presently under investigation. It consists of the administration of mutagenic agents to elevate the mutation rate of the virus above the maximum level compatible with virus infectivity, without mutagenizing the host cells. Since low amounts of virus are extinguished more easily, the combination of a mutagen and inhibitor was more efficient than a mutagen alone in driving virus to extinction. Here we show that foot-and-mouth disease virus replicating in cell culture can be extinguished more easily when the inhibitor guanidine is administered first, followed by the mutagenic agent ribavirin, than when both drugs are administered simultaneously. Interfering mutants that contribute to extinction were active in the presence of ribavirin but not in the presence of guanidine. This observation provides a mechanism for the advantage of the sequential versus the combination treatment. This unexpected effectiveness of a sequential treatment is supported by a theoretical model of virus evolution in the presence of the inhibitor and the mutagen. The results can have an application for future lethal mutagenesis protocols and for current antiviral treatments that involve the antiviral agent ribavirin when it acts as a mutagen.
| The capacity of rapidly multiplying parasites to adapt to the changing environment of their host organisms is a formidable obstacle for the control of the diseases they bring about, and an unsolved problem in medical and veterinary practice. RNA viruses are remarkably adaptable due to their elevated mutation rates and quasispecies dynamics [1]–[4]. A means to compromise the replication of highly variable RNA viruses is to enhance their mutation rates above the maximum level compatible with expression of their genetic program and completion of their life cycle. The existence of an error threshold for genetic stability was predicted by quasispecies theory, that established a correlation between the average error rate and the complexity of the genetic information that can be reproducibly maintained [1]–[3],[5]. This concept has been supported by additional theoretical treatments [5]–[18]. RNA virus genomes and other RNA genetic elements have limited length (encoded information), and, accordingly, they tolerate the high mutation rates that they display during replication [19]–[21]. If in the course of virus replication the error rate is elevated above the error threshold, the result should be the extinction of the virus, in a process that has been termed lethal mutagenesis. Extensive experimental evidence has documented viral extinction upon replication of RNA viruses in the presence of mutagenic agents, in particular mutagenic nucleoside analogues [22]–[35].
In an exploration of variables that could influence extinction of the picornavirus foot-and-mouth disease virus (FMDV), it was noted that low viral fitness and low viral load favored extinction [33],[36]. As a consequence, combinations of mutagenic agents and antiviral inhibitors were more effective than mutagenic agents alone in driving viral populations towards extinction [36]–[39]. The advantage of a combination of a mutagenic agent and an antiviral inhibitor in lethal mutagenesis was in agreement with the requirement of combination treatments to delay or prevent selection of escape mutants when employing non-mutagenic antiviral inhibitors, as supported by theoretical studies and clinical observations [40]–[46].
The nucleoside analogue ribavirin [1-(β-D-ribofuranosyl)-1H-1,2,4-triazole-3-carboxamide] (R) is a licensed antiviral agent shown to be mutagenic for a number of RNA viruses [47]–[53], currently used in investigations on lethal mutagenesis. Viral mutants with decreased sensitivity to R have been described for several viruses, including the picornaviruses poliovirus (PV) and FMDV [54]–[59]. The presence of a mutation that decreased the sensitivity of FMDV to R resulted in extinction failure when the virus was replicated in the presence of high concentrations of R [60]. Extinction was achieved with an alternative mutagenic treatment with 5-fluorouracil [60]. Therefore, it is important to understand the molecular events that underlie virus extinction by mutagenic agents and to explore protocols that avoid selection of escape mutants.
Populations of FMDV and the arenavirus lymphocytic choriomeningitis virus (LCMV) subjected to lethal mutagenesis underwent a decrease of specific infectivity (with loss of infectivity preceding loss of RNA replication capacity), and pre-extinction, mutagenized RNA interfered with replication of infectious RNA [27],[61]. These results were further supported by simulations in silico, and suggested that interference by defective (non-infectious but replication-competent) genomes could contribute to loss of infectivity. These defective genomes were termed defectors, and their involvement in loss of infectivity led to the proposal of the lethal defection model of virus extinction, [27],[61], a model which was further strengthened by additional theoretical studies [62]. In agreement with this proposal, specific capsid and polymerase mutants of FMDV exerted complementation (at early times post-electroporation) and interference (at late times post-electroporation) when co-electroporated into cells together with infectious FMDV RNA [63]. Related FMDV mutants that were not competent in RNA replication did not exert any detectable interference [63]. In line with these observations, replication of drug-resistant poliovirus was inhibited by trans-acting, dominant negative mutants [64]. Therefore, three critical interconnected parameters have been identified as playing a role in viral extinction by lethal mutagenesis: mutation rate, interference by defector genomes, and viral load; the latter can be decreased by the presence of antiviral inhibitors [33],[61],[63],[65].
The combined use of mutagenic agents and antiviral inhibitors poses an evolutionary and a virological riddle to the system. First, the mutagenic agent, that is included in the drug cocktail, can favor the generation of viral mutants resistant to the inhibitors co-administered with the mutagen [36],[38]. Second, if intracellular replication is necessary for mutants to exert interference [63], the presence of an antiviral inhibitor together with the mutagenic agent may jeopardize the interfering activity of mutants by preventing or diminishing their accumulation during mutagenesis. However, if inhibitor-resistant mutants are generated, they could contribute defector genomes induced by the mutagen. Here we address this key issue using FMDV, specific interfering capsid and polymerase FMDV mutants, the inhibitor of FMDV replication guanidinium hydrochloride (GU), and R.
R eliminates FMDV from persistently and cytolytically infected cells in culture, and its mechanism of action includes mutagenesis of the viral RNA [59],[66],[67]. Failure to extinguish FMDV by a sequential treatment with either fluorouracil or R first, followed by GU, was systematically associated with selection of GU-resistant FMDV mutants, with amino acid replacements in non-structural protein 2C [36],[37],[60] (Perales et al., unpublished results). Here we show that, contrary to expectations from antiviral designs with non-mutagenic inhibitors, a sequential treatment first with GU and then with R can be more effective in driving the virus to extinction than the administration of GU and R in combination. Quantification of the interference exerted by specific FMDV mutants on the replication of standard FMDV suggests that the molecular basis of the advantage of the sequential treatment is that interfering FMDV mutants can exert a suppressive activity in the presence of R, when GU is not present. However, the presence of GU at the same time than R in the combination treatment would jeopardize the interfering activity of the mutants by inhibiting their replication, thereby contributing to virus survival. The results are supported by a theoretical model that predicts the evolution of a viral population in the presence of a mutagenic agent and an antiviral inhibitor. From a practical point of view, the proposed treatment protocol has the advantage that the simultaneous administration of two drugs is avoided, and that low mutagen doses may be sufficient to effect viral clearance.
To compare a sequential versus combination treatment involving GU and R, we first determined the maximum GU concentration that permitted at least two passages of FMDV pMT28 (virus rescued from an infectious transcript; see Materials and Methods) with a decrease of viral load, prior to dominance of GU-resistant mutants [36],[37]. The results (Figure 1) indicate that there is a recovery of virus infectivity in the presence of GU concentrations ranging from 8 to 18 mM, but not 20 mM, and this is compatible with the dominance of GU-resistant mutants, as we have previously shown [36],[37]. GU concentrations of 16, 18 and 20 mM led to a decrease in virus progeny production for at least two consecutive passages, and, therefore, these GU concentrations were chosen to compare the efficacy of a sequential versus combination treatment involving GU and R.
A sequential treatment first with GU and then with R was compared with treatment with R or GU alone, and with a combination treatment in which the inhibitor and the mutagenic agent were administered simultaneously. Drug doses and times of exposure were comparable in the different treatment protocols. To this aim, FMDV pMT28 populations were subjected either to one passage in the presence of GU (16, 18 or 20 mM) followed by three passages in the presence of R (5 mM), or to four passages in the presence of a combination of GU (16, 18 or 20 mM) and R (5 mM), or at least five passages in the presence of either R (5 mM) or GU (16, 18 or 20 mM) alone (Figure 2A). The sequential treatment is abbreviated as +GU,+R, and the combination treatment as [+GU+R]. The results (Figure 2B, C, D) show that the decrease in virus titer and in viral RNA levels was more rapid when GU was applied first and then R sequentially, as compared with the other treatment protocols. The second most effective treatment was achieved with the administration of GU and R at the same time, in a combination treatment [+GU+R]. At each passage, the virus titer and virus RNA levels attained one or more logarithms lower level with the sequential +GU,+R treatment than with the combination treatment (Figure 2B, C, D) (p<0.008 for virus titer in all cases; p<0.0001 for viral RNA levels in all cases; Student's t test). The decrease in virus titer and viral RNA levels was more accentuated as the GU concentration was increased. To ascertain that the decrease in the viral replication correlates with the extinction of the viral population, RT-PCR amplifications for every passage were performed. FMDV is considered extinct when no infectivity and no viral RNA can be detected in the cell culture supernatant, using a highy sensitive RT-PCR protocol, and after 3 blind passages in the absence of any drug (see Materials and Methods for further details). The advantage of the sequential treatment was evidenced by loss of FMDV-specific genetic material, indicative of virus extinction, at the same or at earlier passages for sequential than either combination treatment or treatment with GU or R alone (Figure 2E). The results using GU or R alone showed a recovery in virus titer and RNA levels for GU concentrations of 16 mM and 18 mM, whereas virus was not extinguished until passage 6 using R alone.
The advantage of the sequential +GU,+R over the combination [+GU+R] treatment was surprising in view of the previously established (and broadly accepted) requirement of a combination therapy to minimize or prevent selection of viral mutants resistant to antiviral agents [40]–[46]. We suspected that the critical difference could be the involvement of a mutagenic agent in the treatment. Mutants generated in the course of R mutagenesis may play an important role in the extinction of RNA viruses during lethal mutagenesis treatments [27],[62],[63]. Thus, one interpretation of the lower infectivity and viral loads as a result of the sequential +GU,+R treatment over the combination [+GU+R] treatment is that the interfering mutants generated by R-mutagenesis might have been suppressed by GU when this inhibitor is added at the same time. To investigate this possibility, we used previously characterized interfering and non-interfering capsid and polymerase mutants of FMDV [63]. BHK-21 cells were co-electroporated with standard, infectious FMDV RNA and combinations of either interfering FMDV mutants or of non-interfering FMDV mutants, in the absence or presence of GU and R. The concentrations of GU and R that decreased at least one logarithm the infectious progeny production at 3 hours post-electroporation were previously determined (data not shown). The results (Figure 3 and Table 1) show that the interfering combination of capsid mutant Q2027A and polymerase mutant MD exerted an interference in the presence of R at all times post-electroporation tested. In contrast, in the presence of GU, the interference decreased significatively at late times post-electroporation (Figure 3A). The average interference (of the values at 3, 5 and 6 h post-electroporation) exerted by the mutant combination Q2027A+MD was 81% in the absence of GU and 61% in the presence of GU. This decrease is statistically significant (p = 0.004 for the 20% decrease with a 95% confidence interval of 29.6% to 10.4%; Student's t test). In contrast, the average interference exerted by the same mutant combination was 81% in the absence of R and 94% in the presence of R. This increase is statistically significant (p = 0.005 for the 12% increase with a 95% confidence interval of 6.37% to 19.58%). When expressed as the interference measured at each time point (Table 1) the results indicate strong suppresion of interference by GU (but not by R) at the late time point at which the Q2027A+MD combination exerts its highest effect [63]. The results suggest that GU may inhibit replication of defector genomes, while R may contribute to the generation of additional defector genomes (see Discussion). No significant differences were observed in the virus titers produced in the presence of the combination of non-interfering mutants DMD+D3 in the presence and absence of GU (p = 0,5; Student's t test), as expected. The 2C-coding region of the progeny of the RNAs that were electroporated in the presence of GU was sequenced. No mutations associated with GU-resistance were detected in any sample. Thus, the decrease in interference in the presence of GU, upon replication immediately following electroporation, was not due to replication of GU-resistant mutants. These experimental results suggest that the mechanism by which the sequential treatment is more efficient than the combination treatment in decreasing the viral load and driving a virus towards extinction, is that in the sequential treatment the inhibitor cannot prevent replication of defector genomes that are generated by the mutagenic agent.
The lethal defection model of virus extinction was proposed on the basis of experimental results with FMDV and LCMV and of a computational model with virtual viable and defective genomes that replicated in the course of a persistent LCMV infection of BHK-21 cells [27]. Current evidence suggests that trans-active viral gene products harboring amino acid substitutions may impair functions of standard genomes whenever protein complexes are required for function (homo and hetero-polymers among viral proteins or between viral and host proteins). For further discussion of possible interference mechanisms, see [4],[25],[27],[61],[64],[65]. Using realistic parameters for viral genome replication and mutation rates, the model predicted a decisive participation of defector genomes in the extinction of the viable class of genomes [27],[60]. Therefore, we have now developed another model for the evolution of viral populations that replicate in serial cytolytic infections under increased mutagenesis, and in the presence of an inhibitor of viral replication. We have asked whether this new model predicts the experimental results reported here, in particular an advantage of sequential over combination treatment.
The model considers four different types of individuals in the population: wild-type susceptible to the inhibitor (WTs), defective susceptible to the inhibitor (Ds), wild-type resistant to the inhibitor (WTr), and defective resistant to the inhibitor (Dr), but does not consider any direct interference exerted by the defective genomes on wild type replication (Figure 4). Individuals of the wild-type are able to replicate by themselves when they infect a cell; defective individuals cannot replicate in absence of wild-type individuals. WTs and Ds replicate more slowly in the presence than in the absence of the inhibitor, while WTr and Dr are not affected by the inhibitor. The natural mutation rate during genome replication is w, which for simplicity is equated to the rate of production of defective forms, generated upon replication of the wild type. Individuals resistant to the inhibitor are generated at a rate μ = k w, with k = 10−3 (values for the coefficient k ranging from 10−2 to 10−5 do not affect the results qualitatively). Hence, an increase in the mutation rate has two effects: a larger number of defective genomes are generated, and the probability to develop resistance to the inhibitor increases (Figure 4).
We consider a total of N = 107 cells which can be infected at each passage. The maximum number of virions entering a cell is one. The initial virus replicates inside the cell for a number of cycles G = 5. When the inhibitor is absent, the number of progeny genomes per parental genome is given the value R = 1.5 for wild-type individuals, and r = 1 for defective individuals. When the inhibitor is present, the replicative parameters R and r are multiplied by a factor 0<i<1. The natural mutation rate is set to w = 0.05 in the examples that follow. At passage 0, the population is composed only of individuals sensitive to the inhibitor; resistant individuals will be generated by mutation. Replication inside each cell follows a deterministic process according to the above description. Let us call WTs(g), Ds(g), WTr(g), and Dr(g) the four populations at replication cycle g. In one replication cycle, the number of particles produced is
WTs(g+1) = R i (1−μ−w) WTs(g)
WTr(g+1) = R (1−w) WTr(g)+R i μ WTs(g)
Ds(g+1) = i Ds(g)+R i w WTs(g)
Dr(g+1) = Dr(g)+R w WTr(g)
We define the factor of inhibition (to be given in percent) as fi = 100 * (1−i), such that fi = 0% indicates no inhibition (in the equations above, parameter i = 1) and complete inhibition (or forbidding replication), corresponds to fi = 100% (i = 0). The total population produced by the cell in G viral replication cycles is the result of iterating G times the equations that define the model. The total viral population P after one passage is the sum of the production by the N cells. This final viral population is then used to seed a new ensemble of N cells, to start the next passage. The average number of virions entering each cell is P/N, unless it exceeds one (which is the maximum we allow in these examples).
A natural population evolves through passages with a mutation rate w = 0.05, and in the absence of inhibitor (i = 1). The addition of a mutagen changes the mutation rate to a higher value of w, while the addition of an inhibitor diminishes the replication rate of susceptible particles in an amount i<1. The example given in Figure 5A demonstrates the dynamics of the model in some of the experimental situations reported in the manuscript. When no mutagen and no inhibitor are present, the population evolving under the natural parameters reaches a high titer (measured as the number of wild-type individuals). When a mutagen is added (in practice, the value of w is set to 0.6), extinction after several passages occurs. Different curves show how the population recovers when the treatment with the mutagen is interrupted beginning at passages two, three, and four. This behavior is that observed when GU-escape mutants are selected in populations treated with a mutagen and GU ([36],[37],[60]; Perales et al., unpublished results).
In the second case (Figure 5B), we describe the dynamics when different amounts of inhibitor are used, analogous to the experimental assays presented in Figure 1. Increasing concentrations of inhibitor result in a gradually steeper decrease in the number of progeny particles. However, except for the highest inhibitor concentration tested, resistant mutants are selected, and progeny production finally reaches the level attained in the absence of inhibitor. The simulation agrees with the experimental results (compare Figure 1 and Figure 5B).
Finally, in Figure 5C we represent the dynamics of the population under different combinations of inhibitor and mutagen. The examples are chosen to mimic the experiments presented in Figure 2. Sequential therapy causes extinction faster than combination therapy, and a contributing factor is the appearance and fixation of resistant mutants during the combination treatment, promoted by the mutagen. The superiority of sequential therapy is observed over a broad range of parameters, though it is not generic for this model. The role played by parameters such as the replication rate of the wild type, the number of replication cycles inside a cell, or the relative effect of mutagen and inhibitor and how they interact when applied jointly will be explored in future studies (see also Discussion). Both, the experimental results and the theoretical model (for the parameters used) suggest an approximately ten-fold increase in the viral yield in the case of combination therapy compared to sequential therapy.
One of the major consequences of quasispecies dynamics for pathogenic RNA viruses is that subpopulations of viruses from the mutant spectra, that harbor mutations that decrease the sensitivity to antiviral inhibitors, are rapidly selected (review in [68]). Lethal mutagenesis exploits high mutation rates of RNA viruses [19],[20] to increase the average error rate during viral replication even further, until meaningful genetic information and viral functions deteriorate, and the virus is extinguished [22]–[35].
New nucleoside analogues are currently investigated as possible virus-specific mutagenic agents that could be included in lethal mutagenesis protocols [22], [29], [35], [69]–[71]. In addition to safety issues concerning adverse activity on the host cells (related to mutagenesis of cellular DNA, or other effects), the administration of virus-specific mutagenic base or nucleoside analogues requires careful consideration of protocols when antiviral inhibitors are co-administered with the mutagenic agents [72]. This has become particularly relevant with recent observations that suggest that replication-competent subsets of defective viral genome subpopulations termed defectors may participate in the process of viral extinction [27], [61]–[63]. Interference by defectors was specific for their corresponding standard viruses, was not due to induction of IFN or other unspecific cellular effectors, and it required replication of the interfering genomes [27],[61],[63]. In consequence, the presence in the infected cell of a mutagenic nucleotide together with an antiviral inhibitor may jeopardize extinction because the inhibitor will reduce replication of interfering genomes. This has been addressed experimentally in the present study, and, indeed, GU, but not R, attenuated the interfering activity exerted by a combination of a polymerase and capsid mutant of FMDV. In agreement with the attenuation of interference by GU, the decrease in FMDV infectivity, viral load and attainment of viral extinction are more effective with a sequential +GU,+R treatment than with the combination treatment [+GU+R], or treatment with R or GU alone (Figures 1 and 2). The differences were investigated using a constant R concentration of 5 mM, and GU concentrations in the range of 16 mM to 20 mM. Although eventually all treatment regimes achieved extinction (Figure 2C), the sequential +GU,+R treatment led earlier to low infectivity and RNA levels than the corresponding combination treatment (Figure 2B). In an in vivo scenario of application of lethal mutagenesis, an earlier and sustained decrease in viral load may provide the host immune response with an opportunity to effect viral clearance. An added benefit of a sequential treatment is that toxicity or antagonistic effects, additional to suppresion of interference, derived from simultaneous administration of two drugs, are avoided.
It must be stressed that the benefits of a sequential treatment do not hold for standard non-mutagenic antiviral agents, for which combination treatments are essential to prevent selection of inhibitor-resistant mutants [40],[41],[73]. According to our experimental and theoretical results there are two key influences that favor the inhibitor-mutagen sequential treatment, and that do not operate when only non-mutagenic inhibitors are involved. One is that the mutagenic agent increases the probability of selection of inhibitor-escape mutants, and this probability increases with the viral load. The second influence is that the interfering activity of defector genomes is important to drive the population towards extinction. The administration of the inhibitor will produce a decrease in viral load, that will render the system more susceptible to mutagenesis-mediated extinction, allowing expression of interfering activities associated with the mutagenized spectrum of mutants [33]. No mutations in the 2C-coding regions that confer resistance to GU have been detected in passage 2 of sequential or combination treatment in the presence of 16 mM GU (experiment of Figure 2). Thus, GU-escape mutants were not a factor in the disadvantage of the combination treatment. At present, we cannot exclude that other mechanisms may also contribute to the observed benefits of a sequential inhibitor-mutagen treatment.
The experimental results are supported by a simple model of viral evolution taking into account the minimal ingredients that describe the experimental system. Our numerical results indicate that four different viral types, as discussed, are essential to reproduce the observed dynamics. According to the model, the fast generation and fixation of resistant mutants (controlled by parameter μ, and the number G of replication cycles inside the cell) is the essential mechanism conferring advantage to sequential therapy. In the absence of a mutagenic activity, the model predicts benefits of combination therapy, as in previous models of virus dynamics in connection with drug therapy [40]. Future work will explore the range of parameters that provide the strongest advantages to either therapy, as well as the relevance of other dynamical rules (such as the inclusion of lethal mutations or more detailed relationships between genomic mutations and their effect on fitness) in the behavior of the model system. For example, the current model predicts that as the interfering activity of defectors increases, the advantage of the sequential over the combination treatment is gradually lost (Manrubia et al. unpublished results). It is not known why under strong suppression (unrealistic for mutants generated by random mutagenesis within the quasispecies) the response of the system is that expected for administration of classical, non-mutagenic antiviral inhibitors, and this point is under investigation.
The results reported here can impact current antiviral therapies that involve R, such as the combination of pegylated IFN-α (PEG-IFN-α) and R for treatment of human HCV infections. In these treatments, IFN can have multiple effects at the level of the entire organisms, and also it is not clear whether ribavirin acts as a mutagenic agent, which would imply a lethal mutagenic action, or by other mechanisms, or combination of mechanisms [74]–[82]. For patients who respond to IFN-α treatment, and their HCV load is reduced, a sequential +IFN-α (or PEG- IFN-α),+R treatment may be advantageous over the corresponding combination treatment. Our experimental and theoretical results predict this to be true to the extent that R acts as a mutagenic agent for HCV, and that mutagenesis is its major mechanism of action against HCV in vivo. However, R has multiple affects on cell metabolism [83]–[87] it is not easy to assess which is the contribution of mutagenesis and defector genomes, in the course of treatments of HCV infections [34], [66], [87]–[90]. The first clinical trials, that consisted in the administration of recombinant IFN-α to chronic non-A, non-B hepatitis, resulted in improvement of aminotransferase levels and liver histology [91]. Subsequent treatments, once HCV had been identified, involved IFN-α alone or in combination with R [92],[93]. Some early clinical trials documented benefits of a combination [IFN-α,+R] treatment versus either treatment with IFN-α alone, or sequential treatment first with R and then with IFN-α, in chronic HCV infections [94]. More recent trials established a higher efficacy of PEG-IFN-α over conventional IFN-α, in combination treatments with R [95]. Other trials have compared IFN-α or R monotherapy with combination therapies or sequential therapies involving administration of R first [74], [96]–[99]. In a trial for the treatment of chronic HCV and HIV-1 in doubly-infected hemophiliacs, IFN-α-2b was administered as monotherapy for one month, and then oral R was added to the treatment [100]. To our knowledge, no systematic trials have involved sequential treatment with IFN-α first, and then with R alone, precisely the protocol predicted to be more effective, according to our results.
Exploration of sequential inhibitor-mutagenic treatments for HCV infections may become more relevant in the face of the new generation of specific inhibitors of HCV replication, now at different stages of development for clinical practice [101],[102]. Our prediction should hold also for treatment of other chronic viral infections, such as human hepatitis B virus for which both mutagenic nucleoside analogues and non-mutagenic inhibitors are available for treatment [103],[104]. It is obvious, however, that because of the complexities involved in pharmacological activities in vivo [83]–[87], experiments with animal models are needed to explore whether results in vivo will be those predicted by our model experiments in cell culture, and by the theoretical study.
The origin of BHK-21 cells and procedures for cell growth in Dulbecco's modification of Eagle's medium (DMEM), and for plaque assays in semisolid agar have been previously described [105],[106]. The viruses used in the experiment are the following: FMDV C-S8c1 is a plaque-purified derivative of serotype C isolate C1 Santa Pau-Sp70 [106]. An infectious clone of FMDV C-S8c1, termed pMT28 was constructed by recombining into a pGEM-1 plasmid subclones that represented the C-S8c1 genome, as described [107],[108]. Thus, FMDV pMT28 used in the experiments is the progeny of infectious transcripts that express the standard FMDV C-S8c1. Capsid (Q2027A) and polymerase (MD, DMD and D3) mutants were previously described, and characterized biologically and with an interference index [63],[109]. To control for the absence of contamination, mock-infected cells were cultured and their supernatants were titrated in parallel with the infected cultures; no signs of infectivity or cytopathology in the cultures or in the control plaque assays were observed in any of the experiments.
A solution of guanidine (GU) in DMEM was prepared at a concentration of 50 mM, sterilized by filtration, and stored at 4°C. Prior to use, the stock solution was diluted in DMEM (Dulbecco's modification of Eagle's medium) to reach the desired concentration. For infections of BHK-21 cells with FMDV in the presence of GU, no pretreatment of the cell monolayer with GU was performed. After addition of FMDV and washing of the cell monolayers, infections were allowed to continue in the presence of GU. For each passage 2×106 BHK-21 cells were infected with supernatant of virus from the previous passage (0.2 ml), and the infection allowed to proceed for about 24 h. The multiplicity of infection (MOI) ranged from 1×10−5 to 1×10−1 PFU/cell, and the MOI for each passage can be calculated from the infectivity values given in the experiment (Figure 1). Infections in the absence of GU, and mock-infected cells were maintained in parallel; no evidence of contamination of cells with virus was observed at any time.
A solution of R in PBS was prepared at a concentration of 100 mM, sterilized by filtration, and stored at −70°C. Prior to use, the stock solution was diluted in DMEM to reach the desired R concentration. For infections in the presence of R, cell monolayers were pretreated during 7 h with 5 mM R prior to infection. FMDV C-S8c1 was passaged serially in the absence or in the presence of R (5 mM). After addition of FMDV and washing of the cell monolayers, the infection was allowed to proceed in the presence of the same concentration of R. For each passage 2×106 BHK-21 cells were infected with supernatant of virus from the previous passage (0.2 ml) and the infection continued for about 24 h. The multiplicity of infection (MOI) ranged from 5×10−6 to 1×10−1 PFU/cell, and the MOI for each passage can be calculated from the infectivity values given for each experiment (Figures 2B, C, D). The passage experiments described in Figures 1 and 2 occurred over a 24-hour period, and therefore each passage is comprised of multiple replication cycles. Infections in the absence of R, and mock-infected cells were maintained in parallel; no evidence of contamination of cells with virus was observed at any time.
FMDV was considered extinct when no virus infectivity and no viral RNA that could be amplified by a highly sensitive RT-PCR protocol, could be demonstrated neither in the supernatant of the passage that harbors the putatively extinghuished virus, nor after 3 blind passages in BHK-21 cells, in the absence of any drug. Multiple highly sensitive RT-PCR amplification reactions that yield short cDNAs were carried out to ascertain extinction. Some of the gels that did not show a visible band were overexposed to ascertain absence of detectable DNA. These criteria to consider FMDV extinct [33],[36],[37],[59],[66] have now been extended to show that no infectivity or RT-PCR amplifiable material can be retrieved after passaging of the cells that harbor the putatively extinguished virus. This extension was prompted by the observation that FMDV subjected to hundreds of plaque-to-plaque transfers could lose capacity to form plaques and yet maintain intracellular RNA [110]. It should be noted that infectivity below the level of detection did not necessarily imply extinction (see Figure 2).
Viral RNA was extracted from the medium of infected cells using Trizol (Invitrogen) as previously described [110]. Reverse transcription was performed with AMV reverse transcriptase (Promega), and PCR amplification was carried out using Expand High Fidelity (Roche), as specified by the manufacturers. The 3D-coding region was amplified using as primers oligonucleotide A2SacI (5′- CACACATCGACCCTGAACCGCACCACGA; sense orientation; the 5′ nucleotide corresponds to genomic residue 6581), and oligonucleotide AV4 (5′- TTCTCTTTTCTCCATGAGCTT; antisense orientation; the 5′ nucleotide corresponds to genomic residue 7071). Genomic residues are numbered as described in [111]. Amplification products were analyzed by agarose gel electrophoresis using HindIII-digested Ф-29 DNA as molar mass standards. Negative controls (amplifications in the absence of RNA) were included in parallel to ascertain absence of contamination by template nucleic acids.
Real time quantitative RT-PCR was carried out using the Light Cycler RNA Master SYBR Green I kit (Roche), according to the instructions of the manufacturer and as described previously for FMDV RNA [110]. The 2C-coding region was amplified using as primers oligonucleotide 2CR2 (5′- GGCAAACCCTTCAGCAGTAAG; sense orientation; the 5′ nucleotide corresponds to genomic residue 4924), and oligonucleotide 2CD3 (5′- CGCTCACGTCGATGTCAAAGTG; antisense orientation; the 5′ nucleotide corresponds to genomic residue 5047). Quantification was relative to a standard curve obtained with known amounts of FMDV RNA, synthesized by in vitro transcription of FMDV cDNA (plasmid pMT28). The specificity of the reaction was monitored by determining the denaturation curve of the amplified DNAs. Negative controls (without template RNA and RNA from mock-infected cells) were run in parallel with each amplification reaction, to ascertain absence of contamination with undesired templates.
Plasmid DNA was linearized by cleavage with the appropiate restriction enzymes (pO1K/C-S8c1, and the capsid mutant plasmids with Hpa I and pMT28 derivates with Nde I, as previously described [63]). Then, the plasmids were purified by Wizard PCR Preps DNA purification resin (Promega), and dissolved in RNase-free water. FMDV RNA was transcribed from the linearized plasmids by using the Riboprobe in vitro transcription system (Promega). The mixture contained transcription buffer (Promega), 10 mM dithiothreitol, 0.48 units/µl RNasin, 1 mM each of ribonucleoside triphosphates, 4 ng/µl linearized plasmid DNA, and 0.3 or 0.4 units/µl SP6 or T7 RNA polymerase; it was incubated for 2 h at 37°C. The RNA concentration was estimated by agarose gel electrophoresis, with known amounts of rRNA as markers.
To electroporate BHK-21 cells with RNA transcribed in vitro, subconfluent cells were harvested, washed with ice-cold phosphate-buffered saline (PBS), and resuspended in PBS at a density of about 2.5×106 cells/ml. Aliquots (50–80 µl) of transcription mixture with the apropiate amount of RNA were added to 0.4 ml of cell suspension, and the mixtures were transferred to 2 mm electroporation cuvettes (Bio-Rad). Electroporation was performed at room temperature by two consecutive 1.5 kV, 25 µF pulses using a Gene Pulser apparatus (Bio-Rad), as described [63]. As control, BHK-21 cells were electroporated with 50–80 µl of transcription mixture in PBS to monitor absence of contamination. The cells were then resuspended in growth medium and seeded onto culture plates. At 3, 5 and 6 hours post-electroporation, samples of cells and culture medium were withdrawn and after three cycles of freezing at −70°C and thawing at room temperature, the lysate was stored at −70°C. Mock-coelectroporated cultures were treated in parallel and served as control in the titration of virus infectivity. No evidence of viral contamination was obtained in any of the experiments.
|
10.1371/journal.pgen.1002964 | UTX and UTY Demonstrate Histone Demethylase-Independent Function in Mouse Embryonic Development | UTX (KDM6A) and UTY are homologous X and Y chromosome members of the Histone H3 Lysine 27 (H3K27) demethylase gene family. UTX can demethylate H3K27; however, in vitro assays suggest that human UTY has lost enzymatic activity due to sequence divergence. We produced mouse mutations in both Utx and Uty. Homozygous Utx mutant female embryos are mid-gestational lethal with defects in neural tube, yolk sac, and cardiac development. We demonstrate that mouse UTY is devoid of in vivo demethylase activity, so hemizygous XUtx− Y+ mutant male embryos should phenocopy homozygous XUtx− XUtx− females. However, XUtx− Y+ mutant male embryos develop to term; although runted, approximately 25% survive postnatally reaching adulthood. Hemizygous X+ YUty− mutant males are viable. In contrast, compound hemizygous XUtx− YUty− males phenocopy homozygous XUtx− XUtx− females. Therefore, despite divergence of UTX and UTY in catalyzing H3K27 demethylation, they maintain functional redundancy during embryonic development. Our data suggest that UTX and UTY are able to regulate gene activity through demethylase independent mechanisms. We conclude that UTX H3K27 demethylation is non-essential for embryonic viability.
| Trimethylation at Lysine 27 of histone H3 (H3K27me3) establishes a repressive chromatin state in silencing an array of crucial developmental genes. Polycomb repressive complex 2 (PRC2) catalyzes this precise posttranslational modification and is required in several critical aspects of development including Hox gene repression, gastrulation, X-chromosome inactivation, mono-allelic gene expression and imprinting, stem cell maintenance, and oncogenesis. Removal of H3K27 trimethylation has been proposed to be a mechanistic switch to activate large sets of genes in differentiating cells. Mouse Utx is an X-linked H3K27 demethylase that is essential for embryonic development. We now demonstrate that Uty, the Y-chromosome homolog of Utx, has overlapping redundancy with Utx in embryonic development. Mouse UTY has a polymorphism in the JmjC demethylase domain that renders the protein incapable of H3K27 demethylation. Therefore, the overlapping function of UTX and UTY in embryonic development is due to H3K27 demethylase independent mechanism. Moreover, the presence of UTY allows UTX-deficient mouse embryos to survive until birth. Thus, UTX H3K27 demethylation is not essential for embryonic viability. These intriguing results raise new questions on how H3K27me3 repression is removed in the early embryo.
| Post-translational modifications of histones establish and maintain active or repressive chromatin states throughout cell lineages. Thus, the enzymes that catalyze these modifications often have crucial roles in establishing genomic transcriptional states in developmental decision-making. Histone methylation can stimulate gene activation or repression depending on which residues are targeted. Methylation of histone H3 on Lysine 4 (H3K4me) is an active chromatin modification, while methylation on histone H3 Lysine 27 (H3K27me) is associated with repression of gene activity [1].
The polycomb repressive complex 2 (PRC2) methylates H3K27 [2], [3], [4], [5]. Within this complex, enhancer of zeste homolog 2 (EZH2) catalyzes di and tri-methylation of H3K27. Embryonic ectoderm development (EED) and suppressor of zeste homolog 12 (SUZ12) are additional PRC2 core components indispensible for PRC2 activity [6], [7], [8]. EZH1 is a secondary, less efficient H3K27 methyl-transferase that shares some overlapping redundancy with EZH2 in ES cells and epidermal stem cells [9], [10], [11], [12]. The PRC1 complex is recruited through H3K27 trimethylation for additional histone modification and chromatin compaction [13]. In embryonic stem (ES) cells, PRC2 targets and represses genes essential for developmental events [14], [15], [16], [17]. The promoters of these PRC2 targets typically contain “bivalent” chromatin marks with both active H3K4 and repressive H3K27 methylation [18], [19], [20]. Loss of PRC2 activity de-represses these genes but does not alter ES cell pluripotency [14]. However, mouse mutations in any of the three PRC2 core components are early embryonic lethal with gastrulation defects [7], [21],[22].
H3K27 trimethylation is reversible as a family of histone demethylases catalyzes the removal of this epigenetic mark [23], [24], [25], [26]. JMJD3 (KDM6B) is an autosomal H3K27 demethylase upregulated during specific differentiation events [25], [27]. UTX (KDM6A) is a broadly expressed X-linked H3K27 demethylase that escapes X-inactivation [23], [24], [26], [28]. UTY is the Y chromosome homolog of UTX. Both UTX and JMJD3 demethylate H3K27 di and tri-methyl residues; however, UTY lacks this activity in vitro [26], [29]. Based on cell culture models, UTX and JMJD3 mediated H3K27 demethylation is vital in a wide array of functions including cell cycle regulation, M2 macrophage differentiation, neuronal stem cell specification, skin differentiation, and muscle differentiation [27], [30], [31], [32], [33], [34], [35]. In contrast, the biological function of UTY remains unknown.
Utx and Uty are genetically amenable to delineate H3K27me3 demethylation dependent versus demethylation independent function in mouse development. Comparative amino acid sequence analysis of UTX and UTY reveals 88% sequence similarity in humans (83% identity) and 82% sequence similarity in mouse. Across the annotated JmjC histone demethylase domain, the similarity is at 98% and 97% for human and mouse respectively. In the TPR (tetratricopeptide repeat) domain, the similarity is at 94%. So while UTY is reported to have lost H3K27 demethylase activity, it is remarkably well conserved with respect to UTX. Recent discoveries have revealed that JMJD3 functions in macrophage lipopolysaccharide response and lymphocyte Th1 response through H3K27 demethylase independent gene regulation [36], [37], suggesting that function of this family of proteins is not limited to histone demethylation. It has been hypothesized that X and Y chromosome homologs will escape X-inactivation in instances where the Y homolog has not lost functional activity and male to female dosage remains balanced [38]. Therefore, it is possible that UTX and UTY have functional overlap in H3K27 demethylase independent gene regulatory processes.
A recent publication by Lee et al. characterized heart defects in Utx homozygous embryos [39]. Cell culture experiments suggested that the phenotype resulted from H3K27 demethylase activity. Utx hemizygotes were reported to have a wide range of abnormalities, but it was not clear if any phenotypes overlap with the Utx homozygotes as no comparative data were illustrated. Given that Uty remained intact in these studies, it was not possible to conclude definitively whether Utx demethylase activity was essential for early embryonic development. Furthermore, it is not known whether mouse UTY is capable of H3K27 demethylation. The classification of UTY as having no demethylase activity is based on in vitro assays only. The possibility of in vivo demethylase activity due to other co-factors remains a possibility. Also, mouse UTY has considerable sequence divergence from human UTY. The two proteins are 75% identical overall, and 95% identical in the JmjC demethylase domain. Thus, it is possible that mouse UTY has retained demethylase activity.
In our study, we have generated mouse mutations in both Utx and Uty. Hemizygous Utx mutant male mice (XUtx− Y+) were runted at birth with only a small number surviving to adulthood. In contrast, Utx homozygous females (XUtx− XUtx−) had severe phenotypes mid-gestation, with developmental delay, neural tube closure, yolk sac, and heart defects. Unlike homozygotes, Utx hemizygotes lack mid-gestational cardiovascular defects and are recovered in Mendelian frequencies at E18.5. Furthermore, compound hemizygous male embryos (XUtx− YUty−) carrying mutations of both Utx and Uty phenocopy the Utx homozygotes. Thus, the disparity in hemizygous and homozygous Utx phenotypes is due to compensation by Uty in the hemizygous male embryos. We have utilized an in vivo H3K27 demethylation assay to demonstrate that mouse UTY is not capable of H3K27 demethylation. Additionally, cell culture data indicate UTX and UTY may function in gene activation as both proteins associate with the H3K4 methyl-transferase complex, the BRG1 chromatin remodeler, as well as heart transcription factors. Our results implicate a crucial H3K27 demethylase independent function for UTX and UTY in mouse embryonic development. This is the first ascribed function for UTY, and the first example of developmental redundancy for X and Y chromosome homologous genes. Notably, our data suggest the H3K27 demethylase activity of UTX is not essential for embryonic viability.
We developed mutant mouse lines to assess the contribution of UTX H3K27 demethylase function in mouse development. Two alleles for Utx were obtained from public resources. The BayGenomics gene trap line Kdm6aGt(RRA094)Byg is designated as XUtxGT1 (Figure 1A). RT-PCR and PCR genotyping verified the identity of this allele in both ES cells and mutant mice (Figures S1 and S2A–S2C). Additionally, we obtained the EUCOMM Kdm6a knockout line (project 26585, Kdm6atm1a(EUCOMM)Wtsi), designated as XUtxGT2fl, which inserts a gene trap in intron 2 along with a floxed 3rd exon (Figure 1A). Southern blotting and PCR genotyping verified the identity of this allele (Figures S1 and S2D–S2F). Notably, quantitative RT-PCR comparison of tail RNA from XUtxGT1 YUty+ versus XUtxGT2fl YUty+ mice demonstrated that Utx gene trap 1 is more effective than gene trap 2 (a 96% reduction compared to a 61% reduction in Figures S2C and S2F). Because XUtxGT2fl demonstrated incomplete trapping, the 3rd exon was deleted with Cre recombinase to establish XUtxGT2Δ (containing both the gene trap and deleted 3rd exon, Figure 1A). Deletion of the third Utx exon produces a frameshift and introduction of a translational stop codon when Utx is spliced from exon 2 to exon 4. XUtxGT1 and XUtxΔ are null alleles as UTX protein was eliminated in western blotting of these embryonic lysates (Figure 1B, 1C). Consistent with RT-PCR data, XUtxGT2fl exhibits a reduction but not absence of UTX protein (Figure 1D).
Heterozygous Utx female mice were crossed to wild type male mice to produce hemizygous Utx mutant males. At weaning, the hemizygous XUtxGT1 YUty+, XUtxGT2Δ YUty+, and XUtxGT2fl YUty+ mice all exhibited reductions of 68%, 83%, and 55% respectively from the expected genotype frequencies based on these crosses, yet expected genotype frequencies were observed at embryonic day E18.5 (Table 1). At E18.5, most of the hemizygous Utx males appeared phenotypically normal; however a small percentage of the fetuses exhibited exencephaly. At birth, the hemizygous Utx males were small and exhibited a failure to thrive phenotype. Those males that survived through this phenocritical phase reached adulthood and were fertile. Hemizygous Utx mutant males were runted compared to wild type littermates and remained smaller than controls throughout their lifespan (Figure 2A, 2B). Backcross of the Utx allele onto a C57BL/6J or 129/SvJ background affected postnatal viability, but hemizygous Utx male embryos were still readily obtained at E18.5 (Table S1).
Human UTY lacks demethylase activity based on in vitro assays, so we hypothesized that XUtx− XUtx− homozygous females will phenocopy XUtx− YUty+ hemizygous males in demethylase dependent function (UTX specific), but may demonstrate a more severe phenotype in demethylase independent roles. Homozygous XUtxGT1 XUtxGT1 and XUtxGT2Δ XUtxGT2Δ females were never observed at weaning or embryonic day E18.5 (Table 1), but were observed at expected genotype frequencies at E10.5. However, these embryos were dead and resorbed by E12.5 (Table 1). Notably, at E10.5 all homozygous XUtxGT1 XUtxGT1 and XUtxGT2Δ XUtxGT2Δ females were smaller in size and had open neural tubes in the midbrain region (Figure 3A-ii, iii, vi, vii). Variation in severity of the Utx homozygous phenotypes was observed in mutant embryos, ranging from medium sized with typical E10.5 features (Figure 3A-ii, vi) to much smaller embryos resembling the E9.5 timepoint (Figure 3A-iii, vii). The XUtxGT1 and XUtxGT2Δ alleles failed to complement, as trans-heterozygous XUtxGT1 XUtxGT2Δ female embryos resembled individual homozygous alleles (Figure 3A-viii). Hemizygous XUtxGT1 YUty+ male embryos appeared phenotypically normal at E10.5 (Figure 3A-iv). Homozygous XUtxGT2fl XUtxGT2fl females exhibited a slight reduction in phenotypic severity; about half of the mutant embryos had open neural tubes and some survival to E12.5 (Table 1). To distinguish between embryonic and extraembryonic contribution of UTX towards the homozygous phenotype, we crossed the Sox2Cre transgene into the Utxfl background. In this cross, paternally inherited Sox2Cre expression will drive Utx deletion specifically in embryonic tissue [40]. No XUtxfl XUtxfl, Sox2Cre female embryos were recovered at E18.5, whereas XUtxfl YUty+, Sox2Cre male embryos were recovered at expected frequencies (Table S2). At E10.5, XUtxfl XUtxfl, Sox2Cre embryos produced phenotypes largely identical to Utx homozygotes. In summary, Utx homozygous females demonstrate a significantly more severe embryonic phenotype in comparison to Utx hemizygous males.
Mid-gestational lethality is typically associated with defective cardiovascular development. Accordingly, we observed both heart and yolk sac vasculature/hematopoietic phenotypes in Utx homozygotes. Utx homozygous mutant hearts were small and underdeveloped, and more severe embryos exhibited peri-cardial edema (Figure 3A-ii, iii, vi, vii). The yolk sac vasculature of Utx homozygotes was pale with a reduction in the amount of vascular blood (Figure 3B-ii). In more severe examples, homozygous yolk sacs were completely pale with an unremodeled vascular plexus (Figure 3B-iii). Thus, abnormal cardiovascular function may be a source of lethality and developmental delay in Utx homozygous mutant embryos.
The most likely explanation for the disparity between Utx hemizygotes and homozygotes is that UTY can compensate for the loss of UTX in embryonic development. We tested Utx and Uty expression in embryonic development to assess any overlap in expression patterns. Utx expression was initially gauged utilizing the B-galactosidase reporter in XUtx+ XUtxGT1 and XUtxGT1 XUtxGT1 whole mount E10.5 embryos. Utx was expressed at lower levels throughout the E10.5 embryo with a particular enrichment in the neural tube and otic placode (Figure S3A-ii, iii, iv). In situ hybridization for both Utx and Uty demonstrated similar expression patterns characterized by widespread low-level expression with particular enrichment in the neural tube (Figure S3B-ii, iii, v, vi). Our analysis of publicly available RNA-seq data sets [41], [42] revealed similar low-levels of expression for Utx and Uty.
To determine whether UTY can compensate for the loss of UTX, we obtained the Welcome Trust Sanger Institute gene trap line UtyGt(XS0378)Wtsi, designated as YUtyGT (Figure 4A). This line, inserted in intron 4, traps the Uty transcript in a similar position of the coding sequence as the Utx alleles (compare to Figure 1A). This gene trap line was verified by RT-PCR in ES cells and subsequent mice (Figures S1 and S2G), and it achieved a 99% reduction in Uty expression from XUtx+ YUtyGT mouse tail RNA (Figure 4B). Hemizygous Uty mutant males, XUtx+ YUtyGT, were viable and fertile (Table 1). However, no compound hemizygous XUtxGT1 YUtyGT and XUtxGT2Δ YUtyGT embryos were recovered at E18.5 (Table 1). At E10.5, expected genotype frequencies of XUtxGT1 YUtyGT and XUtxGT2Δ YUtyGT males were observed, but these embryos phenocopied the developmental delay, neural tube closure, cardiac, and yolk sac defects observed in Utx homozygous embryos (Figure 4C-iii, iv).
We performed a more detailed phenotypic assessment of Utx and Uty mutant hearts to scrutinize the extent of phenotypic overlap between XUtx− YUty+, XUtx− XUtx−, and XUtx− YUty− embryos. Analysis of cardiac development in similar sized E10.5 embryos (Figure 5A-i, ii, iii, iv) revealed that Utx homozygotes and Utx/Uty compound hemizygotes failed to complete heart looping (Figure 5A-vi, viii), whereas Utx heterozygotes and hemizygotes were phenotypically normal (Figure 5A-v, vii). Additionally, homozygotes and compound hemizygotes had smaller hearts with a lack of constriction between the left and right ventricles. Sectioning of E10.5 hearts confirmed that Utx homozygotes and Utx/Uty compound hemizygotes have small hearts with a reduction in ventricular myocardial trabeculation and little or no initiation of interventricular septum formation (Figure 5B-ii, iv). The outer ventricular wall of these embryos is much thinner, and the overall number of cardiomyocytes and myocardial structure is severely deficient (Figure 5C-ii, iv). In summary, while mid-gestational hearts appear normal in XUtx− YUty+ hemizygous males, XUtx− XUtx−homozygous females and XUtx− YUty− compound hemizygous males display identical deficiencies in cardiac development. Therefore, UTY compensates for the loss of UTX in hemizygous Utx mutant males, rescuing mid-gestational cardiac phenotypes.
UTX and UTY have redundant function in embryonic development, but it is not known whether mouse UTY is capable of H3K27 demethylation. Two independent publications demonstrated that human UTY has no catalytic activity in H3K27 demethylation in vitro [26], [29]. It is possible that human UTY (and not mouse UTY) has accumulated a specific polymorphism rendering it demethylase deficient. Additionally, in vitro assays remove UTY from its natural cellular context and may lack co-factors required to promote H3K27 demethylation. Therefore, we utilized an intracellular, in vivo demethylation assay, whereby HEK293T cells transiently over-expressing the UTX carboxy-terminus (encoding the JmjC and surrounding domains essential for proper structure and function) exhibit a reduction in H3K27me3 immunofluorescence levels [43]. In our assay, wild type and mutant constructs were expressed at similar levels (Figure S4A), and individual cells expressing similar, medium-high expression levels of each construct were selected for analysis (Figure 6). Expression of Flag-tagged human and mouse UTX demethylated H3K27me3 and H3K27me2, while a mutation known to disrupt activity (H1146A) was unable to demethylate H3K27 (Figure 6A and 6B, Figure S5A). Human UTX expression had no effect on other histone modifications we tested, such as H3K4me2 (Figure S5B). In contrast, neither human nor mouse UTY were capable of demethylating H3K27me3 and H3K27me2 (Figure 6A and Figure S5A). Cells expressing medium-to-high levels of UTY (N>100) never exhibited a reduction in H3K27me3 levels relative to nearby untransfected controls.
Our previous structural analysis of human UTX [43], combined with sequence alignments (Figure 6C and Figure S6), suggested several amino acid substitutions in human and mouse UTY sequences might make them catalytically inactive. We introduced these mutations into the human UTX C-terminal fragment (Y1135C, T1141I, SNR1025NKS, G1172D/G1191S, I1267P, I1267V, and H1329P), and examined their effects on the in vivo demethylation activity. Of all the mutations tested, only the Y1135C and T1143I mutations completely abolished the ability of UTX to demethylate H3K27 (Figure 6B, 6D). Complete loss of activity was similarly caused by mutations of the corresponding residues in JMJD3 (Y1377C and T1385I, Figure 6D). All qualitative data was also confirmed by immunofluorescence quantification (Figure S4B). Y1135 is conserved throughout all H3K27 demethylases (Figure S7), and in the crystal structure [43], it interacts with two of the three methyl groups of the H3K27me3 side chain, as well as N-oxalylglycine (NOG; an analog of the cofactor alpha-ketoglutarate) (Figure 6E). The smaller C947 side chain of mouse UTY would not effectively maintain either interaction. T1143 is conserved throughout H3K27, H3K9, and H3K36 demethylases (Figure S8), and also interacts with NOG (Figure 6E). Its replacement with bulky isoleucine not only removes the hydroxyl group for interaction with alpha-ketoglutarate, but also may sterically hinder its binding. These observations are consistent with the fact that no H3K27 demethylation activity has been detected for mouse UTY, and we therefore conclude that the catalytic domain of mouse UTY has crucial amino acid replacements that render the protein incapable of H3K27 demethylation. On the other hand, we failed to identify why human UTY is catalytically inactive. Notably, restoring the 2 crucial mouse UTY polymorphisms (M-UTY C947Y, I955T) failed to recover H3K27 demethylase activity (Figure 6B). These data suggest that unidentified structural elements in the UTY C-terminal region are also responsible for the lack of H3K27 demethylase activity.
Although human and mouse UTY have lost the ability to demethylate H3K27, they retain considerable sequence similarity with UTX, suggesting a conserved function. To gain more insight into the overlap in UTX and UTY activities, we performed a biochemical analysis of tagged constructs to determine if UTX and UTY can associate in common protein complexes. Co-transfection of Flag tagged UTX or UTY with HA-UTX followed by immunoprecipitation demonstrates that UTX can form a multimeric complex with itself and UTY (Figure 7A). UTX associates with a H3K4 methyl-transferase complex containing MLL3, MLL4, PTIP, ASH2L, RBBP5, PA-1, and WDR5 [23], [44]. To examine incorporation into this complex, we performed immunoprecipitations with Flag tagged UTX and UTY constructs. Both UTX and UTY were capable of associating with RBBP5 (Figure 7B). Thus, UTX and UTY are incorporated into common protein complexes.
To identify common gene targets of UTX and UTY mediated regulation we generated E10.5 mouse embryonic fibroblast (MEF) cell lines containing mutations in Utx and Uty (alleles XUtxGT2Δ and YUtyGT). The gene traps in these MEFs efficiently trapped Utx and Uty transcripts (Figure 7C). These MEFs did not demonstrate differences in levels of global H3K27me3 (Figure S9A). Genome-wide UTX promoter occupancy has been mapped in fibroblasts [30]. Therefore, we screened our mutant MEFs for misregulated genes affected by the loss of both Utx and Uty that had been documented as direct UTX targets. The FNBP1 promoter is bound by UTX [30]. We verified UTX and UTY binding to the Fnbp1 promoter by ChIP (Figure S9B and S9C). Fnbp1 expression was reduced to 68% of WT levels in XUtx− YUty+ MEFs, but was further compromised to 42% in XUtx− XUtx− lines and 48% in XUtx− YUty− MEFs in which all Utx and Uty activity was lost (Figure 7C). Analysis of E12.5 MEFs of a secondary allele (XUtxGT2fl) also demonstrated diminished Fnbp1 expression in both XUtx− XUtx− and XUtx− YUty− MEFs (Figure 7D). Therefore, Fnbp1 expression is positively regulated by both UTX and UTY.
To examine the role of UTX and UTY in Fnbp1 regulation, we performed H3K27me3 ChIP on E12.5 XUtx+ YUty+ or XUtx− XUtx− MEFs (Figure 7E). Quantitative PCR for an intergenic region served as a negative control, while HoxB1 served as a positive control for H3K27me3. Quantitative PCR demonstrated that the Fnbp1 promoter has relatively low levels of H3K27me3 with no additional accumulation in XUtx− XUtx− MEFs (Figure 7E). Alternatively, H3K4me3 significantly accumulated at the Fnbp1 promoter (Figure 7F). Notably, a loss of Fnbp1 H3K4me3 was observed in XUtx− XUtx− MEFs (Figure 7F). Therefore, UTX and UTY appear to function in Fnbp1 activation by regulating promoter H3K4 methylation rather than H3K27 demethylation.
It has been documented that UTX can associate with heart transcription factors and with the SWI/SNF chromatin remodeler, BRG1 [39]. It has been hypothesized that UTX association with these factors mediates H3K27 demethylase dependent and demethylase independent induction of the cardiomyocyte specification program. As UTX and UTY have redundant demethylase independent function in embryonic development, we examined whether UTY can also associate with these proteins. Co-transfection of Myc-UTY with Flag-BRG1 followed by immunoprecipitation demonstrated that UTY associates with BRG1 (Figure 8A). Myc-UTY also co-immunoprecipitated with Flag-NKX2–5, Flag-TBX5, and Flag-SRF (Figure 8B and Figure S10A). Thus, UTY can form the same protein complexes as UTX with respect to BRG1 and heart transcription factors.
To examine function of UTY in directing activation of downstream heart transcription factor targets, we assessed the regulation of one previously characterized target, atrial natriuretic factor (ANF) [39]. Co-transfection of NKX2–5 with a ANF promoter-Luciferase reporter construct demonstrated a significant upregulation in expression off the ANF promoter (Figure 8C). The reporter expression was significantly enhanced when NKX2–5 was co-transfected with UTY (Figure 8C). The level of ANF reporter transcriptional enhancement was relatively weaker with UTY as compared to UTX, but this is most likely due to a reduction in the transfection efficiency of full-length UTY relative to UTX (as demonstrated in Figure 7A and 7B). UTY also significantly enhanced the ANF reporter response to TBX5 (Figure S10B). Finally, ANF expression was significantly affected in the hearts of only XUtx− XUtx− and XUtx− YUty− embryos (with 52% and 57% level of expression respective to XUtx+ YUty+ controls, Figure 8D). XUtx− YUty+ hemizygotes only had a moderate loss of ANF expression (76% expression level respective to XUtx+ YUty+) that was not statistically significant from wild type controls due to the variability in ANF expression. In summary, both UTX and UTY can associate with heart transcription factors to modulate expression of downstream targets.
We have undertaken a rigorous genetic analysis contrasting UTX and UTY function in mouse embryonic development. In alignment with current literature, Utx homozygous females are lethal in mid-gestation with a block in cardiac development [39]. We now demonstrate that Utx hemizygous mutant males are viable at late embryonic timepoints in expected Mendelian frequencies. In fact, approximately 25% are capable of reaching adulthood. Our comprehensive phenotypic analysis of Utx hemizygous males illustrates that these embryos are phenotypically normal at mid-gestation and lack the cardiovascular dysfunction of Utx homozygous females. This stark phenotypic disparity suggests that UTY may compensate for the loss of UTX in the male embryo. Compound hemizygous Utx/Uty mutant male embryos phenocopy the cardiovascular and gross developmental delay of homozygous females, proving that UTX and UTY have redundant function in embryonic development. As we have demonstrated that mouse UTY lacks H3K27 demethylase activity in vivo, the overlap in embryonic UTX and UTY function is due to H3K27 demethylase independent activity. Given the widespread developmental delay and pleiotropy, it is difficult to assess the primary defect and tissue(s) responsible for UTX and UTY redundancy. The presence of functional UTY in Utx hemizygous males is not capable of preventing peri-natal runting and lethality, suggesting that UTX and UTY are not completely overlapping in activity. These later phenotypes could be due to H3K27 demethylase dependent activity of UTX. Furthermore, the lack of phenotype in Uty hemizygotes demonstrates the absence of any essential UTY specific function in mouse development.
The UTY Jumonji-C domain has maintained high conservation in the absence of catalytic H3K27 demethylase activity. JMJD3 mediated regulation of lymphocyte Th1 response requires an intact Jumonji-C domain, but is also not dependent on H3K27 demethylation [37]. Therefore, this domain may be an essential structural protein component, a protein binding domain, or a domain that may demethylate non-histone substrates. UTX and UTY can associate in a common protein complex and can both interact with RBBP5 of the H3K4 methyl-transferase complex. UTX, UTY, and JMJD3 all associate with H3K4 methyl-transferase complexes from multiple mouse and human cell types [23], [25], [44], [45]. The Fnbp1 promoter is bound by UTX, and gene expression is positively regulated by both UTX and UTY in MEFs. Based on our histone profiling at this locus, UTX and UTY affect the deposition of H3K4 methylation, not H3K27me3 demethylation. Therefore, the common UTX/UTY pathway in embryonic development may involve gene activation rather than removal of gene repression. JMJD3 has been linked more directly to transcriptional activation as the protein complexes with and facilitates factors involved in transcriptional elongation [46]. One cardiac target of UTX regulation, atrial natriuretic factor (ANF), was misregulated in ES cell differentiation [39]. Cell culture experiments suggest that ANF may be a target of both H3K27 demethylase dependent and demethylase independent regulation; however, this study could not distinguish UTX versus UTY function in ES cell differentiation. Both UTX and UTY affect the transcriptional response of an exogenous ANF reporter in the presence of heart specific transcription factors, suggesting that UTX and UTY can operate more directly by aiding in transcriptional activation of this gene rather than altering chromatin structure. Consistently, ANF expression was affected in XUtx− XUtx− and XUtx− YUty− embryonic hearts. UTX and UTY can both associate with the SWI/SNF chromatin remodeler BRG1, which has been hypothesized to mediate histone demethylase independent gene regulation, but the relevance and mechanism of this interaction is not known. Drosophila UTX associates with BRM (orthologous to BRG1) and CBP (a H3K27 acetyl-transferase), and the coupling of H3K27 demethylation with H3K27 acetylation may be essential for switching from a silent to active state [47].
Female cells are subject to gene silencing of one X-chromosome (X-inactivation) to balance gene dosage with males. Theory on establishing X-inactivation for X-Y chromosome homologs hypothesizes that the initial entry step is loss of function or expression of the Y homolog to create dosage imbalance [38]. This prediction also dictates that conservation of X-Y homolog function will maintain gene dosage between sexes, and the female X-homolog will not experience pressure to inactivate. Utx and Uty represent a unique paradox to this untested theory; UTY has lost demethylation activity yet Utx escapes X-inactivation. We now demonstrate that UTX and UTY have retained embryonic redundancy, verifying the presumed correlations between X-inactivation escape and functional dosage balance. Zfx, Sox3, and Amelx represent unbalanced X-chromosome genes; they have similar hemizygous and homozygous mutant phenotypes indicating that the Y chromosome homologs have lost redundant function [48], [49], [50], [51], [52], [53], [54]. Zfx and Sox3 are inactivated, while the Amelx inactivation status is unknown [55], [56]. Of all mouse X and Y chromosome homologs, only Utx, Kdm5c, and Eif2s3x are known to escape X-chromosome inactivation [28], [56], [57], [58]. Interestingly, both KDM5C (SMCX) and its Y chromosome homolog, KDM5D (SMCY) have retained catalytic activity in demethylation of H3K4 di and tri-methyl residues [59], [60], [61], [62]. In contrast to Utx X-chromosome escape driven by demethylation independent redundancy, Kdm5c may escape inactivation due to demethylation dependent redundancy. Our study is the first to demonstrate that an X-Y homologous pair that escapes X inactivation maintains functional conservation, and this escape may stem from an evolutionary benefit to maintain UTY demethylation independent function.
H3K27 demethylases are hypothesized to function in early developmental activation of “bivalent” PRC2 targets by coordinating H3K27 demethylation with H3K4 methylation. The H3K27 demethylation dependent phenotype (UTX specific) of Utx hemizygotes is not apparent until birth. The UTX H3K27 demethylase activity is dispensable for function in C. elegans [63]. Remarkably, the mammalian embryo, having numerous examples of H3K27me3 repression in early development, can survive to term without UTX histone demethylation. It is possible that there is further redundancy between UTX and JMJD3. JMJD3 mutant mice are not well characterized, but have been reported to be peri-natal lethal with distinct features in comparison to Utx hemizygotes [32]. Therefore, it is likely that JMJD3 has distinct targets in development. Overall, the earliest H3K27 demethylation dependent phenotypes for all members of this gene family do not manifest until late embryonic development. This timepoint is much later than the converse early embryonic phenotypes from mutations in the H3K27 methyl-transferase complex [7], . Thus, there appears to be a lack of interplay between H3K27 methylation and demethylation in gene regulation, and the early embryonic removal of H3K27me3 from PRC2 mediated processes (such as ES cell differentiation, reactivation of the inactive X-chromosome, or establishing autosomal imprinting) may involve other mechanisms such as histone turnover or chromatin remodeling. H3K27 demethylases may certainly have crucial roles in the specification of progenitor cell populations of organ systems essential in peri-natal or postnatal viability, and genetic model systems will best assess the functional impact that H3K27 demethylation plays in these processes.
HEK293T were maintained in DMEM supplemented with Glutamine, Pen-Strep, and 10%FBS. Flag-Human UTX (Plasmid #17438) and UTY (Plasmid #17439) were obtained through Addgene [29]. The N-terminus of H-UTX and H-UTY were deleted with QuikChange Lightning (Agilent) as directed producing H-UTX C-terminus 880–1401 (Genbank: NP_066963.2) and H-UTY C-terminus 827–1343 (Genbank: NP_009056.3, an N-terminal His tag was also incorporated into both constructs). Site directed mutagenesis was performed via QuikChange Lightning (Agilent) as directed to produce point mutations. The mouse UTX C-terminus (880–1401) deviated from Human UTX at 2 residues, R1073K and S1263N (According to the Sanger Vega server, the primary Utx transcript Kdm6a-001 encodes for Genbank: CAM27157, we also detected this transcript in E14 ES cell RT-PCR). These changes were created in the H-UTX C-terminus to generate the M-UTX construct. The Flag-tagged mouse UTY C-terminus (692–1212, Genbank: NP_033510.2) was subcloned by RT-PCR of E14 ES cell RNA and introduced into the same vector as the other UTX constructs (PCS2+MT backbone). HA tagged H-UTX was obtained through Addgene [24]. Flag tagged mouse JMJD3 was generously provided by Burgold et al. [27]. Flag tagged BRG1 was obtained through Addgene (Plasmid #19143). Flag tagged NKX2–5 (Plasmid #32969), TBX5 (Plasmid #32968), and SRF (Plasmid #32971) were obtained through Addgene and recombined into DEST26 (Invitrogen). Flag tagged NKX2–5 was also generously provided by Benoit Bruneau [64].
Transfection of HEK293T was accomplished with Lipofectamine 2000 as directed (Invitrogen). Lipid complexes were removed 24 hours post-transfection, and analysis was performed after 48 hours total. Fixation, extraction, and immunofluorescence were performed as described [65]. Immunofluorescence antibodies include anti-Flag (Sigma F3165, 1∶500), anti-H3K27me3 (Millipore 07-449, 1∶500), anti-H3K27me2 (Millipore 07-452, 1∶500), and anti-H3K4me2 (Millipore 07-030, 1∶500). Cells were imaged with Zeiss axiovision software. Image stacks were deconvolved and z-projected. Quantification of H3K27me3 immunofluorescence was performed on deconvolved z-projected stacks (with no pixel saturation in images) using ImageJ software (NIH). The average mean H3K27me3 signal was calculated for untransfected and transfected cells in a given image. For each image, the relative % H3K27me3 was determined, and the average relative % H3K27me3 was calculated for >15 images per construct. For western blotting, nuclear lysates were prepared according to Invitrogen's nuclear extraction protocol. Immunoprecipitations were carried out with 50 µl Flag beads (Sigma A2220) in buffer A as described, using 500 µg (UTY-UTX, RBBP5, BRG1, TBX5, SRF associations) or 1 mg (UTY-NKX2–5 association) of lysate [44]. Immunoprecipitation reactions were boiled off beads and run with 10% input on an 8% SDS-PAGE gel. Histone extractions were prepared as described [66]. Western blotting was performed as described [67] with anti-Flag (Cell Signaling 2368, 1∶4000), anti-RBBP5 (Bethyl Labs A300-109A, 1∶5000), anti-HA (Roche 11867423001, 1∶10000), anti-Myc (Abcam ab9132, 1∶5000), anti-H3K27me3 (Millipore 07-449, 1∶2000), anti-H3 (Millipore 06-755, 1∶5000), and anti-UTX [24].
E10.5 and E12.5 MEFs were generated by removal of the head and interior organs of respective embryos. The remaining body was passed through a 20G needle 6× and plated in DMEM supplemented with Glutamine, Pen-Strep, and 15%FBS. After 3 passages, RNA was isolated with Trizol, and cleaned with an RNeasy kit (Quiagen). RNA from 3 distinct WT XUtx+ YUty+ MEF lines was compared to 3 XUtxGT2Δ XUtxGT2Δ lines on an Illumina bead array (University of Tennessee Health Science Center). All genes significantly decreased in XUtxGT2Δ XUtxGT2Δ MEFs were cross-referenced to the list of UTX bound promoters in human fibroblasts [30]. These genes were analyzed by qRT-PCR (Bio-Rad SsoFast EvaGreen, CFX96 real time system) in all mutant MEF combinations to identify UTY regulated genes. ChIP was performed on these MEFs according to Rahl et al. [68]. MEFs (5×106 cells) were sonicated by a Branson Sonifier at 15% duty cycle (0.7 s on 0.3 s off). ChIP was performed with anti-H3K27me3 (Millipore 07-449, 10 µl), anti-H3K4me3 (Abcam ab1012, 5 µl), anti-Myc (Abcam ab9132, 10 µl), anti-UTX (Santa Cruz H-300, 50 µl), or Rabbit IgG (Sigma, I5006) and qPCR was performed as described above.
We received the ANF promoter-Luciferase reporter construct from Benoit Bruneau [69]. This construct was co-transfected in the presence of NKX2–5 or TBX5 with or without UTX or UTY. Luciferase activity was measured using the Promega Dual Luciferase Reporter Assay System on the Promega Glomax Multi Detection System. All readings were normalized to a Renilla Luciferase control that was co-transfected with all samples.
Kdm6aGt(RRA094)Byg (XUtxGT1), Kdm6atm1a(EUCOMM)Wtsi (XUtxGT2fl), and UtyGt(XS0378)Wtsi (YUtyGT) ES cells were obtained from BayGenomics (through MMRRC), EUCOMM, and SIGTR (through MMRRC) respectively. All ES cells were injected into C57BL/6J host blastocysts for chimera generation. Chimeras were crossed to CD1 to assess germline transmission, and were maintained on either a mixed CD1 background or were backcrossed to 129/SvJ or C57BL/6J. Sox2Cre and RosaFlp transgenes were obtained from The Jackson Laboratory [40]. The VasaCre transgene was developed by Gallardo et al. [70].
All mouse experimental procedures were approved by the University of North Carolina Institutional Animal Care and Use Committee. Utx homozygous data was generated either by crosses between Utx hemizygous males and Utx heterozygous females, or by crosses between XUtxGT2fl YUty+ VasaCre males and XUtx+ XUtxGT2Δ heterozygous females. XUtxGT2fl YUty+ VasaCre males were utilized because of an initial difficulty in generating XUtxGT2Δ YUty+ males and due to the efficient and specific activity of VasaCre in the male germline [70]. Utx hemizygous phenotypic data was developed from the previously mentioned homozygous crosses or through crosses between a WT male and heterozygous Utx female. Compound hemizygous Utx/Uty embryos were generated by crossing heterozygous Utx females with hemizygous Uty males. Embryos were PCR genotyped from yolk sac samples for Utx and were sexed by a PCR genotyping scheme to distinguish Utx from Uty. All primer sequences are available upon request.
Histology samples, in situ hybridization, and LacZ staining were performed as described [71]. In situ hybridization probes were generated to be identical to previous literature [72].
|
10.1371/journal.pntd.0005007 | Predictors of Post-operative Mycetoma Recurrence Using Machine-Learning Algorithms: The Mycetoma Research Center Experience | Post-operative recurrence in mycetoma after adequate medical and surgical treatment is common and a serious problem. It has health, socio-economic and psychological detrimental effects on patients and families. It is with this in mind, we set out to determine the predictors of post-operative recurrence in mycetoma. The study included 1013 patients with Madurella mycetomatis causing eumycetoma who underwent surgical excision at the Mycetoma Research Centre, Khartoum, Sudan in the period 1991–2015. The clinical records of these patients were reviewed and relevant information was collected using a pre-designed data collection sheet. The study showed, 276 patients (27.2%) of the studied population developed post-operative recurrence, 217 were males (78.6%) and 59 were females (21.4%). Their age ranged between 5 to 70 years with a mean of 32 years. The disease duration at presentation ranged between 2 months and 17 years. The majority of the patients 118 (42.8%) had mycetoma of 1 year duration. In this study, students were the most affected; 105 (38%) followed by workers 70 (25.4%), then farmers 48(17.3%). The majority of the patients were from the Central Sudan 207 (75%), Western Sudan 53 (19.2%) while 11 patients (4%) were from the Northern part. Past history of surgical intervention performed elsewhere was reported in 196 patients (71.1%). Family history of mycetoma was reported in 50 patients (18.1%). The foot was the most affected site, 245 (88.7%), followed by the hand seen in 19 (6.8%) patients and 44 (4.5%) had different sites involvement. Most of the patients 258 (93.5%) had wide local surgical excisions while 18 had major amputation. The model predicted that the certain groups have a high risk of recurrence, and these include patients with disease duration greater than 10 years and extra-pedal mycetoma. Patients with disease duration between [5–10] years, with pedal mycetoma, who had previous surgery, with positive family history and underwent wide local surgical excision. Patients with disease duration [5–10] years, with pedal mycetoma, had previous surgery, with no family history but presented with a disease size (> 10 cm), were non- farmers and underwent wide local surgical excision. Other groups are patients with disease duration (≤5 years), with pedal mycetoma, age <59 years, living in the Western /Eastern / Southern regions of the Sudan and with positive family history and had wide local surgical excision. Also included patients with disease duration (≤5 years), with pedal mycetoma, aged <59 years, living in the northern or central region, with no family history but presented with a disease size >10 cm, working as farmers or students and underwent wide local surgical excision. In conclusion, these groups of patients need special care to reduce the incidence of post-operative recurrence with its morbidity and detrimental consequences. In depth studies for the other predisposing factors for post-operative recurrence such as genetic, immunological and environmental factors are needed.
| Post-operative recurrence in mycetoma is a thoughtful problem. It has numerous undesirable medical, health, socio-economic and psychological impacts on the affected patients and their families, communities and health authorities in endemic regions. It is an important motive for patients to drop out follow up and treatment incompliance and hence the inclination of patients for traditional medical treatment. The factors predicating this phenomenon were not studied previously. However, patients’ characteristics and clinical presentation can partially offer an explanation. Thus the present study was set out to understand the predictive ability of some clinical factors on predicting the post-operative recurrence of eumycetoma. The present study had showed young farmers with small sized pedal mycetoma, with short disease duration, who are residing in endemic areas, with no family history and who underwent wide local excision are most likely to remain disease free. We can also concluded that, adequate surgical treatment conditions are obligatory to achieve good outcome and to reduce recurrence. Appropriate health education programmes to encourage early presentation to medical care are essential to reduce the postoperative recurrence rate with its detrimental impacts.
| Eumycetoma is a chronic granulomatous destructive and mutilating subcutaneous fungal infection [1, 2]. The disease is endemic in many tropical and subtropical regions in what is known as the Mycetoma Belt [3, 4]. It has many medical, health, socio-economic detrimental bearings on the affected patients and communities [5, 6]. Currently there are no accurate data on its prevalence and incidence globally, likewise the infection route, susceptibility or resistance [7, 8].
The patient usually presents with small painless subcutaneous mass which gradually increases in size and spreads along the different tissue planes which eventually causes massive damage, destruction and loss of function of the affected body parts [9, 10]. The extremities are affected most but any site can be affected [11, 12].
Effective management of mycetoma depends mainly on accurate diagnosis. This in turn depends on identification of the type of mycetoma and extent of the disease through a meticulous clinical interview, clinical examinations and a battery of investigations. The later includes various imaging techniques, organism identification using grains culture, phenotypic morphological identification, molecular techniques and cyto-histopathological identification [13,14,15,16].
The management usually involves a combination of surgery and prolonged antifungal therapy [17]. The surgical treatment ranges from wide local surgical excision (WLE), repeated debridement and amputation. Early small lesions are amenable to cure with good prognosis. However, the majority of patients present late with advanced disease and such patients are difficult to cure and frequently relapse after apparently adequate treatment with a high morbidly [18, 19, 20].
Post-operative recurrence in mycetoma is a frequent problem and its explanation is an enigma. It has many impacts on the patients and health authorities in endemic areas.
With this background, this study was conducted to understand the clinical predictors of post-operative recurrence of eumycetoma. It also aims to identify the interactions between the different predictive factors in an attempt to develop a predictive model for eumycetoma post-operative recurrence based on the most important clinical factors identified. However, this study is presented with some limitation such as the retrospective and the single-center experience nature.
This retrospective descriptive study was conducted at the Mycetoma Research Centre (MRC), Khartoum, Sudan. It included 1013 patients with confirmed Madurella mycetomatis eumycetoma who underwent surgical treatment in the period 1991 and 2015.
Two hundred seventy sex of these patients (27.2%) developed post-operative recurrence. The diagnosis of eumycetoma was confirmed by clinical interview, meticulous clinical examinations, ultrasound and conventional X-ray examination of the affected part, grains culture, lesion aspirates cytological examination and histopathological examination of the surgical biopsies.
The clinical records of these patients were carefully reviewed and the data was collected using a pre-designed data collection sheet.
The association between clinical factors and mycetoma post-operative recurrence as a target/ outcome variable was investigated. Clinical predictive factors were selected and reformatted using the available domain knowledge provided by expertise in the field of mycetoma. These factors were: age, gender, residence, disease site and duration in years, occupation, family history, previous surgery and type of surgery. These characteristics are shown in Table 1.
In this study, missing data were small and arbitrary (less than 5%). Therefore, the Markov chain Monte Carlo (MCMC) method was used assuming a multivariate normality [21].
In this study, machine learning algorithms; Decision Tree (DT) and Random Forest (RF) were utilized for predicting mycetoma post-operative recurrence assuming unknown data mechanism [22].
Data was partitioned as 70% for training the algorithms with the remaining 30% of the data kept for the validation purpose. Models were trained with 5-fold cross-validation to avoid model over fitting and to ensure model stability [23].
Model performance was evaluated using model accuracy, Positive Predictive Value (PPV), Negative Predictive Value (NPV) and Area under the Receiver Characteristic Curve (AUC) [24]. AUC measures the discrimination ability of the model on predicting the class levels of target variable based on the predictive factors. It is a single scalar value that represents the excepted performance of receiver operating characteristic (ROC) curve.
The study included 1013 patients with Madurella mycetomatis caused eumycetoma patients who underwent surgical treatment at the Mycetoma Research Centre, Khartoum, Sudan in the period 1991–2015. The study documented that, 276 patients (27.2%) developed post-operative recurrence, [Table 1].
The study population were 727 males (71.8%) and 286 females (28.2%). Their age ranged between 5 and 70 years with a median of 23 years. Most of the patients, 625 (61.7%) were in the age group 18–39 years, 199 (18.6%) in the age group 31–59 years and 169 (16.7%) were less than 18 years old at presentation. The disease duration at presentation ranged between 2 months and 17 years. The majority of patients, 839 (82.8%) had mycetoma of less than five years duration, 171 (14.3%) had mycetoma with a duration ranged between 5 to 10 years and 29 (3%) had mycetoma for more than 10 years.
In this study, students were affected most; 398 (39%) followed by farmers 156 (15.4%). The majority of the patients were from the Central Sudan 817 (80.7%), Western Sudan 117 (11.5%) while 48 patients (4.7%) were from the Northern part.
The history of previous surgical treatment performed elsewhere was reported in 566 patients (56%). Family history of mycetoma was reported in 132 patients (13%). The foot 868 (85.7%) was the most affected site followed by the hand seen in 84 patients (8.3) and 61 patients (6%) had mycetoma at different parts.
Most of the patients, 962 (95%) had wide local surgical excisions while 51 patients (5%) had major amputation. The amputation included above and below knee, below elbow and Syem’s amputations.
The study showed that, 276 patients (27.2%) developed post-operative recurrence of whom 217 were males (78.6%) and 59 were females (21.4%). Their age ranged between 5 year and 70 years with a mean of 32 years [Table 1]. The disease duration at presentation ranged between 4 months and 19 years. The majority of the patients 118 (42.8%) had mycetoma of 1 year duration. In this study, students were affected most; 105 (38%), followed by workers 70 (25.4%), then farmers 48 (17.3%). The majority of the patients were from the Central Sudan 207 (75%), Western Sudan 53 (19.2%) while 11 patients (4%) were from the Northern part.
Past history of surgical intervention performed elsewhere was reported in 196 patients (71.1%). Family history of mycetoma was reported in 50 patients (18.1%). The foot was the most affected site 245 (88.7%) followed by the hand seen in 19 (6.8%) patients and 12 (4.5%) had mycetoma at different parts.
258 (93.5%) of the patients had wide local surgical excisions while the rest of patients had major amputation. Amputation included above and below knee, below elbow and Syem’s amputations.
Factors predicating post-operative recurrence were not studied previously. However, patients’ characteristics and clinical presentations can partially offer an explanation. The post-operative recurrence rate in this study was 27.2%, which is disappointing rather high. This post-operative recurrence necessitates repeated surgical excisions leading to massive tissue damage and resulting in healing by fibrosis. This is commonly associated with deformities, disabilities and loss of functions. Many efforts must be undertaken to reduce this high rate and improve the lives of these patients.
Long disease duration proved to be an important recurrence predictor in this study. Mycetoma patients have many unique clinical features and one of these is patients’ late presentation. This is due to the lack of health education, patients’ low socio-economic status and the scarcity of medical and health facilities at remote rural regions where mycetoma is endemic [1]. Thus, good health education is crucial to encourage patients early reporting for medical advice and treatment to increase cure rate and reduces recurrence rate.
The late presentation is commonly associated with massive tissue damage, deformities, destruction and fibrosis and commonly the causative organisms are usually locked in these formed fibrous tissue [25]. All these factors contributes to the incomplete and difficult surgical excisional procedures and thus high recurrence rate.
Another important finding in this study is that, patients with long standing extra-pedal mycetoma are at high risk of post-operative recurrence. Mycetoma surgery requires a bloodless field facilitated by a tourniquet for complete surgical excisions. This is not feasible in extra-pedal mycetoma and may explain the high recurrence. This is supported by the fact that patients with mycetoma in the extremities have higher chance of disease free postoperatively. Most probably the use of tourniquet and a bloodless field in extremities surgery will enable a good and adequate surgical excision. Furthermore, in long-standing, extra-pedal mycetoma, the causative organisms usually spread freely and widely in the different tissue planes thus inducing massive chronic granulomatous inflammatory tissue, fibrosis and cavitation [13]. This process will shelter and lock the organisms and hence the inability of reaching them medically or surgically.
Optimal surgical excision conditions are prerequisite for good outcome. Clinical observations highlighted that, incomplete surgical excision, performed under local anesthesia by inexperienced surgeons in a poor surgical setting in rural areas is an important cause for recurrence [1]. Local excision under local anaesthesia is contraindicated in mycetoma as wide spread disease is common and local anaesthesia will not permit complete excision. No doubt good surgical experience and reasonable equipped facilities are perquisite for good outcome.
Another high-risk group are elderly patients with pedal mycetoma of short disease duration, reside in the Western, Eastern and Southern Sudan and with positive family history of mycetoma. In this group although the disease duration was short, the age may be an important contributory factor in recurrence presumably due to decreased immunity leading to local disease spread and making complete surgical excision not feasible. Furthermore, these patients were not from mycetoma endemic areas and hence not exposed to low grade subclinical infection that boosted their immune system.
The study also showed that, patients with pedal mycetoma of 5–10 years duration with previous surgical excisions and positive family history of mycetoma were likely to develop recurrence. In this group, the previous surgery may had led to spread of the causative organisms along the different tissue planes and had induced more fibrosis and cavitation which make complete excision not feasible.
The size of the lesion at presentation seems to be a good predictor of recurrence in this study. Patients with lesions of more than 10 cm and regardless of other factors were at a high risk to develop recurrence. The wide spread of disease along different tissue planes and bone which make complete surgical excision impossible may be the explanation.
The short disease duration associated with pedal mycetoma in young age group who underwent wide local excision seems to be disease free postoperatively. This can be explained by the fact that, short disease duration is commonly associated with small lesion that can be excised completely in addition to a competent immune system in the young patients. This is logical, as with the long-standing disease the chance for the infection to spread along the different tissue planes, fibrosis, and cavitation are quite common rendering complete surgical excision impossible.
An interesting observation documented in this study is that, farmers with mycetoma lesions larger than 10 cm were more liable to be disease free postoperatively. There is no clear explanation for this but perhaps their occupation exposes them to low grade subclinical infection due to repeated exposure to the causative organisms thus enhancing their immune responses.
There is a strong association between the absence of family history of mycetoma and disease free-state observed among the studied patients. No clear explanation can be postulated but patients with family history of mycetoma may be genetically prone to develop the infection and recurrence and that make disease eradication difficult however, further studies are needed to confirm that. However, the whole family members may be sharing the same environment conditions suitable to acquire the infection and recurrence.
In conclusion, young farmers and students with pedal mycetoma of small size, with short disease duration, who were residing in endemic areas, with no family history and who underwent wide local excision were most likely to remain disease free. Adequate surgical treatment conditions are obligatory to achieve good outcome. Appropriate health education programmes to encourage early presentation to medical care are essential to reduce the postoperative recurrence rate with its detrimental impacts. In depth studies for the other causes for post-operative recurrence such as genetic, immunological and environmental factors are needed.
Ethical clearance was obtained from Soba Hospital Ethical Committee. Patients’ informed consents proved to be unnecessary in this study.
|
10.1371/journal.pbio.1001968 | The Chromatin Assembly Factor 1 Promotes Rad51-Dependent Template Switches at Replication Forks by Counteracting D-Loop Disassembly by the RecQ-Type Helicase Rqh1 | At blocked replication forks, homologous recombination mediates the nascent strands to switch template in order to ensure replication restart, but faulty template switches underlie genome rearrangements in cancer cells and genomic disorders. Recombination occurs within DNA packaged into chromatin that must first be relaxed and then restored when recombination is completed. The chromatin assembly factor 1, CAF-1, is a histone H3-H4 chaperone involved in DNA synthesis-coupled chromatin assembly during DNA replication and DNA repair. We reveal a novel chromatin factor-dependent step during replication-coupled DNA repair: Fission yeast CAF-1 promotes Rad51-dependent template switches at replication forks, independently of the postreplication repair pathway. We used a physical assay that allows the analysis of the individual steps of template switch, from the recruitment of recombination factors to the formation of joint molecules, combined with a quantitative measure of the resulting rearrangements. We reveal functional and physical interplays between CAF-1 and the RecQ-helicase Rqh1, the BLM homologue, mutations in which cause Bloom's syndrome, a human disease associating genome instability with cancer predisposition. We establish that CAF-1 promotes template switch by counteracting D-loop disassembly by Rqh1. Consequently, the likelihood of faulty template switches is controlled by antagonistic activities of CAF-1 and Rqh1 in the stability of the D-loop. D-loop stabilization requires the ability of CAF-1 to interact with PCNA and is thus linked to the DNA synthesis step. We propose that CAF-1 plays a regulatory role during template switch by assembling chromatin on the D-loop and thereby impacting the resolution of the D-loop.
| Obstacles to the progression of DNA replication forks can result in genome rearrangements that are often observed in cancer cells and genomic disorders. Homologous recombination is a mechanism of restarting stalled replication fork that involves synthesis of the new DNA strands switching templates to a second (allelic) copy of the DNA sequence. However, the new strands can also occasionally recombine with nonallelic repeats (distinct regions of the genome that resemble the correct one) and thereby cause the inappropriate fusion of normally distant DNA segments; this is known as faulty template switching. The chromatin assembly factor 1 (CAF-1) is already known to be involved in depositing nucleosomes on DNA during DNA replication and repair. We have found that CAF-1 is also involved in the recombination-mediated template switch pathway in response to replication stress. Using both genetic and physical assays that allow the different steps of template switch to be analyzed, we reveal that CAF-1 protects recombination intermediates from disassembly by the RecQ-type helicase Rqh1, the homologue of BLM (people with mutations that affect BLM have Bloom's syndrome, an inherited predisposition to genome instability and cancer). Consequently, the likelihood of faulty template switch is controlled by the antagonistic activities of CAF-1 and Rqh1. We thus identified an evolutionarily conserved interplay between CAF-1 and RecQ-type helicases that helps to maintain genome stability in the face of replication stress.
| The maintenance of genome stability requires a complex network to coordinate multiple pathways, including DNA replication, repair, and recombination in a chromatin context. Replication stress, including obstacles to replication fork progression, has emerged as a major source of genome instability that fuels cancer development and underlies chromosome modifications observed in genomic disorders [1]. Deciphering the control of repair pathways occurring at replication forks remains of crucial importance to understanding of the mechanisms underlying genome rearrangements.
Homologous recombination (HR) is an evolutionarily conserved mechanism that promotes DNA repair and contributes to accurate and complete DNA replication [2]. When fork progression is disrupted by DNA damage or a fork obstacle, HR mediates the nascent strands to switch templates to resume DNA synthesis. Template switch occurs either at the three-way branched junction of the fork to restart it or between sister-chromatids to fill in single-stranded DNA (ssDNA) gaps left behind the moving fork [3],[4]. This last pathway is referred to as error-free postreplication repair (PRR) [5].
Faulty replication restart events are one of the causal mechanisms of genome instability. When control of allelic recombination fails, nascent strands at a blocked replication fork can recombine with a nonallelic homologous repeat and initiate DNA synthesis on a noncontiguous template, thus resulting in the fusion of noncontiguous DNA segments and genome rearrangements [6]–[8]. This mechanism is referred to as faulty template switch and is proposed to drive genome rearrangements in cancers cells (e.g., chromothripsis) and in genomic disorders (e.g., complex rearrangements such as triplication-associated inversions) [1],[9]. Faulty template switch between homologous repeats is reminiscent of nonallelic HR (NAHR). In yeast, inverted repeats are particularly prone to faulty template switching [10]. Recently, both HR and error-free PRR have been reported as mechanisms of faulty replication leading to fusion of inverted repeats in human cells [11]. Thus, the mechanisms of faulty template switch appear evolutionarily conserved.
HR has been extensively studied in the context of double-strand break (DSB) repair, but only a few studies have addressed the mechanisms of template switch [2],[4],[12]. With the assistance of HR mediators (such as Rad52 in yeast), the recombinase Rad51 nucleates onto ssDNA covered by RPA to form a nucleoprotein filament. After the search for homology, the Rad51 filament invades a homologous DNA duplex to pair the invading ssDNA with the complementary strand, whereas the noncomplementary strand is displaced. The resulted three-stranded intermediate is a type of joint molecule (JM) called a displacement loop (D-loop) in which the 3′ end of the invading strand primes DNA synthesis. At replication forks, extension of the D-loop by DNA synthesis might permit the restoration of a functional replisome, thus ensuring the completion of DNA replication [13]. In the context of DSB repair, the capture of the second DNA end results in the formation of a later JM called a double Holliday junction (dHJ) whose resolution by cleavage leads to crossover (CO) formation, a source of chromosome rearrangements associated with NAHR [1],[12],[14]. Several DNA helicases/translocases have been shown to be involved in the prevention of mitotic CO by preventing D-loop formation or its disassembly—among them, Srs2, FANCM, and RecQ-type helicases [15]–[23]. Whether the outcomes of template switch at replication forks are also regulated by helicase-dependent D-loop dismantling is unknown.
HR occurs within DNA packaged into chromatin that needs to be disassembled and then restored after the recombination event is completed [24]. Chromatin remodeling factors help in relaxing chromatin and in providing access to DNA damage signaling and repair machineries at damaged sites, but how chromatin restoration is coupled to HR remains poorly understood. The chromatin assembly factor 1, CAF-1, is a histone H3-H4 chaperone that promotes DNA synthesis-coupled chromatin assembly during DNA repair and DNA replication [25]–[28]. CAF-1 is a three-subunit complex conserved throughout evolution, and the three CAF-1 subunits in Schizosaccharomyces pombe are called Pcf1 (SPBC29A10.03c), Pcf2 (SPAC26H5.03), and Pcf3 (SPAC25H1.06), which correspond, respectively, to the p150, p60, and p48 in mammalian cells [29]. The large subunit of CAF-1, p150, interacts with PCNA, thus targeting CAF-1 to DNA synthesis sites at which CAF-1 and Asf1 (anti-silencing factor 1) cooperatively assemble chromatin onto newly synthesized DNA in a PCNA-dependent manner [30]–[37].
In response to DNA damage, the large subunit of CAF-1 and the heterochromatin factors HP1 (heterochromatin protein 1) are targeted to mammalian HR foci, within which they promote the resection of DSBs and thus the recruitment of HR factors such as Rad51 [38]–[41]. After completion of DNA repair, CAF-1 and Asf1 restore nucleosomal organization at DNA damage [26]–[28],[30],[42]–[44]. In budding yeast, CAF-1 and Asf1 are dispensable for DSB repair by HR but necessary for the restoration of the chromatin state, a step required to turn off checkpoint activation [45],[46]. It is suggested that CAF-1 primes DSB repair and then switches to an active histone chaperone mode to restore chromatin at DNA damage [24]. Whether CAF-1 regulates other HR pathways such as template switch and whether it impacts repair fidelity is unknown.
Here, we identified that fission yeast CAF-1 acts in a HR pathway alternative to PRR when cells replicate a damaged template. We revealed functional and physical interactions between CAF-1 and the RecQ-type helicase Rqh1, the fission yeast BLM homologue. Using a conditional replication fork obstacle, we report a novel chromatin factor-dependent step during HR-mediated template switch: CAF-1 counteracts the disassembly of D-loop intermediates by Rqh1. As a consequence, the likelihood of faulty template switch is controlled by the antagonistic roles of CAF-1 and Rqh1 in D-loop stability. The protection of the D-loop requires the three CAF-1 subunits and its ability to interact with PCNA, showing that CAF-1 stabilizes the D-loop at the DNA synthesis step. Thus, CAF-1 and Rqh1 act coordinately to maintain genome stability in response to replication stress. We propose that CAF-1 plays a regulatory role during template switch by assembling chromatin on the D-loop and thereby impacting its resolution.
We wanted to ascertain whether CAF-1 was involved in replication-coupled DNA repair. We focused on cell resistance to the alkylating agent methyl-methane sulfonate (MMS) that creates DNA lesions blocking fork elongation and known to induce template switch events [4]. The deletion of pcf1 (pcf1-d, SPBC29A10.03c) did not affect cell sensitivity to MMS compared to wild-type (wt) cells. However, combined with genetic backgrounds in which error-prone PRR (the bypass of DNA lesions by translesion synthesis, rev1-d, SPBC1347.01c) or error-free PRR (rad8-d, SPAC13G6.01c, Rad8 being the homologue of budding yeast Rad5) were defective, pcf1-d resulted in an increased cell sensitivity to MMS, compared to each single mutant (Figure 1). A similar genetic interaction was observed with Srs2 (SPAC4H3.05), a helicase involved in error-free PRR [47]. These data suggest that CAF-1 acts in a replication-coupled DNA repair pathway but independently of the error-prone and error-free branches of PRR. We then asked whether CAF-1 could act in the Rad51 (SPAC644.14c)-dependent HR pathway. We found that the double mutant pcf1-d rad51-d exhibited only a modest increased sensitivity to MMS compared to the single mutant rad51-d (Figure 1). These data indicate that CAF-1 may operate in the Rad51-dependent replication-coupled DNA repair pathway but may not function entirely through the HR pathway.
To decipher the role of CAF-1 in replication-coupled HR, we made use of a polar replication fork barrier (RFB), which is genetically encoded by a DNA sequence called RTS1 bound by the protein Rtf1 whose expression is regulated at the transcriptional level via the use of the nmt41 promoter. In the presence of thiamine in the media, Rtf1 is not expressed and the RTS1-RFB is not active; after at least 16 h following thiamine removal, Rtf1 is expressed and the RTS1-RFB is induced [48]. In the t-ura4 <ori construct, a single RTS1-RFB is located near an efficient replication origin (ori 3006/7, on chromosome III) to allow the block of forks emanating from this origin and moving toward the telomere, the main replication direction of the ura4 locus (Figure S1A). Blocked replication forks are restarted by HR and independently of DSBs [3],[7].
Perturbed replication-coupled chromatin assembly, due to CAF-1 and Asf1 deficiency, leads to a higher susceptibility of replication forks to collapse and thus an increased level of genome instability in budding yeast [49]–[52]. We thus asked whether a defect in CAF-1 affects the activity of the RTS1-RFB and the early steps of HR at replication forks. At the t-ura4 <ori locus, which contains a single fork barrier, the analysis of replication intermediates (RIs) by bidimensional gel electrophoresis (2DGE) showed that the RTS1-RFB was as efficient in the absence of CAF-1 (i.e., in either pcf1-d, pcf2-d, or pcf3-d null mutant) as in the wt strain (Figure S1B–C). Also, Rad52, the main HR factor (SPAC30D11.10), was recruited to the fork barrier in the absence of CAF-1 to the same extent as in the wt strain (Figure S1D). Our data indicate that the RTS1-RFB was functional and prone to recruit HR factors in the absence of CAF-1.
We then made use of another construct, t> ura4 <ori, that contains two RTS1 sequences integrated at both sides of ura4 (Figure 2A) [48]. A third RTS1 sequence is present at its natural location, near the mat locus on the chromosome II. Given the orientation of the RTS1 sequences relative to the main replication direction of each locus, the RTS1 sequence at the centromere (cen)-proximal side of ura4 behaves as a strong RFB, whereas the two other RTS1 sequences have poor RFB activity. Occasionally during replication restart (in ∼2%–3% of the cell population/generation), HR-mediated template switch results in nascent strands inappropriately invading the RTS1 sequence located in the vicinity of the blocked fork on chromosome III or the one located further away on chromosome II. Such faulty template switches lead to chromosomal rearrangements including inversions and large palindromic chromosomes, as well as the loss of the ura4 marker (Figures 2A and S2A) [3],[48],[53]. Importantly, chromosomal rearrangements were also observed in response to MMS treatment and in the absence of active RTS1-RFB, showing that conditional fork barriers are relevant models for the generation of rearrangements initiated by template switch [7]. Also, replication restart and template switch-mediated rearrangements occur independently of PRR, making the RTS1-RFB assay particularly useful to decipher the role of CAF-1 in replication-coupled HR [7],[53].
In the t> ura4 <ori strain, fork arrest at the cen-proximal side of ura4 leads to an 11-fold increase in the loss of the ura4 marker (Table 1 and [53]). The loss of ura4 corresponds to the deletion of ura4 on chromosome III (genomic deletion) or translocation to chromosome II; each event can be distinguished by PCR (Figure 2A–B). Genomic deletion and translocation result from HR-mediated faulty template switches between the three dispersed RTS1 sequences (Figure 2A) [53]. Consistent with this, both fork-arrest–induced genomic deletion and translocation are dependent on Rad52 (Table 1, Figure 2B, and [53]). The rad52-d mutant experiences a loss of viability upon induction of the RTS1-RFB, a phenotype not observed in strains deficient for individual CAF-1 subunits (Figure 2C). However, we observed that a defect in CAF-1 leads to a 3- to 5-fold reduction in the rate of fork-arrest–induced ura4 loss, compared to the wt strain (Table 1). PCR analysis showed that both genomic deletion and translocation induced by the active RTS1-RFB were affected: In the pcf1-d mutant, these events were reduced by 6- and 10-fold, respectively, compared to the wt strain (p<0.0001) (Figure 2B–D). Our data indicate that, surprisingly, faulty template switch at blocked forks requires CAF-1. As Rad52 is efficiently recruited to the RTS1-RFB in the absence of CAF-1 (Figure S1D), this suggests that CAF-1 promotes template switch downstream of the recruitment of HR factors.
In the t> ura4 <ori strain, fork arrest at the cen-proximal side of ura4 induces stalled nascent strands to switch template and to recombine with the opposite RTS1 sequence located on the telomere-proximal side of ura4 (Figure 3A) [3]. This template switch event leads to a stable and early JM (JM-A), which consists of a D-loop structure. Then, the approaching opposite fork is stalled by the RTS1-RFB, leading to a second template exchange and the formation of a later JM (JM-B), which contains HJ-like structures (Figure 3A–B). Thus, the D-loop structure is the precursor of the HJ-like intermediate. Both types of JMs are detectable by 2DGE and are dependent on Rad52 [3]. The resolution of HJ-like structures leads to chromosomal rearrangements: acentric and dicentric isochromosomes or chromosomes in which ura4 has switched orientation (Figures 3A and S2A). Chromosomal rearrangements can be detected by pulse field gel electrophoresis (PFGE) and restriction fragment length analysis (RFLA), followed by Southern blotting [3].
We further investigated the role of CAF-1 during template switch by these physical assays. The intensity of both types of JMs was severely decreased in strains deficient for individual CAF-1 subunits (Figure 3B–C). In addition, all types of chromosomal rearrangements, the products of resolution of HJ-like structures, were reduced by 2- to 3-fold in CAF-1 defective strains (p<0.003) (Figures 3D–E and S2B–C). Thus, the decreased intensity of HJ-like structures could not be explained by a faster cleavage of these structures. Because CAF-1 does not prevent HR factor recruitment at blocked forks, we rather envisioned that D-loop intermediates are formed but dismantled more quickly in the absence of CAF-1, thus resulting in a decreased level of HJ-like intermediates. To test this hypothesis, we analyzed genetic interactions with mus81 (SPCC4G3.05c). Fission yeast Mus81 is an endonuclease involved in the cleavage of HJs [54]. As previously reported, HJ-like structures, but not D-loop intermediates, accumulated in mus81-d cells, thus resulting in cell death upon induction of the RTS1-RFB (Figure S3) [3]. Consistent with a faster dismantling of the D-loop and less HJ-like structures being produced in the absence of CAF-1, the deletion of pcf1 rescued the sensitivity of mus81-d cells to the induction of the RTS1-RFB (Figure S3A–B). Analysis of JMs by 2DGE confirmed that the intensity of both JMs remained decreased in the double mutant compared to the single mutant mus81-d (Figure S3C–D). The data are consistent with the hypothesis that HJ-like structures are formed less often in the absence of CAF-1 due to faster dismantling of D-loop intermediates.
To reinforce this last hypothesis, we investigated genetic interaction with the recombinase rad51 required to promote D-loop formation. In the absence of Rad51, chromosome rearrangements are produced without D-loop formation, probably via the single strand annealing function of Rad52 [3]. We reasoned that if CAF-1 stabilizes Rad51-dependent D-loop intermediates, its function in promoting template switch should rely on a functional Rad51 pathway. The type and level of chromosome rearrangements observed in the double mutant pcf1-d rad51-d was similar to those of the single rad51-d mutant, showing that rad51 and pcf1 are epistatic (Figures 4A and S4A–B). The data are consistent with CAF-1 acting in the Rad51 pathway to promote HR at replication forks, likely downstream of the formation of D-loop intermediates.
To extend our conclusion of CAF-1 acting in the Rad51 pathway to prevent D-loop disassembly, we performed genetic analysis. In both fission and budding yeast models, the concomitant inactivation of a RecQ-helicase and Srs2 results in a pronounced slow growth phenotype or cell death, a phenotype rescued by the deletion of rad51 [47],[55],[56]. It has been proposed that this synthetic sickness/lethality results from the accumulation of unresolved JMs that impinge on cell fitness. Deleting either pcf1 or pcf2 led to a marked rescue of the slow growth phenotype of the rqh1-d srs2-d strain, although to a less extent than the deletion of rad51 (Figure 4B–C). These data are consistent with CAF-1 acting in the Rad51 pathway and promoting template switch at replication forks by stabilizing D-loop intermediates.
We investigated the mechanism by which CAF-1 prevents D-loop dismantling. Several helicases have been implicated in D-loop dissociation including Srs2, Fml1, and Rqh1 [15]–[23]. We found no evidence of Fml1 (SPAC9.05) promoting template switch at the site-specific arrested fork (unpublished data). Given the synergistic sensitivity of the double mutant pcf1-d srs2-d to MMS, we first analyzed the interactions between CAF-1 and Srs2 using our model system for template switch. We previously reported that Srs2 promotes JM formation and chromosomal rearrangements formed by template switch [3]. We found that strains defective for both CAF-1 and Srs2 showed a reduced level of chromosome rearrangements, similar to those observed in each single mutant (Figure S4). Thus, CAF-1 and Srs2 might act in the same pathway promoting template switch at replication forks.
The human RecQ helicase BLM and the large subunit of CAF-1 (p150) physically interact to coordinately promote cell survival to replication stress [57]. We found that the double mutant pcf1-d rqh1-d was more sensitive to MMS than the single rqh1-d mutant, rqh1 (SPAC2G11.12) being the fission yeast homologue of BLM (Figure 1). Also, pcf1-d rqh1-d was more sensitive to camptothecin (CPT, a topoisomerase 1 inhibitor) than each single mutant, whereas the deletion of pcf1 suppressed the sensitivity of the single mutant rqh1-d to hydroxyurea (HU), a ribonucleotide reductase inhibitor that depletes dNTP pools and stalls replication forks (Figure S5). Co-immunoprecipitation experiments showed that Rqh1 and Pcf1 physically interact (Figures 5A and S6A). Thus, functional interactions between CAF-1 and Rqh1 to promote cell resistance to replication stress are evolutionarily conserved. RecQ helicases prevent genome instability by promoting the dissolution of early (D-loop) and late (double HJs) JMs [54]. We previously have proposed that Rqh1 limits genome instability at replication forks by disassembling Rad51-dependent D-loops [3]. In the RTS1-RFB assay, HJs formed between RTS1 repeats cannot branch migrate in vitro and thus cannot be resolved by dissolution. Accordingly, HJ-like intermediates did not accumulate in rqh1-d cells compared to wt cells (Figure 5B–C, panels 7 and 10) [3]. We analyzed whether Rqh1 could be responsible for D-loop dismantling in the absence of CAF-1. In a pcf1-d rqh1-d and in a pcf2-d rqh1-d strain, the level of both JMs was restored to those observed in either rqh1-d or wt cells (Figure 5B–C). To verify that the stability of JMs are restored in vivo and does not result from an in vitro artifact during DNA manipulation, DNA samples were cross-linked prior to extraction. In such conditions, the lack of JMs in the absence of CAF-1 was confirmed (Figure 5B, panel 5), showing that JMs are unstable in vivo. Also, the intensity of JMs was restored to a wt level by deleting rqh1 (Figure 5B, panels 6 and 9), showing that Rqh1 is responsible for the lack of JMs in the absence of CAF-1 in vivo. Consistently, both ura4 inversion and acentric chromosomes, the resolution products of HJ-like structures, were restored to wt levels in strains defective for CAF-1 and Rqh1 (Figures 5D–E and S6B). Our data establish that Rqh1 disassembles the D-loop in the absence of CAF-1, and we propose that CAF-1 promotes template switch at the replication fork by counteracting D-loop disassembly by Rqh1.
We analyzed the level of genomic deletion and translocation that result from faulty template switch between RTS1 sequences on chromosomes II and III [53]. Following induction of the RTS1-RFB, deleting pcf1 resulted in a 6 and 10 times reduction in the rate of genomic deletion and translocation, respectively, compared to the wt strain (Figure S6C–D). In contrast, the rate of these events was reduced by only ∼1.8 times by deleting pcf1 in the absence of Rqh1 (Table 1 and compare rqh1-d and pcf1-d rqh1-d strains on Figure S6C). PCR analysis showed that translocation and deletion events were decreased in pcf1-d cells compared to wt cells (Figure 2B), but not when rqh1 is deleted (Figure S6D). Altogether our data reveal that the likelihood of faulty template switch is controlled by the antagonistic roles of CAF-1 and Rqh1 in processing the D-loop.
CAF-1 mediates replication-coupled chromatin assembly and interacts with the heterochromatin factor Swi6 (SPAC664.01c, the human HP1 homologue) to assist the maintenance of heterochromatin and silencing during S-phase [29]. A strain mutated for swi6 exhibited no defect in the accumulation of the acentric chromosome following the activation of the RTS1-RFB, indicating that the role of CAF-1 in template switch is unlikely to involve heterochromatin (Figure S7).
Deposition of histone H3-H4 onto newly synthesized DNA by CAF-1 requires, in vitro, its three subunits and its ability to interact with the replication factor PCNA (SPBC16D10.09) [28],[31]–[33],[35],[37],[44]. The three strains pcf1-d, pcf2-d, and pcf3-d exhibited a similar phenotype: fewer faulty template switches and a faster dismantling of the D-loop (Figures 2 and 3). A strain in which the three subunits of CAF-1 have been inactivated showed a decreased level of acentric chromosomes, one of the products of JM resolution, similar to those observed in each single mutant (Figure 3D–E). Thus, the role of CAF-1 in promoting template switch is not specific to a single subunit but necessitates the three subunits to act in the same HR pathway.
In budding and fission yeast, the large subunit of CAF-1 contains only one canonical PCNA interacting peptide (PIP box). We mutated the key residues to alanine to generate a mutant of pcf1 unable to interact with PCNA (pcf1-PIPmut) (Figure 6A). Co-immunoprecipitation showed that mutating the PIP box of Pcf1 severely impaired the interaction of Pcf1 with PCNA without affecting its interaction with Pcf2 (Figure S8A). The interaction of Pcf2 with PCNA was also dependent on the PIP box of Pcf1 (Figure S8A). Thus, expressing Pcf1-PIPmut leads to the formation of a CAF-1 complex unable to interact with PCNA. As expected, mutating the PIP box of Pcf1 led to a loss of Pcf1 foci in S-phase cells and a loss of co-localization with replication factories (labeled with a CFP-tagged version of PCNA) (Figure S8B). Thus, the canonical PIP box of Pcf1 is sufficient to target CAF-1 into replication foci and expressing Pcf1-PIPmut is likely to impair replication-coupled chromatin assembly by CAF-1. Then, we investigated the phenotype of the pcf1-PIPmut strain. First, the stability of JMs was impaired and consistently the level of the acentric chromosome (one of the products of JM resolution) was reduced as in the pcf1-d strain (Figure 6B,C,D). Second, the rates of genomic deletion and translocation induced by the active RTS1-RFB were similarly decreased in pcf1-PIPmut and pcf1-d cells, compared to the wt strain (Table 1, Figure 6E–F). Thus, mutating the PIP box of Pcf1 is sufficient to mimic the deletion of Pcf1. Thus, the role of CAF-1 in promoting template switch by preventing Rqh1-dependent dismantling of the D-loop requires the full complex and the capacity to interact with PCNA.
We have discovered a novel chromatin-factor–dependent step during HR-mediated template switch, involving CAF-1: the protection of the D-loop from disassembly by the RecQ-type helicase Rqh1. First, CAF-1 promotes an HR-dependent and replication-coupled repair pathway, independently of the error-prone and error-free branch of PRR. Second, using a genetic assay that selects for recombination events at replication forks, we establish that CAF-1 promotes template switch by counteracting D-loop disassembly by Rqh1. Third, Rqh1 and CAF-1 physically interact. Consequently, the likelihood of faulty template switch is controlled by the opposite activities of CAF-1 and Rqh1 in processing the D-loop. Finally, the D-loop protection by CAF-1 requires the full complex and its interaction with PCNA, but not the heterochromatin factor Swi6. Our data are thus consistent with a model in which D-loop extension by DNA synthesis is coupled to histone deposition by CAF-1. We propose that the newly assembled nucleosomes on the D-loop display a substrate less favorable to antirecombinase Rqh1 action, thus protecting the D-loop from being dismantled (Figure 6G, black line). This pathway does not exclude other mechanisms such as a negative interference of CAF-1 with Rqh1 activity either directly or via an additional nonhistone factor. This second mechanism would also require PCNA-mediated localization of the CAF-1 complex during D-loop extension (Figure 6G, dashed green line).
Beyond its role in chromatin restoration at DNA damage sites, roles for CAF-1 in recombinational DNA repair pathways have been reported [24],[45],[46]. Budding yeast CAF-1 protects against DSBs by acting both in HR and nonhomologous end-joining pathways [58],[59]. A defect in CAF-1 also leads to a decreased efficiency of DSB-induced recombinational repair in drosophila [60]. More recently, a genetic screen has identified CAF-1 as promoting break-induced replication, a one-ended invasion HR pathway that occurs when the homology between the broken end and the donor DNA molecules is limited to one broken arm [61]. In mammals, CAF-1 acts in both the early and late steps of HR-mediated DNA repair by promoting the resection of DSBs and the recruitment of HR factors and then the restoration of chromatin state when repair is completed [24].
Here, we report that CAF-1 promotes replication-coupled DNA repair independently of the error-prone and error-free branches of PRR. The sealing of ssDNA gaps left behind moving forks involves template switches mediated by the error-free branch of PRR [4],[5]. This damage tolerance pathway requires Rad5, Rad51, and the ubiquitination of PCNA. Our data place CAF-1 in an alternative Rad51-dependent template switch pathway. Consistent with this, replication restart and chromosome rearrangements mediated by template switch at site-specific arrested forks occur independently of the ubiquitination of PCNA [7],[53]. We propose that CAF-1 acts in Rad51-dependent template switches occurring during replication restart. We identified the underlying mechanism: CAF-1 stabilizes the D-loop by preventing its disassembly by the helicase Rqh1. Consequently, the likelihood of faulty template switch, a type of NAHR causing chromosomal rearrangements, is controlled by the antagonistic activities of CAF-1 and Rqh1 at the D-loop: CAF-1 stabilizing the D-loop and Rqh1 promoting its disassembly. Functional interplays between CAF-1 and Rqh1 in response to replication stress are evolutionarily conserved (see below).
In mammals, CAF-1 primes HR events at DNA damage by promoting the end-resection of DSBs and thus the recruitment of HR factors such as Rad51 [24]. Then, CAF-1 might switch towards its histone chaperone mode to restore chromatin after the completion of the HR event. Here, we report a novel step at which CAF-1 promotes HR: By preventing D-loop disassembly, CAF-1 impacts the resolution of the subsequent HR event. Thus, the role of CAF-1 during HR might be more dynamic than previously anticipated, not only acting in the early and final steps, but having potential roles all along the HR process. We propose that CAF-1, and potentially chromatin assembly coupled to the DNA synthesis step of the HR event, is an important regulatory point of template switch.
Defects in the RecQ-type helicase BLM lead to Bloom's syndrome, a human disorder associating genomic instability and cancer predisposition. Functional interactions between BLM and the p150 large subunit of CAF-1 have been previously reported in response to replication stress [57]. Here, we identified that interplays between CAF-1 and BLM are evolutionarily conserved in fission yeast. The large subunit of CAF-1, Pcf1, and Rqh1 physically interact and act in a coordinated way to promote survival and maintain genome stability in response to replication stress. Importantly, we uncovered the underlying mechanism. Using genetic and physical assays that allow the analysis of the individual steps of HR-mediated template switch at a single replication fork, we found that the impaired stability of D-loop intermediates due to a CAF-1 defect results from the activity of Rqh1. CAF-1 thus counteracts D-loop dismantling by Rqh1. The RecQ helicase family is also involved in the rescue and stability of stalled forks, however we excluded interplays between CAF-1 and Rqh1 in this process [62]–[64]. First, the site-specific arrested fork is stable and prone to recombination events in both single and double mutants. Second, CAF-1 acts downstream of D-loop formation by Rad51.
We hypothesize that nucleosome assembly on the D-loop is promoted by the interaction of CAF-1 with PCNA and that the nucleosomal nature of the D-loop prevents disassembly by Rqh1. We cannot exclude that the interaction with PCNA simply serves to recruit CAF-1 to the D-loop where CAF-1 could either directly counteract Rqh1 action or trigger the recruitment of an additional factor counteracting Rqh1 activity (Figure 6G). In human cells, BLM inhibits CAF-1–mediated chromatin assembly coupled to DNA repair [57]. Through physical interactions, Rqh1 could also mediate CAF-1 recruitment to the D-loop on which Rqh1 could inhibit chromatin assembly by CAF-1. However, such hypotheses are not sufficient to account for all our observations. Indeed, in the absence of CAF-1 and of any potential histone deposition on JMs, the D-loop is disassembled faster by Rqh1, thus rather suggesting a model in which CAF-1 counteracts Rqh1 activity.
Preventing Rqh1-dependent D-loop dismantling requires the three subunits of CAF-1 and its interaction with PCNA as for optimal histone deposition onto newly replicated DNA in vitro [25],[31],[37]. Therefore, we propose that CAF-1 prevents D-loop disassembly by promoting histone deposition onto the D-loop. We could not confirm this hypothesis by generating CAF-1 mutated forms unable to interact with histones, as CAF-1 binds histone H3-H4 by multiple interactions: Each subunit interacts directly with histones and independently of the two other subunits. The human p150 interacts with histone H3-H4 via an acidic domain of 350 residues containing the KER and ED domains. The third subunit p48 interacts with the N-terminal domain of histone H4, and deleting this domain is not sufficient to abolish in vitro chromatin assembly [65]. Thus, the complexity of the protein interface between CAF-1 and histones currently limits our ability to genetically impair the interaction of CAF-1 with histones.
Although the sole absence of CAF-1 does not confer cell sensitivity to MMS, our data place CAF-1 in a Rad51-dependent template switch pathway by stabilizing D-loop intermediates. Thus, redundant pathways must exist for D-loop stabilization. On the other hand, CAF-1 is critical for faulty template switch events occurring between repeated sequences, a type of NAHR. Protection of the D-loop by CAF-1 during extension by DNA synthesis might provide a mechanism that allows the stabilization of the heteroduplex. This CAF-1–dependent D-loop stabilization might be critical when the homology between DNA molecules is limited (e.g., in NAHR), but alternative mechanisms of stabilizing the heteroduplex likely exist in the case of allelic HR. For example, the ability of Rad51 to branch migrate a single HJ behind the initial point of strand invasion provides the opportunity to extend the heteroduplex without DNA synthesis [66]. Such a mechanism can operate when the two recombinant molecules share a substantial length of homology. The confined length of homology in case of faulty template switch (∼900 bp for the RTS1 sequence compared to unconfined length of homology between sister chromatids) might restrict the effectiveness of this process. Thus, CAF-1 modulates Rad51-dependent template switch, but because of alternative pathways to stabilize the D-loop, a defect in CAF-1 does not completely eliminate template switch. Further investigations are necessary to explore other mechanisms of D-loop stabilization.
In conclusion, CAF-1 promotes Rad51-mediated template switch events at replication forks by counteracting Rqh1-dependent D-loop dismantling. We propose that when CAF-1 switches towards its histone chaperone mode to promote histone deposition, this pathway impacts the resolution of the subsequent template switch event and thus genome stability. In mammals, HR is one of two replication fork maintenance pathways that fuse inverted repeats to mediate chromosome rearrangements, especially in the absence of BLM [11]. Given that functional interactions between CAF-1 and BLM in response to replication stress are evolutionarily conserved, it is possible that the role of CAF-1 in preventing D-loop disassembly is conserved in mammals and might account for the genetic instability associated with Bloom's syndrome.
Strains used were constructed by standard genetic techniques and are listed in Table S1.
The rate of ura4 loss (presented in Table 1), genomic deletion, and translocation was determined as previously reported, as well as PCR analysis of 5-FOA–resistant cells [53]. Statistical significance was detected using the nonparametric Mann–Whitney U test.
RIs were analyzed by 2DGE as reported [3]. Zymolyase-treated cells were embedded in agarose plug, treated with proteinase K, and washed several times in TE. After restriction digestion by AseI, RIs were enriched on BND cellulose columns, precipitated, and separated by 2DGE, according to [67], using 0.35% and 0.9% agarose gel for the first and second dimension, respectively. Quantification of RIs was performed as reported, using a phosphor-imager (Typhoon-trio) to detect 32P-probed signal. Briefly, fork termination and JM signal were quantified as a percentage of stalled fork signal. DNA samples were cross-linked using tri-methyl psoralen (Trioxsalen, Sigma) as follows: 2.109 cells were washed twice in water and resuspended in 20 ml of cold water and placed into an 8.5-cm-diameter glass petri dish on ice. Cells were mixed with 1 ml of Trioxsalen at 200 µ/ml, incubated on ice for 5 min in the dark, and mixed every minute. Cells were then exposed to UV-A (365 nm) for 90 s at a flow of 50 mW/cm2. Analysis of restriction fragments by electrophoresis in denaturing conditions showed that 2–3 inter-crosslinks were formed every 500 bp (unpublished data).
Chromosomal rearrangements were analyzed by PFGE or Southern blot as previously reported [3],[48]. Rad52-GFP enrichment at the RTS1-RFB was performed as previously reported using the primer listed in Table S2 and using a polyclonal anti-GFP antibody (A11122 from Life Technologies) [48].
For immunoprecipitation experiments, the following procedure was used, based on the method published by [29]: 5.108 cells were washed in cold water, resuspended in 400 µl of EB buffer (50 mM HEPES High salt, 50 mM KOAc pH 7.5, 5 mM EGTA, 1% triton X-100, 1 mM PMSF, and anti-protease), ribolysed with glass beads, and centrifuged. The supernatant was recovered and an aliquot of 50 µl was kept as INPUT control. Then, 2 µl of anti-GFP (A11122 from Life Technologies) or anti-Myc (c-Myc 9E10: sc-40 from Santa Cruz Biotechnology) antibody was added to 300 µl of the supernatant and incubated for 1 h at 4°C on a wheel. Then, 40 µl of prewashed Dynabeads protein-G (Life Technologies) was added and then incubated at 4°C overnight. Beads were then washed twice 10 min in EB buffer before separating proteins on acrylamide gel for analysis by Western blot with appropriate antibodies.
Cells were grown in filtered minimal medium (EMM) containing glutamate and implemented in amino acids and bases. Around 5.106 to 1.107 cells from an exponential culture were centrifuged at low speed (1,500 rpm for 1 min) and then resuspended in 1 ml of fresh filtered media. A drop of 1 µl was dropped onto a microscopy agarose slide containing a layer of 1.4% agarose dissolved in filtered media. Cells were observed with a LEICA DMRXA microscope equipped of an oil immersion 100× objective, with a numerical aperture of 1.4 and coupled to a COOLSNAP HQ camera (Roper Scientific, USA). The filters used were a FITC filter to collect GFP signal, CFP for CFP signal, and YFP for YFP signal. Images were taken with the Z-stack (3D) parameterized at 15 slices and were analyzed using METAMORPH (Roper Scientific, USA) and Image J software.
|
10.1371/journal.pntd.0007722 | Prevalence and environmental determinants of cutaneous leishmaniasis in rural communities in Tigray, northern Ethiopia | In Ethiopia guidelines for diagnoses and treatment of leishmaniases are available, but only a few hundred people are diagnosed and receive treatment. A field study has been carried out to determine the status and environmental determinants of cutaneous leishmaniasis (CL) and assess the degree of awareness of the rural communities in affected areas in Tigray, northern Ethiopia.
Following a reconnaissance survey that identified endemic foci, a cross sectional door-to-door survey was conducted in 2009 in five rural communities around the towns of Adigrat and Hagereselam in Tigray. In total 9,622 residents of 1,721 households were clinically screened and household heads interviewed regarding the determinants of infection. The χ2 test and logistic regression were used to determine differences in prevalence between localities, age and sex, and to identify environmental determinants of infection.
The overall prevalence of localized CL was 2.3% (highest 4.7%), with marked inter-village differences. Another 20.9% had scars from previous infections. While risk was sex-independent, prevalence was significantly higher in the 0–9 (4.5%) and 10–19 (2.5%) age groups and predominantly involved the face (82.1%) and upper limbs (13.1%). Nearly 11% of the households had one or more cases of CL and this was associated with proximity to hyrax habitats. All interviewees were knowledgeable about the lesions but ignorant of the disease’s mode of transmission and its association with hyraxes.
The study established that CL is an important public health problem in the study communities, and has been so for a while, as demonstrated by the widespread presence of scars. CL in Tigray appeared to be predominantly of zoonotic nature, mainly transmitted in peri-domestic habitats in proximity to hyrax habitats. Integrated interventions, including awareness creation, are highly recommended.
| Cutaneous leishmaniasis (CL) is a skin infection, transmitted by sandflies. It is most common in Ethiopia, but so far only a few hundred people have received treatment. Five rural villages in Tigray Region, in the north of Ethiopia, were visited to assess the status and determinants of CL. In a door-to-door survey 9,622 residents of 1,721 households were examined and interviewed. A total of 222 had active lesions, an average prevalence of 2.3% CL. Children (up to 9 years old) and teenagers (age 10–19) were more affected than other groups. Most active lesions were found in the face and on arms. Almost 11% of the households had one or more cases of CL and this was associated with proximity to habitats of hyrax, intermediate hosts of the disease. A total of 2009 people (20.9%) showed scars from earlier infections. The findings show how widespread the disease is in the north of Ethiopia and provide some first insights into the environmental factors that influence transmission.
| Ethiopia is among the 98 leishmaniasis endemic countries of the world with both the visceral and cutaneous forms prevalent in the country [1]. An estimated 0.2 to 0.4 million visceral leishmaniasis (VL) cases and 0.7 to 1.2 million cutaneous leishmaniasis (CL) cases occur each year globally [2].CL due to Leishmania aethiopica has long been the most widespread skin disease in the highlands of Ethiopia [3,4,5] principally affecting the poor rural communities. A recent study indicated that an estimated 29 million people are at high risk of CL in the central highlands of the country [6]. Estimates of the annual incidence range from 20,000 to 50,000 cases yearly [7,8], but only a few hundred cases are actually reported [8]. Stable foci are maintained by hyraxes (small herbivorous mammals Procavia capensis and Heterohyrax brucei), and the parasite is transmitted among hyraxes and humans by female phlebotomine sandflies (Phlebotomus longipes and P. pedifer) that feed mainly on hyraxes and share their habitat [9,10,11]. There are three clinical forms of CL due to L. aethiopica: localized CL (LCL), mucosal cutaneous leishmaniasis (MCL), and diffuse cutaneous leishmaniasis (DCL). Although not fatal, persistent LCL, MCL and DCL are disfiguring [12] and may bring long-term psycho-social problems, particularly in young women.
CL is the most neglected even among the neglected tropical diseases (NTDs) in the country. Exact figures on the magnitude of CL in Ethiopia are lacking both nationally and by regional state. The first guideline for diagnosis, treatment and prevention of VL was produced in 2006, and updated with the inclusion of CL in June 2013. However, only very few health centres (eight) diagnose leishmaniasis and as of 2014, only 342 CL cases were treated in VL treatment centres [13]. Diagnosis of CL involves clinical assessment and confirmation with microscopic examination of skin lesion sample. Antimonials are approved for CL treatment in the selected health centres, but most cases are treated traditionally using plants and local application of heat (with hot iron or charcoal fire as per local practice) [8]. There is no leishmaniasis vector control program. Distribution of insecticide-treated nets (ITNs) and insecticide spraying in the context of malaria control may have some impact on phlebotomines in lowland localities where VL is also endemic. On the whole, there is limited evidence, nor control efforts of CL in the country. Outbreaks of CL are not uncommon [14].
The risk of HIV / CL co-infection is also a serious threat as it increases the burden of CL by causing severe forms that are more difficult to manage [15]. Absence or limited access to diagnosis and treatment for CL further increases the urgency for epidemiological surveillance of the disease in the country.
Reports pertaining to CL in Ethiopia date back to 1913, but the disease appears to be around much longer considering the presence of vernacular names in every language where the disease is endemic. Despite its long recognized endemicity [1,3], information on the epidemiology of CL in Ethiopia is still incomplete [16]. Tigray is one of the regions in northern Ethiopia where the status of CL is still unknown. This is mainly due to absence of epidemiological field studies in the region. Due to absence of diagnosis and drugs to treat CL, even reports from health facilities indicating its mere presence in the region have been scarce [15]. This study was initiated following our observation of a steady flow of severe CL cases to a newly established (2005) Italian dermatological unit that started to provide treatment at Ayder Referral and University Teaching Hospital in Mekelle, the capital of the region. With the aim to identify active CL foci and map the pattern of distribution of the disease in the region, a large-scale study has been in progress since late 2008. Part of this study was a door-to-door survey (see Supporting information S1 STROBE checklist) undertaken in five rural communities of Tigray to assess the extent of CL presence and in high risk areas identify potential environmental risk factors, the results of which are presented here.
The study was approved by the Research Ethics Review Committee (RERC) of College of Health Sciences at Mekelle University under reference number CHS/790/DN-16. Permission from the district and respective village authorities was also obtained. Informed oral consent was obtained from the head of the household selected for the study and for those with active lesions signed consent was sought from the guardians. The data were anonymised before analysis by replacing birth dates by age range and household details by name of the subdistrict.
The study was carried out in March and April of 2009 in villages around the towns of Adigrat and Hagereselam in the Tigray National Regional State in northern Ethiopia. The region covers a surface area of about 50,000 km2 and borders with the Sudan in the west and Eritrea in the north (Fig 1). Its nearly 4.3 million people are predominantly rural (80.5%) and engaged in subsistence rain-fed agriculture [17]. The region has a diverse topography, with peak highlands (8%), midlands (39%) and lowlands (53%). Its altitude varies from about 200 meters above sea level (masl) in the north east to almost 4165 masl in the south west. The climate is semi-arid, and the rainfall pattern is mainly unimodal (June to September) but erratic (200 to >1000 mm annually). The regional average annual temperature is about 18.0°C but varies greatly with altitude [18].
The first study site lies to the east of the town of Adigrat. Adigrat is the administrative centre of the eastern zone, located at about 900 km north of Addis Ababa and 120 km north of Mekelle, the regional capital. Locally known for its CL endemicity, the study area comprises of three adjacent subdistricts (kebeles, peasant associations) demarcated by huge gorges: Golea-Genahti, Sasun-Bethawariat, and Kumasubuha (Table 1). The altitude of the study sites ranges from 2248–2650 masl with mild to high temperatures and rainfall is on average 400 mm per year. The area is characterized by deeply incised plateaus, dominated by limestone and sedimentary rock formations providing ideal habitats for hyraxes. The area has moderate to high density population with heavily deforested plains, predominantly scattered bush, and acacia trees. Cactus grows wild in the backyards of most homes. The soils are sandy and of low fertility.
The second study site is located around the market town of Hagereselam (altitude 2650 masl), the administrative centre of Degua-Tembien district 50 km west of Mekelle. It consists of villages belonging to two subdistricts: Mahbereslassie and Michael-Abya (Table 1). The district is about 1033 km2 with an average population density of about 108 persons/km2 [17]. The area has a stepped landscape where flats alternate with steep escarpments. As part of an environmental rehabilitation and reforestation programme nearly 10% of the total area of the district is closed off from people and livestock. Such ‘exclosures’ are mostly located on steep and degraded slopes. The climate of the district is locally classified as Dega (lower highland; 2200–2800 masl), receiving an average annual rainfall of around 700mm. The mean annual temperature is about 15–16°C.
The epidemiological status of leishmaniasis in the study area is not known, owing to the absence of field-based research and limited access to treatment. Based on university hospital records between February 2005 and February 2009, 21 cases of CL were reported from the Adigrat area and 23 cases from the study villages around the town of Hagereselam. VL has never been reported from these highland localities.
The subdistrict villages were selected based on information from medical records of Ayder Referral Hospital and personal experience. In absence of reliable boundary and demographic data at the time, the largest possible sample size was sought by considering every other household in each selected subdistrict. Accordingly, a total of 1721 households were surveyed from the study areas, namely: 437 in Golea-Genahti, 378 in Sasun-Bethawariat, 511 in Kumasubuha, 210 in Mahbereslassie, and 185 in Michael-Abya.
Household members’ socio-demographic, clinical and household data as well as information pertaining to environmental determinants, knowledge on mode of transmission and prevention methods of CL were recorded by a regularly supervised research team consisting of trained health officers guided by a community member knowledgeable on CL. The questionnaire is added as supporting information (S1 Text).
The skin lesions of CL are well identified by the community in its vernacular name “Gizwa”. The required information was sought from the household head or eligible adult present during the visit. Information, collected at each household, included housing type and demographic variables such as age, sex, occupation, duration of residence, and number of animals owned (livestock, dogs). We also observed and recorded the geographic features of the household, including proximity to caves, gorge, and hyrax habitats.
Household heads were also asked for the presence of members with active skin lesions or scars due to CL or other causes. When present, members with evidence of CL were summoned and examined for lesions and scars. For each case, information was recorded on the number, type (LCL, MCL, and DCL) and location of lesions and scars, as well as determinants of infection, such as history of travel and outdoor sleeping. Healed lesions that are depressed, none-or hypo-pigmented, but shiny at the periphery and with rubbery borders, were diagnosed as past CL and referred to as ‘scars’. Papular, nodular or ulcerative lesions with or without satellite lesions and mostly located in uncovered parts of the body were clinically diagnosed as LCL (‘lesions’). Lesions involving the mucosa was considered as MCL while multiple non-ulcerative nodular lesions, often bigger in size from those lesions of LCL were operationally defined as DCL. People with lesions were considered as cases of active CL and this was used to calculate prevalence. Samples of skin snips (n = 51) were taken from a subsample of active cases, smeared on microscope slides, stained with Giemsa and examined for presence of Leishmania amastigotes. The remaining active cases were referred for treatment to Ayder Referral Hospital in the capital and reports of tissue smears of those who visited the hospital were sought from the laboratory records. This confirmed that active CL was caused by Leishmania parasites, most probably L. aethiopica.
Frequencies and proportions were used for the descriptive analysis of the data. The χ2 test was used to determine any statistically significant difference in disease prevalence between age groups, sexes and areas and a P-value <0.05 was considered significant. Any association between the presence of lesions and environmental and host factors was sought using logistic regression. As the flight range of most sandflies is estimated at 300m from their breeding sites [22], this distance was taken as the cut off point for the analysis of environmental factors such as proximity to caves, gorges and hyrax habitats. The data were analysed using SPSS (Statistical Package for Social Sciences, version 16). The full data set is added as supporting information (S1 Data).
A total of 9622 inhabitants in 1721households were surveyed in the selected subdistricts. Males (50.01%) and females (49.99%) were represented almost equally. All study participants resided in the area for more than three years and none of them travelled out of their area within the last 6 months during the study period. However, three people with active CL claimed to have travelled to neighbouring CL endemic subdistricts prior to the appearance of lesions. Prevalence of active CL was 2.3%, with an additional 20.9% of the population showing scars. All active lesions observed during the study were of the localized type (LCL). Of the 1721 households sampled, nearly 11% (188) had one or more cases of active CL, with a total of 60% that had either lesions or scars. Of the 188 households with active CL, 159 (84.6%) households had one case, 25 (13.3%) had two cases, and 4 (2.1%) had three or more cases.
There was a marked difference in prevalence between the study localities. The highest prevalence (active lesions) was observed in Mahbereslassie subdistrict (4.7%) in Degua-Tembien district and in Kumasubuha (2.7%) in Saesie-Tsaedaemba district (Table 2). Most scars, indicative of past infections, were found in Kumasubuha (34.4%). Statistically significant differences were observed between the five study sites, both in the prevalence of lesions (χ2 = 45.860; df = 4; p < 0.001) and scars (χ2 = 628.080; df = 4; p < 0.001).
In the present study, males showed slightly higher rates of active lesions (2.5%; 119/4812) than females (2.1%), but the difference was not statistically significant (χ2 = 1.17, df = 1, p = 0.762). Age specific active prevalence was significantly higher in the 0–9 years olds (4.5%; χ2 = 86.96; df = 3, p < 0.001) than for those more than 10 years old (average 1.6%; χ2 = 86.96; df = 3, p < 0.001). Of those with active lesions, 71.6% (159/222) were under 15 years of age and nearly 20% (44/222) were less than 6 years. The youngest person with active CL was eleven months old.
The majority (82.1%; 207/252) of lesions were found on the face, in which the cheeks and nose were the most affected (Table 3). A similar pattern of distribution was observed with regard to scars. The number of lesions or scars per individual ranged from 1 to 7 (Fig 2). Nearly 90% (199/222) of active CL cases had single lesions and 93.5% (1879/2009) of those with healed lesions had single scars. Most of the single lesions (90%) were of the ulcerative type, with indurated margins and necrotic base, appearing as reddish plaques with irregular borders covered by a firmly adherent crust. Nearly 82% of the active lesions had developed less than two years ago. Out of the 51 active CL skin scraping smears examined, amastigotes were found in 31 (60.8%) of them. All tissue smears, of twenty individuals with lesions who visited the hospital, were found to be positive for amastigotes raising the total percentage to 71.8%.
Several environmental and host related factors were assessed for their association with active CL cases using univariate and multivariate logistic regression (Table 4). Significant variations were observed in prevalence of active CL among age groups, study villages, and physical features of household location, including the presence of hyraxes, caves, and gorges within 300 m of the residence. Individuals in the age group of 0–9 years were nearly five times (OR = 4.71; 95% CI: 3.13–7.1) and those aged 10–19 years were 2.5 times (OR = 2.54; 95% CI: 1.66–3.89) more likely to have CL compared to individuals in the age group of 30 years and above. However, those in the age group 20–29 years were 18% (OR = 0.82; 95% CI: 0.4–1.69) less likely to have active CL. Livestock ownership (OR = 1.65; 95% CI: 1.001–2.71), presence of hyraxes (OR = 4.15; 95% CI: 2.64–6.53), gorges (OR = 3.63; 95% CI: 2.39–5.52), and caves (OR = 3.22; 95% CI: 2.32–7.47) in the vicinity were highly associated with the presence of active CL. Accordingly, participants who lived near hyrax colonies were 4.2 times and those living near caves were 3.2 times more likely to be infected by CL compared to those living far away. Similarly, those living within 300 m of a gorge were found to be 3.6 times more likely to get CL than those far away. Besides, those who slept outdoor were two times (OR = 2.04; 95% CI: 1.29–3.21) more likely to get CL than those who slept inside. Similarly, individuals residing in households with livestock were found to be 65% times more likely to be infected with CL than those without.
When compared to those residing in Michael-Abya subdistrict, inhabitants living in Mahbereslassie were found to be nearly 5.3% (OR = 5.26; 95% CI: 2.58–10.71) times more likely to be infected with CL; in Kumasubuha this was 3% (OR = 2.96; 95% CI: 1.48–5.91), in Golea-Genahti 2.1% (OR = 2.1; 95% CI: 1.01–4.24), and 70% (OR = 1.70; 95% CI: 0.81–3.57) in Sasun-Bethawariat subdistrict (Table 4).
Except for livestock ownership, multivariate logistic regression analysis also showed the significant effect of the environmental and host factors on the odds of being positive for CL (Table 4). Accordingly, the age groups 0–9 years (AOR = 4.67; 95% CI: 3.10–7.04; p< 0.001) and 10–15 years (AOR = 2.48; 95% CI: 1.62–3.81; p<0.001), outdoor sleeping (AOR = 2.67; 95% CI: 1.17–2.67; p = 0.007) and location of residences in proximity to caves (AOR = 2.62; 95% CI: 1.26–2.62; p = 0.001), gorges (AOR = 2.43; 95% CI: 1.57–3.75; p < 0.001) and hyrax habitats (AOR = 2.35; 95% CI: 1.43–3.86; p = 0.001) was highly associated with the presence of active CL.
All interviewed household heads were aware of CL and the great majority (98.7%) identified CL lesions. However, they were invariably ignorant of the disease’s association with hyraxes and its mode of transmission. Nearly all (99.8%) respondents claimed that CL is treated using traditional medicine, such as herbs, holy water (“Tsebel”) and local heat application, in that order (Table 5). Only 4 (0.2%) of the participants claimed the presence of modern treatment for CL. In some villages, farmers made use of the huge piles of hyrax pellets as manure, yet, all the interviewed said that hyraxes eat crops and vegetables and so hyraxes are considered an agricultural pest.
The present study revealed that the localities around the towns of Adigrat and Hagereselam are important CL foci in the region. All cases were of the localized type (LCL) and occurred mostly in the face, with the cheeks in particular. Leishmania aethiopica should be the etiologic agent as in a follow-up study conducted in a neighbouring subdistrict it was isolated as the main Leishmania species causing CL in the area [23], as in other parts of the country. The overall prevalence rate of active LCL in the study communities ranged from 0.9–4.7%, for scars this was 3.9–34.4% (Table 2). This is indicative of the public health significance of CL in these areas previously and currently. Despite good awareness and recognition of CL, understanding the mode of transmission lacked. This may have contributed to the high prevalence of the disease in the study villages and holds a risk of further spread in the future, since no active detection and treatment strategies are in place. Moreover, knowing that CL is transmitted by biting sandflies does not necessarily lead to adequate prevention strategies [24]. Integration with malaria control by distribution of ITNs and indoor residual spraying is unlikely as CL is predominantly prevalent in the malaria free highland areas.
The observed absence of gender sensitivity to infection by CL is consistent with several studies conducted elsewhere in the country. The predominance of CL in the young (0–9 years old) is also well known and resonates with countrywide data (e.g. [25,26], who found 8.5% and 7.1% respectively). In established endemic areas, CL prevalence typically increases with age up to 15 years, after which prevalence levels off, presumably because of the acquisition of immunity. In this study, the occurrence of the disease almost in equal proportion in both sexes, with large numbers of women and children infected, including those under one year, probably reflects that CL transmission may have occurred in peri-domestic habitats, where sandfly exposure is most equally distributed among individuals. This also explains why sleeping outdoors is an important (peri-domestic) risk factor. This is further corroborated by the fact that Phlebotomus longipes, the proven vector of CL in the Ethiopian highlands [9], was collected both from indoor and predominantly outdoor locations (compounds of houses) in a follow-up study conducted in the neighbouring subdistrict [23]. In peri-domestic settings, sandflies rest in cool, dark and humid corners of animal shelters or human dwellings [27,28]. As with rodent burrows, peri-domestic areas also provide ready access to bloodmeals in addition to shelter and suitable breeding grounds in decaying organic matter (manure).In line with previous reports from different parts of the country [9,23,29,30,31], the presence of CL cases in households was closely associated with the presence of hyrax colonies in the vicinity, indicating that the disease is mainly of zoonotic nature. The intimate ecological association of rocky hyraxes and the sandfly species Phlebotomus longipes and P. pedifer, the proven vectors of L. aethiopica induced CL in Ethiopia, is well established [30,31]. Two species of hyrax (Procavia capensis and Heterohyrax brucei) are the widely incriminated reservoir hosts of L. aethiopica in the country [30]. Livestock was also a risk factor, as P. longipes readily feeds on cows [9] and manure provides breeding habitat.
In the present study, some of the most important environmental determinants for CL occurrence were location of households near gorges and on rocky hillsides and presence of caves nearby. This is consistent with previous reports by Ashford [9] and Lemma et al. [3], who pointed out that gorges and escarpments, rock cliffs and mountainous areas constitute favourable environments for the reservoir host (hyrax). These features result in steep slopes, identified as a risk factor for CL by Seid et al. [6]. The presence of livestock and their manure further creates favourable breeding grounds for the sandfly vectors in peri-domestic environments.
Overall, this study establishes that important foci of CL exist in Tigray, northern Ethiopia, with children and young adults being the most affected. The data further highlight that the disease is predominantly of zoonotic nature, and mainly transmitted in peri-domestic habitats where hyraxes prevail in the vicinity. Apart from their role as reservoirs of CL, the status of hyraxes as agricultural pests needs to be determined. This offers perspectives for environmental transmission control complementing minimal efforts in passive and active case detection, treatment, reporting, and data analysis. Reservoir control, i.e. small-scale eradication of hyraxes in the proximity of dwellings, would be effective, especially combined with fogging of their habitats to reduce sandfly densities. Vector control alone is unlikely to be effective. In addition, regular education to children and adults on the transmission and prevention of CL is recommended. All control efforts should be evaluated periodically, with their impact on incidence, building on essential active case detection and monitoring.
Finally, although the study was carried out in 2009, we believe that our results still pertain to the present situation of CL in Ethiopia as there has been no progress in advancing treatment or prevention activities. Following our study, an awareness creating international consultative meeting was held by WHO and Federal Ministry of Health in 2011 and the first National Neglected Tropical Disease (NTD) Master Plan was launched In June 2013, to achieve WHO NTD elimination and control targets by 2020. As of September 2015, although the Federal Ministry of Health has managed to mobilize support to implement mass drug administration in 84% - 100% of the endemic districts for other NTDs (trachoma, onchocerciasis, lymphatic filariasis, soil-transmitted helminths and schistosomiasis), there has been no progress in advancing treatment or prevention activities in CL, owing to the absence significant domestic or international donors to support CL intervention activities. A follow-up study in 2013 (not part of this investigation) and further studies in progress indicate an even greater magnitude both in terms of spread and in level of prevalence of CL in the region [23]. Unfortunately, today CL remains neglected even among the NTDs.
|
10.1371/journal.pgen.1004951 | Modeling of the Human Alveolar Rhabdomyosarcoma Pax3-Foxo1 Chromosome Translocation in Mouse Myoblasts Using CRISPR-Cas9 Nuclease | Many recurrent chromosome translocations in cancer result in the generation of fusion genes that are directly implicated in the tumorigenic process. Precise modeling of the effects of cancer fusion genes in mice has been inaccurate, as constructs of fusion genes often completely or partially lack the correct regulatory sequences. The reciprocal t(2;13)(q36.1;q14.1) in human alveolar rhabdomyosarcoma (A-RMS) creates a pathognomonic PAX3-FOXO1 fusion gene. In vivo mimicking of this translocation in mice is complicated by the fact that Pax3 and Foxo1 are in opposite orientation on their respective chromosomes, precluding formation of a functional Pax3-Foxo1 fusion via a simple translocation. To circumvent this problem, we irreversibly inverted the orientation of a 4.9 Mb syntenic fragment on chromosome 3, encompassing Foxo1, by using Cre-mediated recombination of two pairs of unrelated oppositely oriented LoxP sites situated at the borders of the syntenic region. We tested if spatial proximity of the Pax3 and Foxo1 loci in myoblasts of mice homozygous for the inversion facilitated Pax3-Foxo1 fusion gene formation upon induction of targeted CRISPR-Cas9 nuclease-induced DNA double strand breaks in Pax3 and Foxo1. Fluorescent in situ hybridization indicated that fore limb myoblasts show a higher frequency of Pax3/Foxo1 co-localization than hind limb myoblasts. Indeed, more fusion genes were generated in fore limb myoblasts via a reciprocal t(1;3), which expressed correctly spliced Pax3-Foxo1 mRNA encoding Pax3-Foxo1 fusion protein. We conclude that locus proximity facilitates chromosome translocation upon induction of DNA double strand breaks. Given that the Pax3-Foxo1 fusion gene will contain all the regulatory sequences necessary for precise regulation of its expression, we propose that CRISPR-Cas9 provides a novel means to faithfully model human diseases caused by chromosome translocation in mice.
| Many cancers carry recurrent chromosome translocations, which often result in the formation of fusion genes that are directly involved in the tumorigenic process. Alveolar rhabdomyosarcoma, a muscle tumor in children, is typified by a translocation that fuses the PAX3 gene on chromosome 2 to the FOXO1 gene on chromosome 13. For translocation to occur both genes need to break and the disparate ends need to fuse via a process called non-homologous end joining. We determined that physical proximity of Pax3 and Foxo1 in mouse muscle progenitor cells (myoblasts) facilitates fusion gene formation. Because Pax3 and Foxo1 in the mouse are in an opposite orientation, we used a chromosome engineering strategy to invert the orientation of Foxo1 so that upon translocation a productive Pax3-Foxo1 fusion gene is created. Co-localization of the Pax3 and Foxo1 loci is higher in fore limb than in hind limb myoblasts. Simultaneous induction of a targeted double strand DNA break in each gene by CRISPR-Cas9 nuclease generated more fusion genes in fore limb than in hind limb myoblasts. Thus, gene proximity facilitates fusion gene formation. We propose that CRISPR-Cas9 nuclease can be used for the precise modeling of chromosome translocations of human cancer in mice.
| Rhabdomyosarcoma (RMS) is the third most common soft-tissue sarcoma in children with an annual incidence of five new cases per million. It accounts for 5–8% of all pediatric cancer. RMS belongs to the family of small round blue cell tumors of childhood and exhibits histological features of skeletal muscle. Two major histological subtypes of RMS can be distinguished, embryonal (E-RMS) and alveolar (A-RMS). E-RMS has its highest incidence in infants and young children whereas A-RMS is more frequent in older children and adolescents. A-RMS has a more aggressive clinical behavior with early dissemination, a poor response to chemotherapy, frequent relapses, and a 5-year failure-free survival of 65% after treatment [1]. A-RMS is found predominantly in the extremities (42%), parameningeal (17%), head and neck (11%) and other locations (21%) [1] including the trunk, perirectal and perianal areas [2, 3]. Cytogenetically A-RMS is distinguished from E-RMS by one of two recurrent chromosome translocations: t(2;13) or t(1;13), which result in fusion of PAX3 or PAX7 to FOXO1, respectively [4].
In spite of multiple attempts to identify the cell of origin in which the t(2;13) occurs the question remains unanswered. It was shown previously that transcription occurs at a few hundred discrete nuclear sites called transcription factories [5]. Some genes frequently involved in a recurrent chromosome translocation (MYC and IGH in B lymphoid progenitors, TMPRSS2 and ERG or ETV1 in prostate cancer, RET and H4 in in radiation-associated papillary thyroid cancer) co-localize to the same transcription factory [6–9]. Initial chromosome conformation capture experiments in activated mouse B cells suggested that physical proximity of the IGH and MYC loci is a minor contributor to the frequency of chromosome translocation [10]. However, combined high resolution Hi-C mapping and genome-wide translocation sequencing in transformed mouse pre-B cells found good coincidence between chromosomal translocation and spatial proximity [11]. A possible driver of double strand DNA breaks might be the co-localization of replication stress-induced early replication fragile sites (ERFSs) with highly expressed gene clusters [12]. Though it was demonstrated that ectopic expression of PAX3-FOXO1/Pax3-Foxo1 can transform mouse mesenchymal stem cells in vitro [13] as well as Myf6+ myofibers in vivo [14] in view of the above these cell types seem unlikely hosts for the chromosome translocation given that they do not express Pax3. In fact, the suggestion that Myf6+ myofibers might be the host of the PAX3-FOXO translocation was recently rectified [15].
In contrast, Pax3 is expressed in activated myoblasts upon muscle injury or in growing muscles during normal development [16]. Moreover, PAX3-FKHR, in cooperation with loss of p16INK4A expression, transforms both fetal and postnatal primary human skeletal muscle cell precursors [17]. Together these observations suggest that translocation might occur in a population of activated myoblasts that express PAX3 (PAX3+). It has been shown that Pax3 expression differs among different muscles in the mouse [18, 19]. There are many more Pax3+ cells in fore limb than in hind limb muscles [19]. Muscle satellite cells from the masseter and soleus did not express Pax3 while only 7% of those from the extensor digitorum longus (EDL) did. In contrast 49% of satellite cells from the biceps were Pax3+. In addition, most ventral trunk muscles were Pax3-positive and 64% of satellite cells from the diaphragm expressed Pax3. Importantly, primary myoblast cultures of Pax3+ satellite cells remain Pax3+, while Pax3- satellite cells from hind limb remain negative [19].
Studies addressing the relation between spatial chromosome proximity and translocation have been performed in cells of the B-lymphoid lineage or of hormone-responsive lineages mostly using transformed cell lines [6, 7, 9]. Recently CRISPR-Cas9 nuclease (Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR)/CRISPR-associated systems) [10] was used to engineer human tumor-associated translocations [20]. To answer the question if locus proximity of Pax3 and Foxo1 in low-passage primary mouse myoblasts contributes to the frequency of Pax3-Foxo1 fusion gene formation we used the CRISPR-Cas9 system to induce double strand DNA breaks (DSBs), which spurred by non-homologous end joining repair (NHEJ) produce chromosome translocations between these two loci. We used synthetic single-guide RNAs (sgRNA) to program Cas9 to induce DNA double-strand breaks (DSBs) in Pax3 and Foxo1 [21–23]. Unlike the human PAX3/7 and FOXO1 genes, mouse Pax3/7 and Foxo1 are in opposite orientation on their respective chromosomes (1, 4, and 3). Compared with human chromosome 13, Foxo1 is part of an inverted 4.9Mb syntenic region on mouse chromosome 3. Although a recurrent complex inversion/translocation event involving the oppositely oriented ETV6 and c-ABL genes in humans gives rise to the ETV6-ABL fusion gene in some myeloid and lymphoid malignancies, the frequency of this event is extremely low [24]. Therefore, to successfully generate a CRISPR-Cas9-mediated Pax3-Foxo1 fusion gene we used chromosomal engineering via Cre recombinase-mediated genetic alterations to create a mouse in which the Foxo1 containing 4.9 Mb syntenic region is inverted (Foxo1-inv+/+ mice). Previously, Cre recombinase-mediated inversions of large fragments of chromosomes have been used to create balanced chromosomes [25–29].
We show that myoblasts isolated from fore and hind limb keep their Pax3-expressing identity and co-localization of Pax3 and Foxo1 loci strongly correlates with the level of Pax3 expression and generation of a CRISPR-Cas9 induced t(1;3), which is more frequent in fore limb myoblasts. Our Foxo1inv+/+ mice will be a valuable tool for studying mechanisms underlying the initial stages of the A-RMS implicated chromosome translocations resulting in development of better animal models for this pediatric cancer and other human diseases caused by chromosome translocations.
Since close physical proximity of translocation partners might facilitate chromosome translocation, we determined if Pax3 and Foxo1—the translocation partners in A-RMS—are co-localized in actively proliferating low-passage primary mouse myoblasts. DNA-FISH analyses of the Pax3 and Foxo1 loci in interphase nuclei of primary limb myoblasts of newborn pups, after one week in culture showed 13% co-localization, which was significantly higher than in similarly cultured MEFs (2%, the background of this method; Fig. 1A). We hypothesized that co-localization of Pax3 and Foxo1 loci in myoblasts reflects the percentage of Pax3+ cells in the original newborn muscles. To test this hypothesis we isolated myoblasts from hind and fore limbs of newborn pups and compared the frequency of co-localization of Pax3 and Foxo1 loci in proliferating myoblasts from these two sources. It was shown previously with a Pax3 knock-in reporter gene that many more satellite cells in fore limb muscle express Pax3 than in hind limb muscle [19]. In accordance, the percentage of co-localized Pax3 and Foxo1 loci was notably higher in fore limb than in hind limb myoblasts in 8 independent experiments (Fig. 1A–C). In addition, Q-RT-PCR of RNA from these myoblasts confirmed that expression of Pax3 was six-fold higher in fore limb myoblasts (Fig. 1D). These results are in agreement with the published observation that satellite cells maintain their Pax3+ identity upon activation in vitro. Expression of other genes such as Foxo1 and Pax7 was similar in the two types of myoblasts (Fig. 1D). The results for Pax3 expression were reproducible given that a number of independent experiments produced similar data (S1 Fig.). Because diaphragm was shown to contain the highest number of Pax3+ myoblasts [19], we compared by FISH the co-localization of the Pax3 and Foxo1 loci in myoblasts isolated from fore limb, hind limb, and diaphragm of the same adult mouse. Indeed, diaphragm myoblasts showed a higher co-localization of the two loci (20%) than fore limb (11%) or hind limb (9%) myoblasts.
The mouse Foxo1 gene is located on chromosome 3 in a 4.9 Mb DNA fragment (ch3:52,059,615–56,995,963) that is syntenic with human chromosome 13 (ch13:41,254,213–34,463,185) but positioned in the opposite orientation (Fig. 2A). This places Foxo1 in the mouse in a reverse transcriptional direction with respect to that of the Pax3 or Pax7 genes. To engineer a mouse capable of acquiring productive Pax3/7-Foxo1 fusion genes via a simple balanced t(1;3) or t(4;3), we performed two consecutive rounds of ES cell targeting in which we introduced two pairs of non-compatible LoxP sites at either border of this syntenic region with the goal to create a Cre-recombinase mediated permanent inversion of the 4.9Mb DNA fragment (S2 Fig.). Without inversion there would only be two ways to produce a productive fusion: 1) Via a translocation in which the resulting chromosomes would carry a double centromere and no centromere, respectively, an option likely to be non-viable in primary myoblasts and 2) Via a complex inversion/translocation event as described for the human ETV6-ABL fusion gene [24], a rare event, which likely would reduce the frequency of fusion gene formation below detectable levels.
The centromeric border of the mouse/human syntenic region is located 15 kb upstream of the Foxo1 start codon (Fig. 2B). To precisely target this border in ES cells we used recombineering in E. coli [30] to modify the RP24–391O12 BAC (bacterial artificial chromosome) clone, so that it carries non-compatible mutant 511-ILoxP and wtLoxP sites [31] flanking the hph (hygromycin B resistance) and tk (HSV1-thymidine kinase) selectable marker genes (Figs. 2B, S2). The precise targeting of the border of the syntenic region minimizes the chance of disturbing any potentially important regulatory sequences that might affect Foxo1 expression (Fig. 2B, top). We targeted ES cells with linearized RP24–391O12-LoxP-hygro-TK BAC DNA and counter selected hygromycin B resistant clones carrying random integrations by screening for the presence of vector sequences remaining on either side of the insert. Colonies containing such vector segments were discarded [32]. The remaining clones were subjected to the ‘loss-of-native-allele’ assay using real-time quantitative PCR [33]. For copy number control of stably integrated target DNA we used the telomeric border of the syntenic region as a reference. In total 273 clones were analyzed, two of which contained a single copy of the wild type locus (Fig. 2C). These clones were submitted to FISH analysis and karyotyping which confirmed the presence of only two native signals on chromosome 3 when hybridized with a wild type BAC RP24–391O12 probe (Fig. 2D). For consecutive targeting of the telomeric border of the syntenic region we selected clone XIIB3, which had a 100% normal diploid karyotype.
For targeting of the telomeric border of the syntenic region we followed the same strategy and engineered a BAC clone carrying the 511-ILoxp-Neo-TK-wtLoxP cassette inserted at the precise syntenic border (Fig. 2B, bottom). The recombinant RP23–422I13-LoxP-Neo-TK BAC was linearized in such a way that only very short vector fragments remained at either side of the insert. After targeting in ES cells, analysis with the ‘loss-of-native-allele’ assay of 48 clones proved sufficient to obtain the desired recombinant. Two clones carrying a single copy of the wild type telomeric locus (Fig. 2E) were analyzed by FISH using the RP24–391O12 and RP23–422I13 BAC probes. One of them (13D3) showed two native signals on chromosome 3 with either BAC probe (Fig. 2F). This clone had a 90% normal diploid karyotype and we next determined if it carried cis- or trans-targeted borders of the 4.9 Mb syntenic region.
To discriminate between these two possibilities we transiently transfected a Cre recombinase plasmid into the double-targeted 13D3 ES cells and DNA isolated from the pool of electroporated cells was analyzed by PCR using only forward or reverse primers from both targeted borders. Both PCRs produced bands indicating that the pool contained cells carrying the 4.9Mb inversion. The same pool of cells was counter selected with FIAU and 23 resistant clones were analyzed by PCR (Fig. 3A). Sixteen clones harbored the 4.9Mb inversion and two of these were selected, A6 and C5, which had a 100% and 93% normal karyotype, respectively. Inversion of the 4.9Mb region in these clones was subsequently confirmed by FISH analysis (Fig. 3B) using the RP24–391O12 and RP23–422I13 BAC probes. The chromosome containing the inversion showed split hybridization signals while the wild type chromosome produced contiguous signals with these probes. These ES cell clones were used to generate chimeric mice that transmitted the inversion of the Foxo1 syntenic region to heterozygous Foxo1-inv+/- offspring.
Foxo1-inv+/- mice were fertile and produced Foxo1-inv+/+ offspring at the expected Mendelian frequency. Foxo1-inv+/+ animals did not exhibit any obvious phenotypic abnormalities and showed normal fecundity and life span. Moreover, western blot analysis confirmed that Foxo1-inv+/+ primary myoblasts and wild type myoblasts expressed equal amounts of Foxo1 protein (Fig. 3C) and co-localization of the Pax3 and Foxo1 loci was equal in Foxo1-inv+/+ and wild type myoblasts (8%, S5 Fig.). Finally, DNA-FISH analysis of Foxo1-inv+/+ fibroblasts with RP24–391O12 and RP23–422I13 BAC probes confirmed that both chromosomes 3 carried the 4.9Mb inversion (Fig. 3D).
Nuclear receptor-induced chromosomal proximity of TMPRSS2 and ERG or TMPRSS2 and ETV1 promotes the occurrence of nonrandom ligation sites upon translocation between these partner genes, thereby generating unique breakpoint “hot spots” [6]. It is possible that translocations in A-RMS are non-random and occur predominantly at sites, coming in close proximity during co-regulated expression. We hypothesized that directing DSBs to sites in mouse Pax3 and Foxo1 homologous to those in PAX3 and FOXO1 in an ARMS cell line carrying a t(2;13) might increase the chance of generating a t(1;3) in proliferating Foxo1-inv+/+ myoblasts after Cas9 induced DSBs. We chose to mimic the breakpoints of the widely used ARMS cell line RH30 (S3 and S4 Figs.). Alignment of human and mouse Pax3 and Foxo1 sequences mapped the RH30-like breakpoints at positions 78105273 on mouse chromosome 1 and 52300558 on mouse chromosome 3 (Fig. 4A). We chose unique protospacer sequences followed by a 5’-GGT PAM as close as possible to the RH30-like breakpoints in both Pax3 and Foxo1 (Fig. 4B). Cas9 introduces DSB three nucleotides downstream of the two PAM sequences, which would result in DSBs between nucleotides 78105248 and 78105247 on chromosome 1 in intron 7 of Pax3 and between nucleotides 52300541 and 52300542 (coordinates in the non-inverted sequence) on chromosome 3 in intron 1 of Foxo1 (Fig. 4B).
For gene delivery to the myoblasts we cloned the human codon optimized Cas9 (hCas9) into the pCL20C [34] lentiviral vector downstream of the MSCV promoter and upstream of an IRES-YFP fluorescent marker (Fig. 4C). In order to express two different sgRNAs form a single vector we constructed a second pCL20C dual sgRNA vector in which the Pax3-specific sgRNA was driven by the human U6 promoter and the Foxo1-specific sgRNA by the mouse U6 promoter (Fig. 4C,D).
We first determined that with our current batch of serum maximum co-localization of Pax3 and Foxo1 occurred at 7–8 days of culture after myoblast isolation. This time point synchronized with Cas9 and sgRNA expression should therefore maximize the probability of introducing DSB in closely positioned Pax3 and Foxo1 loci. Hence 24 hours after isolation we transduced primary fore and hind limb myoblasts of Foxo1-inv+/+ pups, Foxo1-inv+/+ MEFs and fore limb myoblasts from wild type mice with Cas9 lentivirus (Fig. 4C). After FACS sorting for YFP, cells were expanded and transduced with lentivirus expressing the RH30-like sgRNAs at day 7 after isolation and with an SV40 large T antigen expressing lentivirus at day 8. The latter was done to prevent senescence of the myoblasts during puromycin selection and allows subsequent expansion of the culture. To detect the Pax3-Foxo1 fusion DNA fragments from Cas9/sgRNAs expressing myoblasts and MEFs we used the Pax3-RH30F (forward) and Foxo1-RH30R (reverse) primers for PCR analysis, which are positioned downstream and upstream of the putative Cas9-induced Pax3 and Foxo1 DSBs (Fig. 5A). PCR amplification of DNA from 104 cells produced bands of 250 bp or shorter in Cas9/sgRNAs expressing hind limb and fore limb Foxo1-inv+/+ myoblast (Fig 5B, lanes 2 and 4). However, no product was detected upon PCR amplification of DNA from 104 hind and fore limb Foxo1-inv+/+ myoblasts not treated with Cas9/sgRNAs (Fig. 5B, lanes 1 and 3) or from 104 Cas9/sgRNAs expressing Foxo1-inv+/+ MEFs or wild type fore limb myoblasts (Fig. 5B, lanes 5 and 6). As a control we verified that the difference in translocation frequency between myoblasts and MEFs was not caused by differences in CRISPR-Cas9’s accessibility to chromatin, given that Pax3 is not expressed in MEFs. The CRISPR-Cas9 breakpoint in Pax3 falls within a MaeIII restriction endonuclease site and that in Foxo1 within a DdeI site. Therefore we PCR amplified the Pax3 and Foxo1 fragments spanning the breakpoints and digested them with MaeIII or DdeI. This showed that 96% (Pax3) and 97% (Foxo1) of the PCR products of CRISPR-Cas9 treated myoblasts were resistant to MaeIII or DdeI digestion, whereas in CRISPR-Cas9 treated MEFs these numbers were 72% for both enzymes (Fig. 5C,D). Thus, there was no great difference in chromatin accessibility. Moreover, the Pax3 and Foxo1 chromatin in MEFs was equally accessible to CRISPR-Cas9, despite the fact that Foxo1 is and Pax3 is not expressed in these cells.
Cloning of the CRISPR-Cas9 induced fusion DNAs, followed by sequencing of 45 individual clones of each of the PCR products, produced 39 and 34 translocation breakpoint sequences from fore and hind limb myoblasts, respectively. This identified 6 different breakpoint sequences from fore limb and 3 different breakpoint sequences from the hind limb myoblasts. This represents the minimal number of translocation events per 104 cells (Fig. 5E, top and bottom). Taking into account the percentage of locus co-localization (Fig. 5F) these numbers translate to a minimal translocation frequency of 1 in 150 in fore limb and 1 in 200 in hind limb myoblasts, respectively. The only sequence in common between the fusion fragments from these two types of myoblasts was the cleanly re-ligated fusion, without missing or added base pairs. The other 7 (5 from fore limb myoblast and 2 from hind limb myoblast) were all unique and carried NHEJ-mediated deletions varying from 6 to 71 bp. Superimposed on the deletion, two of the clones also contained randomly added base pairs. Notably, three additional breakpoint sequences obtained from an independent experiment (S5 Fig.) were different from the 7 shown in Fig. 5E and underline the mutation-prone repair of the NHEJ DNA-repair machinery during the translocation event. Together these results show excellent correlation between the frequency of translocation, co-localization, and expression of the Pax3 and Foxo1 loci in primary myoblasts. It was highest in fore limb myoblasts, lower in hind limb myoblasts and undetectable in MEFs. Although wild type myoblasts show the same frequency of locus co-localization as Foxo-inv+/+ myoblasts (S6 Fig.), the opposite orientation of Foxo1 prevented the formation of a productive Pax3-Foxo1 fusion gene. Next we performed RT-PCR on equal amounts of total RNA from fore limb and hind limb myoblasts to detect the Pax3-Foxo1 fusion mRNA. In support of the higher frequency of chromosome translocation in fore limb myoblasts, we were able to RT-PCR amplify the Pax3-Foxo1 cDNA from these myoblasts (Fig. 5G) but not from the hind limb myoblasts using an equal amount of input RNA (not shown). Sequence analysis of the cDNA confirmed the correctly spliced Pax3 exon 7-Foxo1 exon 2 fusion (Fig. 5G).
To further characterize the t(1;3) we repeated the experiment in Foxo1-inv+/+/Ink4a-ARF-/- myoblasts. Due to loss of a functional p53 pathway Ink4a-ARF-/- myoblasts do not senesce during further experimental manipulation. Based on the Pax3 and Foxo1 co-localization data at the time of induction of the t(1;3) (11% in fore limb myoblasts and 7% in hind limb myoblasts) we assumed that the frequency of translocation events in these myoblasts should not be lower than in the myoblasts used in Fig. 5, i.e. at least 6 independent translocation events per 104 fore limb myoblasts. This frequency is too low for further molecular and functional analyses. To enrich the cell pool for the t(1;3) carrying cells, we evenly distributed 104 cells between the wells of three 96-well plates (on average 30 cells per well). PCR analyses of the DNA of 95 wells from the first plate identified 3 potentially t(1;3)-enriched cell pools (S7 Fig.). Pool 1E10 was lost during the freeze-thawing cycle but FISH analyses detected the reciprocal t(1;3) in 64% of pool 1G3 metaphase cells (Fig. 6A–C) and in 4% of pool 1D10 metaphase cells. Both the derivative chromosomes 1 and 3 were detected in all t(1;3) positive cells, confirming that the translocation was reciprocal.
To determine if the t(1;3) resulted in expression of Pax3-Foxo1 protein we immunoprecipitated three cell lysates each of the t(1;3)-negative (1H3) and t(1;3)-positive (1G3) pools with either an anti-Pax3 or an anti-Foxo1 antibody. The Pax3 IPs were then immunoblotted with the anti-Pax3 antibody, showing the Pax3 and Pax3-Foxo1 bands (Fig. 6D, Pax3/Pax3 panel), or with anti-Foxo1 antibody showing only the Pax3-Foxo1 band (Fig. 6D, Pax3/Foxo1 panel). Similarly, immunoblotting the Foxo1 IPs with anti-Pax3 antibody again showed the Pax3-Foxo1 fusion protein (Fig. 6D, Foxo1/Pax3 panel) while immunoblotting with the Foxo1 antibody showed both Foxo1 and the fusion protein (Fig. 6D, Foxo1/Foxo1 panel). This confirmed that the engineered t(1;3) expressed the fusion protein, which allowed us to assess if it affected the expression of Pax3-Foxo1’s transcriptional targets. We performed RNA-seq analysis comparing the mapped sequence reads of presumed PAX3-FOXO1 target genes [2] in the 1G3 pool (64% Pax3-Foxo1 positive) with those in the 1H3 pool (Pax3-Foxo1 negative) (S1 Table). This showed that roughly half the targets of PAX3-FOXO1 were correctly up or down regulated in the 1G3 pool. The same comparison with a PAX3-FOXO1 expression signature obtained with the ectopic PAX3-FOXO1 expressing ERMS cell line RD [35], also showed coincident regulation of half the targets (S2 Table), suggesting that the t(1;3) generated fusion protein is active.
For the precise modeling of human recurrent chromosome translocations and their impact on disease development in mice, reenactment of the actual translocation would be the closest possible recapitulation of the sequence of events in humans. Until now such reenactment was a daunting task as the translocation would require introduction of LoxP [36, 37] or Frt recombination sites into both translocation partners via homologous recombination in ES cells, followed by expression of Cre or Flp recombinase to create DSBs that would mediate the translocation. As shown by others [20] and here, the availability of the CRISPR-Cas9 system has paved the way to implementing this approach without such major technical or time investment. Given the high homology between mouse and human genes and their regulatory sequences, this approach is likely to include all sequences that are important for the precise regulation of the mouse fusion gene as it occurs in humans. The first and only published model for ARMS [38] in which expression of a conditional Pax3-Foxo1 knock-in fusion oncogene is induced by a Myf6 driven Cre had a low incidence and long latency of tumor development, requiring the presence of two Pax3-Foxo1 alleles on a Trp53-null or Ink4a/Arf-null background. One reason for this might be that the level of expression of the fusion oncogene in this KI model is inadequate for shorter latency tumor development. An argument against this possibility is that a high level of PAX3-FOXO1 expression induces cell death [39], most likely due to transcriptional activation of the Pax3-Foxo1 pro-apoptotic target gene Noxa1 [40]. Unlike other studies [41, 42], the KI Pax3-Foxo1 gene contained some Foxo1 genomic sequences that allowed expression of the fusion gene in adult mice, but despite their presence the construct might lack sequences that mediate human-like regulation of fusion gene expression, which in turn might be crucial for efficient tumor development.
In agreement with published data [19] we established that co-localization of Pax3 and Foxo1 in our culture system was higher in forelimb than in hind limb myoblasts, which coincided with higher Pax3 expression in forelimbs. Due to experimental variability the percentage of co-localization of the two loci varied in 8 independent experiments, but co-localization in the fore limbs was always higher than in the hind limbs. Therefore our myoblast model represents a graded system to determine if these features contributed to the frequency of chromosome translocation in low passage primary myoblasts upon introduction of targeted DSBs. To perform these experiments and to eventually develop a precise mouse model of ARMS, the transcriptional orientation of Foxo1 on chromosome 3 needed to be inverted. We followed the Cre-dependent one-way inversion of a DNA fragment in mice as was previously demonstrated by Schnütgen and colleagues [43]. To avoid disturbing the transcriptional regulation of the inverted Foxo1, we decided to invert the mouse/human 4.9 Mb syntenic region encompassing Foxo1, rather than the gene itself. Although the centromeric border of this region is only 15 kb upstream of Foxo1, we reasoned that all important Foxo1 regulatory sequences should be contained within this region otherwise it would not be syntenic with human FOXO1 on chromosome 13q14.1. Although we did not analyze the detailed expression of Foxo1 in Foxo1-inv+/+ mice during pre- and postnatal life, the animals did not show any obvious phenotypic abnormalities. In addition, they had a normal lifespan, normal fecundity, and the level of Foxo1 protein expression and co-localization of the Pax3 and Foxo1 loci in myoblasts were identical to those of wild type mice. Together these observations made the Foxo1-inv+/+ myoblasts suitable for our translocation experiments.
To determine if the level of co-localization of Pax3 and Foxo1 in primary myoblasts affected the frequency of chromosome translocation between these loci upon induction of targeted DSB, we transduced the cells with Cas9 and dual sgRNA expressing lentiviruses. Combining the three genes into a single lentiviral vector failed to produce viral particles. We targeted the CRISPR-Cas9 DSBs to sequences in Pax3 and Foxo1 that mediated the t(2;13) in the A-RMS cell line Rh30. Both breakpoints are present in sequences conserved between the mouse and human genes, suggesting that they occurred in non-redundant sequences that might bind factors with a role in expression regulation of both genes. Currently we do not know if this affects the frequency of translocation, which is a possibility that can be tested in future by choosing sgRNAs targeting non-conserved sequences within the target Pax3 and Foxo1 introns. We found excellent positive correlation between the frequency of the t(1;3) and the percentage of locus co-localization using FISH analysis. This also correlated with the level of Pax3 expression, which is much higher in fore limb than hind limb myoblasts and absent in MEFs, while Foxo1 expression is ubiquitous. Given that the frequency of CRISPR-Cas9 induced DSBs in Pax3 and Foxo1 is comparable in myoblasts and MEFs, it is the proximity of the loci in these primary cells that facilitates trans-chromosomal ligation producing the two expected derivative chromosomes during NHEJ DNA repair. The derivative chromosome 3 produced correctly spliced Pax3-Foxo1 mRNA, encoding active Pax3-Foxo1 protein that up/down-regulated expression of approximately half the presumed PAX3-FOXO1 targets in the 64% Pax3-Foxo1-positive cell pool (S1 Table). The genes compiled in this table are differentially expressed in ARMS versus ERMS tumors or have been identified by forced expression of PAX3-FOXO1 in different cell lines, including NIH3T3 cells, MEFs, SAOS2 cells and C2C12 cells ([2] and references therein). Because the cell background affects the range of PAX3-FOXO1 target gene expression [44], none of the published scenarios reflect expression of Pax3-Foxo1 in primary p16/Arf-/- mouse myoblasts. Possibly this is the reason for the 45% match of reported PAX3-FOXO1 up or down regulated genes. Comparison with genes up or down regulated in the ERMS cell line RD transduced with PAX3-FOXO1 retrovirus [35] showed 52% coincident regulation (S2 Table). Clearly, the t(1;3) generated Pax3-Foxo1 protein in mouse myoblasts is active and changes the expression of target genes in an ARMS-like manner.
One would expect that the frequency of translocation in myoblast that show co-localization of the two translocation partners would be the same irrespective of the source of myoblasts. We found a frequency of 1/150 and 1/200 in fore and hind limb myoblasts, respectively, which we believe does not represent a difference given the uncertainty of how many translocation events actually took place (we can only count those that give distinguishable fusion products). Our results in mouse myoblasts suggest that human myoblasts can be a cell of origin for the PAX3-FOXO1 translocation as they would provide a favorable environment for the translocation to occur, i.e. expression of both genes and spatial co-localization. It is curious that A-RMS is more frequent in the lower than in the higher extremities in humans, as reported by Neville and co-workers [45]. This apparent inconsistency with our mouse data might be explained by the possibility that humans may not have a difference in the distribution of PAX3 expression in the upper and lower extremities. In addition, the muscle mass and presumably the number of satellite cells in the lower extremities in humans is much higher than in the upper extremities, hence increasing the number of translocation-competent cells and frequency of translocation.
By using CRISPR-Cas9 nuclease we showed that targeted chromosome translocations could be induced with high efficiency. Unlike other approaches that have relied on induction of chromosome translocation using γ−irradiation, DSB-inducing chemicals, or the lymphoid cell-specific gene rearrangement machinery, CRISPR-Cas9 can be employed in any cell type. Due to its specificity the system is suitable for use in vivo in cell culture or in mice. Application of this system will greatly facilitate the development of mouse models that precisely recapitulate chromosome translocation-induced human diseases.
A complete list of E. coli strains used for this work can be found in S1 Protocol.
BAC clones RP24–391O12 (centromeric border of the 4.9 Mb syntenic region) and RP23–422I13 (telomeric border of the 4.9 Mb syntenic region) were purchased from the BACPAC Resource Center (BPRC), Children’s Hospital Oakland Research Institute in Oakland, California, USA (http://bacpac.chori.org).
The complete list of PCR Primers and oligonucleotides can be found in S1 Protocol.
A modified pNeoTKLoxP was recombineered into BAC RP23–422I13 (telomeric border of the syntenic region). In pNeoTKLoxP we replaced the wild type (wt) LoxP site downstream of the TK gene with the 511-ILoxP sequence (annealed oligonucleotides TK-511-ILoxP and TK-511-ILoxP-C). Then, via recombineering, we introduced the EM7 promoter upstream of the Neo gene. We therefore transformed electrocompetent DY380 Ecoli cells, containing the wtLoxPNeoTK-511-IloxP plasmid with the TK-EM7-Neo fragment (ends of the annealed oligonucleotides TK-EM7 and EM7-NeoC had been filled-in with Klenow DNA polymerase (Invitrogen) following the manufacturer’s protocol). For recombineering we followed the protocol posted on the Frederick National Laboratory for Cancer Research web site: http://ncifrederick.cancer.gov/research/brb/protocol/Protocol1_DY380.pdf. A short 5’-arm (annealed phosphorylated oligonucleotides 5-tel-s and 5-tel-s-C) was cloned downstream of 511-IloxP and a short 3’-arm was cloned upstream of wtLoxP (annealed phosphorylated oligonucleotides 3-tel and 3-tel-C).
A modified pBSLoxPTKhygro plasmid (kind gift from Drs. M. Roussel and F. Zindy, SJCRH) was recombineered into BAC RP24–391O12 (centromeric border of the syntenic region). In this construct we inserted a 511-ILoxP sequence upstream of the TK-promoter-Neo sequence. Since the activity of TK promoter in prokaryotic cells was sufficient to ensure Hygromycin B resistance, we did not introduce the bacterial EM7 promoter in this construct. A short 5’-arm (annealed phosphorylated oligonucleotides 5-cent-s and 5-cent-s-C) was cloned upstream of the 511-IloxP site and a short 3’-arm was cloned downstream of wtLoxP (annealed phosphorylated oligonucleotides 3-cent and 3-cent-C).
The pCL20c-MSCV-IRES-YFP vector backbone was generated by replacing GFP of pCL20c-MSCV-GFP [46] with I-YFP from MSCV-I-YFP [38]. hCas9 [23] was then cloned downstream of MSCV into pCL20c-MSCV-IRES-YFP. The mU6 fragment was generated by PCR using pSicoR-GFP (Addgene, Cambridge, MA, USA) and cloned downstream of hU6 in pLKO.1 (Addgene, Cambridge, MA, USA). The cassette containing the human and mouse U6 promoters (hU6 and mU6) followed by AgeI and EcoNI cloning sites was cloned upstream of the β-actin promoter of the modified pCL20c vector, containing the β-actin-puro cassette from pJ6.OMEGA.puro [47]. The spacer sequence of hU6 driven sgRNA starts with GG followed by 18 specific nucleotides from the target sequence, and mU6 driven sgRNA starts with GT followed by 18 specific nucleotides from the second target sequence (Fig. 4C). Synthetic ds-DNA fragments, coding Pax3_RH30 sgRNA and Foxo1_RH30 sgRNA were cloned into AgeI and EcoNI sites under control of hU6 and mU6 promoters of pCL20C-hU6-mU6-βact-puro, respectively (Fig. 4C,D).
pCL20C-MSCV-Luc2–2A-LgT was constructed by replacing IRES-YFP with a Luc2–2A-LgT cassette.
Lentivirus was produced as described in [46].
F12 (129SvJ-derived) embryonic stem (ES) cells were electroporated and selected for hygromycin B or G418 resistance using standard procedures. In short, 25–45 μg of linearized BAC DNA was electroporated into 2*107 ES cells followed by selection with 100 μg/ml Hygromycin B or 200 μg/ml G418. RP24–391O12-LoxP-hygro-TK was linearized with PI-SceI (NEB) and RP23–422I13-LoxP-Neo-TK was linearized with NotI (NEB). Drug resistant clones were picked after 7–9 days of selection. DNA from these clones was used for PCR analysis.
Screening of homologously recombined ES cell clones was done by PCR and qPCR. The presence of vector arms remaining on either side of the insert was detected by PCR with primers pTARBAC1–3F and pTARBAC1–3R for 3’-located sequences and pTARBAC1–5F and RP24–5R for 5’-located sequences. The “loss-of-native allele” assay was performed as described in [33] with some minor modification. For quantitative (q)PCR we used SYBR®Green PCR Master Mix (Applied Biosystems). qPCR was performed with the RP24-F and RP24-R primers to determine the copy number of the centromeric locus and the RP23-F and RP23-R primers to determine the copy number of the telomeric locus. Ratios between the copy numbers of the two loci were determined either by a standard dilution curve or by the Δct method.
A double targeted ES cell clone was electroporated with a Cre-expressing plasmid (pMC-CRE) using the Amaxa™ Mouse ES Cell Nucleofector™ Kit (Lonza, Germany) according to the manufacturer’s protocol. After 5 days of selection with 0.2μM of fialuridine (FIAU) (a kind gift of Bristol Myers) ES cells were collected for DNA isolation as a pool or as single clones. Cre-mediated inversion was detected by standard PCR using the RP24-F/RP23-F2, and RP24-R/RP23-R2 primer pairs.
For FISH analyses of Pax3 and Foxo1 co-localization and detection of t(1;3) reciprocal translocation we used BAC probes RP23–260F1 and RP24–391O12. For FISH analyses of targeted ES cells and Foxo1-inv+/+ fibroblasts we used BAC probes RP24–391O12 (centromeric border of the syntenic region) and RP23–422I13 (telomeric border of the syntenic region). BAC probes were labeled with nick translation using either Green (RP23–260F1and RP24–391O12) or Red (RP23–422I13) dUTP (Abbott Molecular). Probes were hybridized to metaphase and/or interphase cells either separately or as a 1:1 mixture in hybridization solution (50% formamide, 10% dextran sulfate, and 2X SSC). Slides were washed in 2X Saline-Sodium Citrate (SSC) buffer containing 50% formamide at 37°C for 5 minutes. Cells were counterstained with DAPI and analyzed using a Nikon E80i fluorescence microscope (Nikon) with a 100× oil immersion objective. Successfully targeted clones showed 2 native signals for the centromeric or telomeric targeted regions. Inverted chromosomes 3 appeared as two linked pairs of red and green signals on interphase cells, each pair representing one end of the inverted chromosome segment. Normal chromosomes 3 appeared as a single loosely paired red and green signal. One hundred interphase nuclei were scored for the presence of co-localization of Pax3/Pax7 and Foxo1 signals. Only nuclei with discernible red and green signals were scored. Fifty metaphase cells from CRISPR-Cas nuclease treated myoblasts were scored for the presence of cells containing the reciprocal translocation between Pax3 and Foxo1a.
Experimental details are provided in S1 Protocol. Position of primers, used for LD-PCR, gel electrophoresis of LD-PCR products and sequences flanking the breakpoint in the Rh30 cell line are shown in S3 and S4 Figs.
Cas9 induced translocation was detected by PCR of chromosomal DNA from 104 cells using Pax3-RH30F and Foxo1-RH30R primers.
For each qRT-PCR reaction we used RNA isolated from either 2.5×103 (data in Fig. 1) or 1.6×103 (data in S1 Fig.) cells. For each RT-PCR we used RNA isolated from 6.7×103 cells. For RT we used SuperScript III First-Strand Synthesis SuperMix (Invitrogen) with an equimolar mix of Pax3R primer and random hexonucleotides and performed the reaction following the manufacturer’s protocol. For the qPCR step we used TaqMan Gene Expression Master Mix (Applied Biosystems). Ratios between gene expression in different cell lines were determined by a standard dilution curve.
Myoblasts (5×106) were lysed in 0.5 ml CHAPS lysis buffer (40 mM HEPES [pH 7.4], 1 mM EDTA 120 mM NaCl, 10 mM sodium pyrophosphate, 10 mM β-glycerophosphate, 0.3% CHAPS, 50 mM NaF, 1.5 mM NaVO, 1 mM PMSF, and 1 tablet of EDTA-free protease inhibitors [Roche] per 10 mL solution) and freeze-thawed 3 times, followed by centrifugation at 20,000 ×g for 10 min at 4°C. After adding 2 μg anti-Pax3 antibody [48] or anti-Foxo1 (C29H4) Rabbit mAb (Cell Signaling) Pax3, Foxo1 and Pax3-Foxo1 were immunoprecipitated overnight at 4°C. Immunoprecipitated material was bound onto 10 μl protein G-coated Dynabeads (Invitrogen) for 90 minutes at 4°C, which were captured using a DYNA-Mag-2 magnet (Invitrogen), washed 4 times with CHAPS buffer, and removed from the beads by heating to 70°C in 1.25xLDS loading buffer (Invitrogen) in CHAPS and separated on pre-cast 4%–12% bis-tris polyacrylamide gels. Western-blotting was performed using the same anti-Foxo1 and anti-Pax3 antibodies.
To enrich for cells harboring the t(1;3), 104 of the Cas9/sgRNAs expressing fore limb myoblasts were evenly distributed over three 96-well plates (on average 30 cells per well). DNA from each cell pool was isolated and analyzed for the presence of t(1;3) translocation using PCR.
Libraries were generated from ~ 500 ng total RNA of the 1H3 (no Pax3-Foxo1) and 1G3 (64% Pax3-FOXO1) cell pools using the Illumina TruSeq Stranded mRNA Sample Preparation Kit. Libraries were sequenced on an Illumina HiSeq 2500 using paired-end 100 bp sequencing chemistry. Paired-end reads from RNA-seq were aligned to the following 4 database files using BWA (0.5.10) aligner: (1) the human GRCh37-lite reference sequence, (2) RefSeq, (3) a sequence file representing all possible combinations of non-sequential pairs in RefSeq exons, (4) AceView database flat file downloaded from UCSC representing transcripts constructed from human ESTs. The mapping results from (2) to (4) were aligned to human reference genome coordinates. In addition, they were aligned using STAR 2.3.0 to the human GRCh37-lite reference sequence without annotations. The final BAM file was constructed by selecting the best alignment among the five map events. We used HTSeq [49] to count the number of fragments that mapped to each gene (Gencode v 15), where each gene is considered as the union of all its exons. Then we normalized the count to FPKM (fragments per kilobase of exons per million fragments mapped) as the expression value of the gene. RNA-seq of both samples produced 55M reads each, with a 20X coverage of 43.561% of the exons in 1H3 and 43.992% of the exons in 1G3.
This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The Institutional Animal Care and Use Committee (IACUC) of St. Jude Children’s Research Hospital (Animal protocol number 209–100171) approved the protocol.
|
10.1371/journal.pntd.0007229 | Culex species diversity, susceptibility to insecticides and role as potential vector of Lymphatic filariasis in the city of Yaoundé, Cameroon | Culex species are widespread across Cameroon and responsible for high burden of nuisance in most urban settings. However, despite their high nuisance, they remain less studied compared to anophelines. The present study aimed to assess Culex species distribution, susceptibility to insecticide, bionomics and role in Lymphatic Filariasis (LF) transmission in the city of Yaoundé.
Mosquito collections were conducted from March to December 2017 using Centre for Disease Control light traps (CDC-LT), human landing catches (HLC) and larval collections. Mosquitoes were identified using morphological identification keys. Mosquitoes from the Culex pipiens complex were further identified using Polymerase Chain Reaction (PCR) to assess the presence of sibling species. Bioassays were conducted with 2–5 day-old unfed females to assess mosquito susceptibility to DDT, permethrin, deltamethrin and bendiocarb following WHO guidelines. Dead, control and surviving mosquitoes from bioassays were screened by PCR to detect the presence of knockdown resistance (kdr) alleles. Pools of mosquitoes were examined by PCR to detect the presence of Wuchereria bancrofti.
A total of 197,956 mosquitoes belonging to thirteen species were collected. The density of mosquito collected varied according to the collection methods, districts and seasons. Culex quinquefasciatus emerged as the most abundant and the only species of the Culex pipiens complex in Yaoundé. Culex species were found breeding in different types of breeding sites including polluted and unpolluted sites. All Culex species including Cx antennatus, Cx duttoni, Cx perfuscus and Cx tigripes were found to be highly resistant to permethrin, deltamethrin and DDT. Culex quinquefasciatus was also found to be resistant to bendiocarb. A high frequency of the West Africa kdr allele was recorded in resistant Cx. quinquefasciatus. Out of the 247 pooled samples of 25 Culex spp. examined for the presence of Wuchereria bancrofti, none was found infected.
The study confirms the high adaptation of Culex species particularly Culex quinquefasciatus to the urban environment and no implication of this species in the transmission of LF in Yaoundé Cameroon. Culex species predominance in urban settings highlight potential transmission risk of West Nile and rift valley fever in Yaoundé.
| Culex species are highly prevalent in both urban and rural settings in Cameroon and are responsible for high nuisance and transmission of pathogens such as Wuchereria bancrofti and arbovirus. Despite the important epidemiological role, that Culex could play, they are still less studied. The current study was conducted to assess Culex species distribution, susceptibility to insecticide and role in W. bancrofti transmission in the city of Yaoundé. Mosquito collection was conducted using three collection methods human landing catches, CDC light traps and larval collection. Once collected, mosquitoes were identified using morphological identification keys and PCR diagnostic tools. They were later processed to determine their infection status. Bioassays with Culex females of 2 to 5 days old were conducted to determine their susceptibility level to different insecticide families. Culex quinquefasciatus emerged as the most abundant species. Up to 13 different culicine species were recorded. Culex species were recorded to be highly resistant to DDT, permethrin and deltamethrin. A high frequency of the West Africa kdr allele was recorded. No mosquito was detected to be infected by LF.
The study confirms the need for further xenomonitoring activities in order to control the risk of outbreaks due to Culex mosquitoes in the city of Yaoundé.
| Culex species are the most widespread mosquito species across the world [1]. They are known to be highly opportunistic feeding on both humans and animals, a behaviour which increases their potential to transmit zoonotic diseases and makes them important threat to public health [2]. Culex have over decades adapted to human made habitats [3]. One of the most important group in the Culex genus is Culex pipiens complex which comprises six members: Cx. quinquefasciatus Say, Cx. pallens Coquillet, Cx. australicus Dobrotworsky & Drummond, Cx. globocoxitus Dobrotworsky, Cx. pipiens Linneaus and Cx. molestus Forskll [4, 5]. Species of the Cx. pipiens complex particularly Cx. quinquefasciatus are widespread and predominant in the urban environment notably in Africa where suitable environmental conditions created by rapid unplanned urbanization is contributing to their proliferation [6–9]. Culex quinquefasciatus can be found in all types of water collections including temporary or permanent stagnant water bodies such as drains, septic tanks, wet pit latrines, organically polluted sites, puddles [10] and has emerged as the most common mosquito species in major African cities [11–13]. In addition to nuisance that Culex species could induce, they also transmit diseases such as Japanese and Saint Louis encephalitis, Rift valley fever, West Nile Virus and lymphatic filariasis (LF) [14, 15]. The later caused by the parasite Wuchereria bancrofti is largely prevalent in Asia and sub-Saharan Africa and is consider as one of the leading causes of long term disability in the World [16–18]. In Cameroon, Lymphatic filariasis is considered to be endemic with mean prevalence level (ICT>1%) estimated at 3.3% countrywide [19]. LF is among the neglected tropical diseases targeted for elimination by the World Health Organization by 2020 using mass drug administration (MDA) [20]. Although direct implication of Culex species in the transmission of LF in West and Central Africa is still not well documented [21, 22], in East Africa, Culex species particularly Cx. quinquefasciatus is known to have a major role in LF transmission [23, 24]. With changing climate associated to increased traffic between East and West African countries and rapid expansion of this species in urban settings, it is becoming crucial to assess the role of Culex species in the transmission of diverse diseases. In most cities in Cameroon Culex are the main species causing the highest nuisance in the population. Household survey conducted in the cities of Douala and Yaoundé indicated that in addition to treated nets, tools such as insecticide spray, coils, screen are permanently used by urban dwellers to fight against mosquito nuisance [25, 26]. In the cities of Douala and Yaoundé, high pyrethoid resistance in An. gambiae populations was reported [27–29], whereas for Culex species there is still not enough data on their bionomic in the urban environment. Data on species composition, spatial distribution, susceptibility to insecticides and implication in diseases transmission are all lacking. This information is of paramount importance in the perspective of integrated vector management and insecticide resistance management [30]. Also, understanding the bionomic and distribution of Culex species could enable understanding the epidemiology of diseases that they transmit and to establish sustainable surveillance and control programmes. The present study assesses the distribution, susceptibility status to insecticides and epidemiological role of Culex species before the implementation of a larval control trial in the city of Yaoundé.
Study site- The study was conducted in Yaoundé (03°52’N; 11°31’E), the capital city of Cameroon from March to December 2017. The city has a population estimated at 2.8 million inhabitants. Yaoundé belongs to Guinean subequatorial climate type, characterized by four distinct seasons: the short rainy season (Mars-June), the short dry season (June-July), the long rainy season (August-November) and the long dry season (November-February). The city receives annually over 1600 mm of rainfall and the annual average temperature is 24°C. Yaoundé is located about 750 m above sea level and surrounded by many hills. Although occurring at very low endemicity, human infection by Wuchereria bancrofti was estimated at 2.3% during surveys conducted between 2009–2010 in Yaoundé and it surroundings [19].
The study was conducted under the ethical clearance N° 2016/11/832/CE/CNERSH/SP delivered by the Cameroon National Ethics Committee for Research on Human Health (CNERSH) Ref N°D30-172/L/MINSANTE/SG/DROS/TMC of 4 April 2017. For human landing catches all adult men who took part in the collection signed a written informed consent form before being enrolled in the study as recommended by the validated protocol and were given free malaria prophylaxis.
Mosquito’s collection and breeding sites characterization- Adult and immature stages of Culicine mosquitoes were sampled in 32 districts of Yaoundé. Culicine collections were undertaken in the context of a big survey intended to assess mosquito distribution and malaria transmission pattern in the city of Yaoundé before a larval control trial and will allow in the future additional analysis with more data. Adult mosquitoes were collected using CDC light traps (CDC-LTs) and Human Landing Catches (HLCs) from 7pm to 6am. All potential larval breeding sites were inspected and positive sites (with at least one Culicine larvae or pupae) recorded. Three dips were undertaken for small breeding sites of less than 1 m2; and 5 to 10 dips were undertaken in breeding sites of more than 1m2. The average larval density (N) was estimated. Once collected larvae were classified according to their stages: early instars larvae (L1&L2) and late instars (L3&L4 and pupa). Other parameters measured included the type of breeding sites sampled (stagnant water pools, gutters, well, tyre print, footprint, pit latrine….), depth, the status organically polluted or not, the distance to the nearest house, the presence/absence of predators, the proportion of water surface covered by vegetation or algae.
Larvae collected were kept in plastic containers and brought to the insectary for rearing. After emergence, adult mosquitoes were identified to species level under a binocular magnifying glass using morphological identification keys [31–33]. For mosquitoes collected using either CDC-LTs or HLC, a subsample of 50 culicine specimens per district was randomly selected for identification during each collection month. All mosquitoes collected were stored at -20°C for further molecular analyses.
Susceptibility tests to insecticides-Bioassays were performed with 2–5 days old females emerging from larval collection. Mosquitoes were tested against permethrin 0.75%, DDT 4%, bendiocarb 1% and deltamethrin 0.05% following WHO guidelines [34].
For each test, batches of 25 mosquitoes per tube were exposed to impregnated papers for 1 hour. The number of mosquitoes knocked down by the insecticide was recorded every 10 minutes during exposure. After exposure, mosquitoes were fed with a 10% glucose solution and the number of dead mosquitoes was recorded 24 hours post-exposure. Mosquitoes used as controls were exposed to untreated papers. The mortality rates were corrected using the Abbot formula [35] whenever the mortality rate of the controls was between 5 and 20%. Susceptibility and resistance levels were assessed according to WHO criteria [34]. At the end of the assay, mosquitoes were classified into three different groups: 98%-100% mortality indicates susceptibility, 80%-97% mortality suggests possible resistance that needs to be confirmed, <80% mortality suggests resistance.
The study objective was to assess culicine species distribution, bionomic and potential role in W. bancrofti transmission in the city of Yaoundé. High Culicine species diversity was recorded with up to 13 species collected. Culex species were the most prevalent and this was consistent with previous studies conducted in Cameroon and across Africa indicating the high adaptation capacity of species of this genus particularly Cx. quinquefasciatus to the urban environment [27, 41–44]. The diversity of culicine species recorded could result from the presence of different landscapes across the city of Yaoundé made up of an alternation of highland and marshland covered with vegetation and exploited for agriculture, lakes invaded by vegetation, and rural environment. It is still unknown whether there is an intense competition between culicine species sharing similar habitats. Species such as Cx. tigripes larvae are known to be predators for early instars of different species. Culex quinquefasciatus emerged after molecular analysis, as the sole member of the Cx. pipiens complex in Yaoundé; its presence was consistent with the known distribution of members of the complex [37].
Species diversity and abundance were all found to vary according to collection methods and seasons. High species diversity was recorded using CDC-LT compared to HLC or larval collection and reflects the high efficiency of CDC-LT method for collecting culicines. The use of CDC-LT has now become common for sampling mosquito populations across the world and has been shown to be particularly effective for sampling Culex mosquitoes [27, 45]. This tool was rather found to underestimate anophelines densities [27, 45, 46]. Both HLC and CDC-LT techniques were used because there was so far no available data on the efficiency of CDC-LT for collecting Culex species from Yaoundé.
Seasonal variations in species composition was detected for mosquitoes collected from breeding habitats, however, no similar trend was detected for mosquitoes collected using CDC-LT or HLC. This likely suggest different breeding habitats preference for culicine species at different periods of the year or the influence of physico-chemical parameters [47, 48] or xenobiotics selection [49] on Culex species distribution. Cx. quinquefasciatus larvae were found to be highly prevalent in polluted sites. It is likely that females of Culex species are more attracted by oviposition cues released by the microbial fauna in this type of habitats. In addition, these habitats are rich in nutrients and could thus reduce competition for resources between species. This could also be because mosquitoes in polluted sites are also frequently exposed to intensive selective pressure induced by pollutants and xenobiotics [27, 50–52], different strategies were reported to promote Culex species adaptation to different ecological constraints. This include the development of resistance or detoxification mechanisms to a large set of insecticides and xenobiotics [53–55], the capacity for eggs to resist desiccation [56] and development of cuticle resistance in larvae [3, 57, 58]. Several Culex species including Cx. quinquefasciatus, Cx. antennatus, Cx. duttoni were found to display resistance to DDT, permethrin and deltamethrin. This is the first time that insecticide resistance in different Culex species is documented in Cameroon. The level of pyrethroid resistance was similar to data recorded for An. gambiae populations in the city of Yaoundé [59, 60]. In addition to the fact that Culex species are known to breed in polluted environment and could thus be affected by xenobiotics selection, the high level of resistance recorded could also result from increased use of LLINs for malaria vector control and pesticides use in agriculture in the city of Yaoundé [27, 61]. Our study also suggested the presence of kdr allele in Cx. quinquefasciatus populations. It is likely that resistance in Culex species is sustained by both kdr mutations and other mechanisms such as the metabolic detoxification machinery [62].
The present study also permitted to evaluate the role of Culex species in LF transmission after mass drug administration (MDA) scale up in Cameroon. Culex quinquefasciatus is the predominant vector of LF in both urban and rural settings in East Africa [3, 23] but less so in Central and West Africa. However, with potential gene flow and changing climate, one cannot rule out that Cx. quinquefasciatius in Central Africa such as in Cameroon may also emerge as LF vector. Furthermore, because of the rapid expansion and predominance of this species in Cameroon cities, it’s potential implication in LF transmission in Yaoundé was examined. Analysis conducted with pool samples of Culex mosquitoes recorded no infection. In Cameroon LF is considered to be endemic with prevalence rates varying from 1 to 8% [19, 63]. It is likely that the prevalence of parasite may have decreased over years due to the implementation of mass drug administration of ivermectin and abendazole to the population since 2009 [19]. So far, five to six rounds of MDA have been successfully conducted in endemic settings across the country and interruptions of LF transmission have been documented in some parts of the country [64]. The fact that only Culex species were screened during this study could have limited the capacity of detecting any ongoing transmission since mosquito species such as An. gambiae and An. funestus are also good vectors of LF [3, 23]. Another important dimension which could explain the absence of W. bancrofti infection in Culex is that the area may have not been endemic for W. bancrofti before the introduction of MDA. Recent studies conducted in Cameroon and DRC suggested that the perceived endemicity of LF established by ICT test in the central African region could result from the presence of Loa filariasis which cross react to the ICT tests which was used to detect W. bancrofti in Central Africa, leading to false positivity [64–66]. During the last decade, several arboviral diseases such as chikungunya, dengue, yellow fever, West Nile, Sindbis, Tahyna, O’nyong-nyong and spondweni virus have been reported in circulation in human adults in both urban and rural settings [67–70]. With the rapid distribution of Culex species in the urban environment, the potential role that these species could play in spreading of these arboviral diseases deserves further consideration.
The present study confirms high abundance of Cx. quinquefasciatus in the city of Yaoundé and high insecticide resistance in most Culex species populations. The study also suggests no transmission of W. bancrofti by Culex species in Yaoundé. In Cameroon, apart from malaria vectors, surveillance activities are not regularly conducted on other vectors of diseases because of lack of funding or technical capacities for these activities. In this context, combining surveillance activities of malaria vectors with other culicine species and strengthening capacities of medical entomologists on taxonomy, sampling, processing and calculation of key entomological indicators for endemic vector borne diseases could be cost effective and will enable better understanding of the distribution and epidemiology of various diseases. This could lead to the establishment of sustainable surveillance systems.
|
10.1371/journal.pmed.1002081 | Duration of Adulthood Overweight, Obesity, and Cancer Risk in the Women’s Health Initiative: A Longitudinal Study from the United States | High body mass index (BMI) has become the leading risk factor of disease burden in high-income countries. While recent studies have suggested that the risk of cancer related to obesity is mediated by time, insights into the dose-response relationship and the cumulative impact of overweight and obesity during the life course on cancer risk remain scarce. To our knowledge, this study is the first to assess the impact of adulthood overweight and obesity duration on the risk of cancer in a large cohort of postmenopausal women.
Participants from the observational study of the Women’s Health Initiative (WHI) with BMI information from at least three occasions during follow-up, free of cancer at baseline, and with complete covariate information were included (n = 73,913). Trajectories of BMI across ages were estimated using a quadratic growth model; overweight duration (BMI ≥ 25 kg/m2), obesity duration (BMI ≥ 30 kg/m2), and weighted cumulative overweight and obese years, which take into account the degree of overweight and obesity over time (a measure similar to pack-years of cigarette smoking), were calculated using predicted BMIs. Cox proportional hazard models were applied to determine the cancer risk associated with overweight and obesity duration. In secondary analyses, the influence of important effect modifiers and confounders, such as smoking status, postmenopausal hormone use, and ethnicity, was assessed. A longer duration of overweight was significantly associated with the incidence of all obesity-related cancers (hazard ratio [HR] per 10-y increment: 1.07, 95% CI 1.06–1.09). For postmenopausal breast and endometrial cancer, every 10-y increase in adulthood overweight duration was associated with a 5% and 17% increase in risk, respectively. On adjusting for intensity of overweight, these figures rose to 8% and 37%, respectively. Risks of postmenopausal breast and endometrial cancer related to overweight duration were much more pronounced in women who never used postmenopausal hormones. This study has limitations because some of the anthropometric information was obtained from retrospective self-reports. Furthermore, data from longitudinal studies with long-term follow-up and repeated anthropometric measures are typically subject to missing data at various time points, which was also the case in this study. Yet, this limitation was partially overcome by using growth curve models, which enabled us to impute data at missing time points for each participant.
In summary, this study showed that a longer duration of overweight and obesity is associated with an increased risk of developing several forms of cancer. Furthermore, the degree of overweight experienced during adulthood seemed to play an important role in the risk of developing cancer, especially for endometrial cancer. Although the observational nature of our study precludes inferring causality or making clinical recommendations, our findings suggest that reducing overweight duration in adulthood could reduce cancer risk and that obesity prevention is important from early onset. If this is true, health care teams should recognize the potential of obesity management in cancer prevention and that excess body weight in women is important to manage regardless of the age of the patient.
| Excess weight has become the leading risk factor for disease burden in high-income countries and has been offsetting or surpassing the decreasing disease burden attributable to tobacco smoking. Excess weight has been linked to the development of several types of cancer.
To date, most studies exploring the relationship between excess weight and cancer risk looked at cross-sectional exposure information on overweight and obesity, i.e., height and weight measured at one point in time. Insights into the dose-response relationship of the cumulative impact of overweight and obesity during the life course on cancer risk remain scarce.
This study examined how the timing, duration, and intensity of overweight and obesity during adulthood impacts on cancer risk, taking into account important information on other factors related to obesity, such as physical activity, diet, smoking, hormone use, and diabetes history.
A total of 73,913 women were included in the study, and 6,301 obesity-related cancers were diagnosed during a mean follow-up of 12.6 y. About two-thirds of all included women were ever overweight or obese during adulthood.
The study found that being overweight for a longer duration during adulthood significantly increased the incidence of all obesity-related cancers by 7% (for every ten-year increase in adulthood overweight duration), of postmenopausal breast cancer by 5%, and of endometrial cancer by 17%. After adjusting for the intensity of overweight (that is, how overweight individuals were), these figures rose to 8% for postmenopausal breast cancer and 37% for endometrial cancer (for every ten years spent with body mass index ten units above normal weight).
How much of their adult lives women are overweight and how overweight they are play important roles in cancer risk. This finding highlights the importance of obesity prevention at all ages and from early onset.
| High body mass index (BMI) has become the leading risk factor for disease burden in high-income countries and has been offsetting or surpassing the decreasing disease burden attributable to tobacco smoking [1]. In the US, currently about 70% of all adults are considered overweight (BMI ≥ 25 kg/m2) and 36% obese (BMI ≥ 30 kg/m2, according to WHO classification), making the US one of the countries with the highest prevalence of obesity [2]. This development has also been found to contribute substantially to the country’s poor international ranking in longevity [3,4]. Continuing increases in obesity prevalence seen over the past decades remain of great public health concern.
Overweight and obesity have been associated with an increased risk of and mortality from cancer and other chronic diseases such as type 2 diabetes and cardiovascular disease [5]. In the US, more than 100,000 cancer cases were attributed to high BMI in the year 2012 alone [6]. Cancers previously linked to high BMI include postmenopausal breast cancer, adenocarcinoma of the esophagus, and pancreatic, colorectal, renal, endometrial, ovarian, and gallbladder cancer [7,8]. To date, most studies of these associations have used cross-sectional exposure information on BMI, i.e., BMI measured at one point in time and typically obtained at recruitment. Yet, insights into the dose-response relationship of the cumulative impact of overweight and obesity during the life course on cancer risk remain scarce. While age-dependent and cumulative effects of weight change have previously been reported to affect the risk of postmenopausal breast cancer [9–11], only a few studies have investigated the association between overweight duration and cancer outcomes [9,12]. It is thus still unclear how different exposure durations of overweight and obesity are associated with cancer development.
In this study, we assessed the impact of adulthood overweight and obesity duration and cumulative intensity on cancer risk in US women, using data from the observational study cohort of the Women’s Health Initiative (WHI). In secondary analyses, we investigated the influence of important effect modifiers and confounders, such as smoking status, postmenopausal hormone use, and ethnicity.
WHI procedures and protocols were approved by the institutional review boards at each participating institution, and all participants provided written informed consent.
The WHI is a large, multi-center prospective cohort study of postmenopausal women. The WHI was designed to have a clinical trial arm and an observational study cohort [13], and recruited postmenopausal women aged 50–79 y at 40 clinical centers nationwide between October 1, 1993, and December 21, 1998, to be followed for the development of diseases that are the most common causes of death, including cardiovascular disease and cancer. Details of the study design and methods have been published elsewhere [13–16]. In total, 93,676 women were enrolled in the observational study, and 68,132 were enrolled in the clinical trial arm (n = 161,808) [16]. For this study, we included all participants from the observational study cohort, except those who reported cancer prior to or at baseline or without data on cancer history (n = 12,827) and women with incomplete follow-up information (n = 411) (Fig 1).
Information on BMI for was obtained from retrospective self-reports at baseline for ages 18, 35, and 50 y, from weight and height measurements at baseline and at 3-y follow-up, and from self-reports at follow-up years 4–8. BMI was calculated by dividing weight in kilograms by height in meters squared. For inclusion in the study, women were required to have valid body weight information from at least three occasions and a valid baseline measurement of body weight and height.
Covariates included baseline information on age (continuous), ethnicity (six categories), education (11 categories), energy intake (in kilocalories, continuous), diet quality (Alternate Mediterranean Diet Score [17], nine categories), physical activity (frequency of moderate exercise per week, four categories), smoking status (never/past/current smoker), and outcome-specific covariates—for pancreatic cancer: diabetes (never/ever diabetic); for colon cancer: diabetes and red meat intake (servings/day); for postmenopausal breast, ovarian, and endometrial cancer: postmenopausal hormone use (never/past/current user), number of full-term pregnancies (continuous), age at first term pregnancy (<20, 20–24, 25–29, 30–34, 35–39, 40–44, 45+ y), and age at menopause (continuous).
Adjudication and outcome ascertainment for the WHI have been described elsewhere [18]. In brief, all outcomes were self-reported annually in the observational study arm. Only first invasive cancers confirmed by adjudication were included in these analyses. All analyses were conducted for the following cancers with convincing evidence of a positive relation with excess BMI [7,8]: colon, rectum, liver, gallbladder, pancreas, postmenopausal breast, endometrium, ovary, kidney, and thyroid, as well as the combination of all of these, termed “all obesity-related cancers.”
The analysis was carried out in two steps. In the first step, BMI was modeled across all ages using a quadratic growth model with random intercept and random slope, incorporating all available BMI information from all included participants [19]. No random coefficient was included for the quadratic term.
BMI information in the year preceding cancer diagnosis for those who developed an invasive malignancy was excluded. This model was developed and adjusted in a stepwise manner by adding ethnicity, education, and baseline physical activity, smoking status, energy intake, and diet quality score to the model. Using this approach, we allowed individuals to have their own BMI trajectory. Using the full model, BMI was predicted from age 18 y until the age at study exit for every cohort member. The obtained predicted age-specific BMI data were then used to compute the following parameters: overweight (BMI ≥ 25 kg/m2) and obesity (BMI ≥ 30 kg/m2) duration in years and weighted cumulative overweight years (OWY) and obese years (OBY). OWY and OBY were calculated by multiplying the duration of overweight and obesity in years by the difference (in BMI units) above normal BMI (≥25 kg/m2) for overweight and above overweight (≥30 kg/m2) for obesity for each age. This allowed us to take the degree of each participant’s overweight and obesity over time into account. An individual with a BMI of 35 kg/m2 for 10 y would thus contribute 100 (= 10 × [35 − 25]) OWY and 50 (= 10 × [35 − 30]) OBY units. The merit of this approach as a better predictor than cross-sectional BMI information alone for many obesity-related outcomes has been described earlier [20]. Overweight and obesity duration were assessed per 10-y increments and OWY and OBY per 100 units.
In the second step of the analysis, Cox proportional hazard models with time since enrollment as the underlying time metric were fitted to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for the relationship between BMI overweight/obesity duration, OWY/OBY, and the risk of developing specific cancers. Overweight/obesity duration and OWY/OBY were treated as continuous, time-varying covariates. Participants were censored at study exit due to death, loss to follow-up, any cancer diagnosis, or end of follow-up (August 29, 2014), whichever occurred first. Three models were fitted for all outcomes with adjustments for several established risk factors for obesity-related cancers. In model 1, adjustment was made for age. Model 2 was additionally adjusted for ethnicity and education. In model 3, smoking status, physical activity, energy intake, and diet quality score were additionally introduced into the model. Cancer-specific adjustments were made in model 4, introducing age at first birth, age at menopause, parity, and postmenopausal hormone use for breast, endometrial, and ovarian cancer; red meat intake and diabetes status for colon cancer; and diabetes status for pancreatic cancer. Possible interactions were tested by fitting models with and without the interaction term, and corresponding p-values were computed from likelihood ratio tests comparing the two models. In order to assess the potentially nonlinear dose-response relationships between overweight duration, OWY, and cancer risk, we used restricted cubic splines with three knots to model those relationships [21]. Nonlinearity was evaluated by testing the null hypothesis that the coefficients of the splines were the same.
In secondary analyses, we assessed a priori interactions by stratifying by postmenopausal hormone use, hysterectomy and/or oophorectomy, ethnicity, diabetes, and smoking status. We also tested the robustness of our findings by predicting BMI trajectories using self-reported height and weight assessments only.
All analyses were carried out using Stata 13.
The final sample for analyses included 73,913 women with a mean follow-up of 12.6 y (standard deviation [SD] = 5.1), during which a total of 6,301 invasive obesity-related cancer cases occurred. Out of all included study participants, 40% (n = 29,770) were never overweight (BMI < 25 kg/m2) during their adult life (age 18 y until study exit), and 60% (n = 44,143) were ever overweight (BMI ≥ 25 kg/m2), almost half of whom (n = 19,654) were also ever obese (BMI ≥ 30 kg/m2) (Table 1). These values were estimated based on BMI data from on average >9 occasions (both measured and self-reported) per study participant. Women who were ever overweight were on average overweight for 31.3 y; those who were ever obese were, on average, obese for 20.6 y. Compared with women who were never overweight, those who were ever overweight or obese were slightly younger at baseline, had a lower education, and were more likely to be African-American (Table 1). While no differences existed in smoking status, women who were ever overweight or obese were less physically active, consumed more calories, had a lower diet quality score, and more often reported ever being diabetic than women who were never overweight. Women who were never overweight were more likely to be using postmenopausal hormones and less often had a hysterectomy.
A longer overweight duration significantly increased the risk of all obesity-related cancers combined (multivariable-adjusted HR per 10-y increment: 1.07, 95% CI 1.06–1.09) (Table 2). The strongest associations were observed for endometrial (HR 1.17, 95% CI 1.12–1.22) and kidney cancer (HR 1.16, 95% CI 1.07–1.26), while no significant associations were found for rectal, liver, gallbladder, pancreatic, ovarian, and thyroid cancer. When taking into account the degree of overweight over time, the association became stronger for all obesity-related cancers combined, with an increase in risk of 12% for every 100 units of OWY. For endometrial cancer, the HR per 100 OWY was 1.37 (95% CI 1.29–1.46). The results for obesity duration were even more pronounced and showed a significant association for all obesity-related cancers combined (multivariable-adjusted HR per 10-y increment: 1.10, 95% CI 1.08–1.12) and individually for colon, breast, endometrial, and kidney cancer, with HRs for every 10 y of obesity ranging from 1.07 (95% CI 1.04–1.10) for breast cancer to 1.23 (95% CI 1.18–1.28) for endometrial cancer. Risks associated with OBY were similar. Further cancer-specific adjustments (for pancreatic cancer: diabetes; for colon cancer: diabetes and red meat intake; for postmenopausal breast, ovarian, and endometrial cancer: postmenopausal hormone use, number of full-term pregnancies, age at first term pregnancy, and age at menopause) did not materially change these results (Table A in S1 Appendix).
A dose-response relationship with increasing overweight duration was found for almost all cancers (Fig 2). The risk of endometrial cancer associated with increasing overweight duration rose in an exponential fashion, but became statistically significant only after 26 y of being overweight. This threshold effect disappeared when the degree of overweight over time was taken into account (Fig A in S1 Appendix). In contrast, the association of overweight duration with colon and postmenopausal breast cancer was linear.
When stratifying the analyses by postmenopausal hormone use, for which statistically significant interactions were found, stronger associations of breast and endometrial cancer with all measures of overweight and obesity duration emerged in women who were never or past users (Table B in S1 Appendix). While the risk of endometrial cancer was statistically nonsignificant in current users, the risk of breast cancer was still elevated in this group, although at a much lower level than in never users (Fig 3). In line with these findings, the incidence rate of postmenopausal breast cancer was high in women who were ever overweight, regardless of their hormone use (Fig B in S1 Appendix). In contrast, among women who were never overweight in adulthood, current hormone users had a higher incidence rate of postmenopausal breast cancer than never or past postmenopausal hormone users. This pattern was somewhat different for endometrial cancer, where the incidence rate in ever overweight current users was similar to that in women who were never overweight and never or past hormone users.
Similar risks of postmenopausal breast, endometrial, and ovarian cancer associated with increasing overweight and obesity duration and intensity were found in women who had ever had a hysterectomy or oophorectomy compared to women who never underwent either of these procedures (Table C in S1 Appendix). Cancer risk related to increasing overweight and obesity duration was slightly higher in non-Hispanic white women than in African-American women for breast cancer (Table D in S1 Appendix), more pronounced in women who ever reported diabetes for colon cancer (Table E in S1 Appendix), and highest in current smokers compared to past and never smokers for colon and pancreatic cancer (Table F in S1 Appendix). Additional analyses using only self-reported BMI assessments confirmed the main findings (Table G in S1 Appendix).
Many previous studies have reported on the association between obesity and cancer, but, to our knowledge, this study is the first to assess the impact of adulthood overweight and obesity duration on the risk of various types of cancer in a large cohort of postmenopausal women. We found that longer durations of overweight and obesity were significantly associated with an increased incidence of obesity-related cancers, postmenopausal breast cancer, and colon, endometrial, and kidney cancer. Taking into account the intensity of overweight over time further increased the risks, and clear dose-response relationships were found. These findings are in line with previous studies on other chronic diseases, which showed that obesity duration is an important and independent predictor of type 2 diabetes [11], cardiovascular disease [10], and all-cause mortality [9]. Underlying biological mechanisms explaining these associations include a higher risk of developing hypertension and insulin resistance [10,11]. Earlier and long-term exposure to overweight and obesity may also increase the risk and severity of chronic inflammation, oxidative DNA damage, and alterations in endogenous hormone metabolism, three key mechanisms that have been found to be associated with increased risk of cancer [22].
We also observed that the risks of postmenopausal breast and endometrial cancer related to overweight and obesity duration were modified by postmenopausal hormone use and were largely attenuated or even eliminated among postmenopausal hormone users. Very high levels of estrogen in women using postmenopausal hormones have been postulated to obscure the effect of obesity and might explain this finding. While similar effect modification was reported in earlier analyses of the WHI observational study [23,24], in the Million Women Study [25–27], and in several other prospective cohorts [28–30], a recent analysis of the WHI clinical trial suggested a remaining positive association between obesity and postmenopausal cancer risk, independent of postmenopausal hormone use [31]. This finding was attributed to the often self-reported height, weight, and hormone use information in observational studies as well as outcome ascertainment bias. Strong associations found for endometrial cancer and its exponential dose-response relationship with overweight duration and intensity confirm risk patterns observed in previous studies [32,33].
Long-term follow-up and many repeated measurements of weight and height are necessary in order to fully capture lifetime overweight and obesity exposure and to quantify corresponding health effects. In our study, we used data from a large cohort of women with numerous repeated measurements and self-reports of weight and height for each individual, and applied a novel approach to estimate adulthood overweight and obesity duration. By modeling individual BMI trajectories across age using growth curves, all observed BMI information for an individual could be used, and unbalanced repeated measurements accommodated. Compared to traditional approaches to handling incomplete time-dependent data, such as linear interpolation and last observation carried forward methods, which may not adequately model BMI trajectories, growth models can incorporate any covariate data that might be important when predicting an individual’s BMI during life course. Using this approach, we were able to quantify the risk associated with both the duration and the intensity of overweight and obesity, as opposed to most previous studies, which focused on cross-sectional BMI information only.
Our study has limitations because of variation in how BMI information was collected and used. Some of this information was obtained from retrospective self-reports, which might be subject to measurement error. Yet, including only self-reported height and weight information when modeling BMI trajectories only marginally altered the results. Data from longitudinal studies with long-term follow-up and repeated anthropometric measures are typically subject to missing data at various time points due to periodical measurements, which was also the case in this study. Using growth curve models, we were able to impute data at missing time points for each participant. In a previous small-scale simulation study focusing on the effects of measurement errors, it was reported that using predicted biomarker values from growth curve models as a time-dependent covariate in Cox regression models—instead of applying a more naïve approach based on periodically observed biomarker values—yielded much less biased estimates for the association between the biomarker and the clinical outcome, especially when the variance in measurement errors was large [34].
BMI is not an ideal measure of body fatness as it does not differentiate tissue type (fatty, lean, bone). It has been suggested that other anthropometric measures such as waist circumference or waist-to-hip ratio better predict obesity-related health outcomes than BMI does [35,36]. These measures were, however, not available in a repeated fashion in the WHI. Additionally, BMI has limitations with regard to its application across different race/ethnic groups and ages [37]. In the US, African-American and Hispanic women have been reported to be more likely to be obese than non-Hispanic white and Asian women, while Asian women have been shown to have a higher percentage of body fat than non-Hispanic white women at similar BMI levels [38]. It might thus be appropriate to use different BMI thresholds for overweight and obesity in different race/ethnic groups. In our study, the risks for postmenopausal breast cancer associated with a longer overweight duration were higher in non-Hispanic white women than in African-American women. The fairly small number of women with an ethnicity other than non-Hispanic white did not allow further investigation of this issue. Furthermore, it is important to note that BMI serves as a proxy for many internal physiological processes that are the actual correlates of cancer development and that these also depend on many other lifestyle and environmental factors. Any residual confounding can therefore not be ruled out.
In summary, this study showed that the risk of cancer associated with overweight and obesity compounds over time, and a longer duration of overweight and obesity during adulthood is associated with increased risks of several cancers. Furthermore, not only the duration but also the degree of overweight seems to play an important role in the risk of developing cancer, especially for endometrial cancer. Although the observational nature of our study precludes inferring causality or making clinical recommendations, our findings suggest that reducing overweight duration in adulthood could reduce cancer risk. If this is true, health care teams should recognize the potential of obesity management in cancer prevention and that excess body weight in women is important to manage regardless of the age of the patient.
|
10.1371/journal.ppat.1001307 | Human Macrophage Responses to Clinical Isolates from the Mycobacterium tuberculosis Complex Discriminate between Ancient and Modern Lineages | The aim of the present study was to determine whether there is a correlation between phylogenetic relationship and inflammatory response amongst a panel of clinical isolates representative of the global diversity of the human Mycobacterium tuberculosis Complex (MTBC). Measurement of cytokines from infected human peripheral blood monocyte-derived macrophages revealed a wide variation in the response to different strains. The same pattern of high or low response to individual strains was observed for different pro-inflammatory cytokines and chemokines, and was conserved across multiple human donors. Although each major phylogenetic lineage of MTBC included strains inducing a range of cytokine responses, we found that overall inflammatory phenotypes differed significantly across lineages. In particular, comparison of evolutionarily modern lineages demonstrated a significant skewing towards lower early inflammatory response. The differential response to ancient and modern lineages observed using GM-CSF derived macrophages was also observed in autologous monocyte-derived dendritic cells and murine bone marrow-derived macrophages, but not in human unfractionated peripheral blood mononuclear cells. We hypothesize that the reduced immune responses to modern lineages contribute to more rapid disease progression and transmission, which might be a selective advantage in the context of expanding human populations. In addition to the lineage effects, the large strain-to-strain variation in innate immune responses elicited by MTBC will need to be considered in tuberculosis vaccine development.
| Mycobacterium tuberculosis is a long-standing human pathogen spread by aerosol transmission between individuals interacting in close social groups. It can be anticipated that the evolution of M. tuberculosis will parallel the evolution of human societies, and the phylogeny as determined by whole genome sequencing of clinical isolates is indeed consistent with emergence of the pathogen with modern humans in Africa and its subsequent dissemination along routes of human migration and trade. The present study was designed to test the hypothesis that the genetic diversity of M. tuberculosis isolates would be reflected in a corresponding diversity in their biological properties. In particular, we explored the interaction of different isolates with the innate immune system, which plays important contrasting roles in initial resistance to infection and in disease transmission. We observed a difference in the innate immune response when we compared isolates belonging to “modern” lineages that have evolved amongst high-density populations in regions of recent massive demographic expansion, with isolates belonging to “ancient” lineages selected in older low-density human populations. Our results provide insights into host-pathogen co-evolution and into fundamental mechanisms underlying the pathogenesis of M. tuberculosis.
| High-throughput sequence analysis has allowed reconstruction of the evolution of human Mycobacterium tuberculosis Complex (MTBC), differentiating the bacteria into six main phylogenetic lineages [1], [2]. Three lineages, including two whose members are known as M. africanum [3], which branched off from a common ancestor at an early stage of evolution, are referred to as evolutionarily “ancient” lineages; three separate evolutionarily “modern” lineages diverged at a later time point [1]. It is proposed that the branches reflect the history of human migration out of Africa, with the current geographic distribution of the different lineages being determined by the expansion and migration of their corresponding host populations.
This phylogeny provides a rational framework to assess whether the genotypic diversity of MTBC is associated with diversity in biological phenotype [4], [5]. Several studies suggest that this may be the case [5], [6]. A comparison of pulmonary TB with TB meningitis in Vietnam demonstrated that strains belonging to the modern Euro-American lineage were significantly less likely to cause extra-pulmonary disease [7]. In contrast, the modern Beijing/W lineage was found significantly associated with extra-thoracic TB in comparison with non-Beijing/W lineages, though no difference was found when looking at other read-outs of virulence such as the amount of cavitation [8]. A recent study in Madagascar found that infection with an ancient lineage induced a significantly higher immune response, measured as interferon-γ production by peripheral blood T cells [9]. Finally, a series of studies, including the report from Madagascar, have described a reduced immune response to members of the “Beijing family” (part of the modern East Asian Lineage 2), and its association with rapid progression to severe disease in humans and experimental animals [9], [10], [11], [12], [13], [14].
Two studies that analyzed individual isolates involved in TB outbreaks concluded that a low inflammatory response was linked to increased virulence [10], [15]. The rationale is that a reduction in innate immune recognition will result in a delay in engagement of the adaptive response, providing the pathogen with a significant advantage during the early stage of infection. The low inflammatory phenotype of M. tuberculosis HN878, a member of the Beijing family implicated in an outbreak in Texas, was reversed by disruption of the gene encoding an enzyme required for biosynthesis of a phenolic glycolipid molecule (PGL) [11]. However, it was shown later that the production of PGL was variable across strains from the Beijing/W lineage [16]. Moreover, the role of this particular glycolipid in the virulence of HN878 could not be reproduced by restoring its production in the genetic background of another modern strain belonging to Lineage 4, highlighting a rather confusing inter- and intra-lineage diversity in the molecular mechanisms of M. tuberculosis pathogenicity. In contrast, the low inflammatory phenotype of M. tuberculosis CAS, responsible for an outbreak in Leicester, was linked to a chromosomal deletion and could be reversed by restoration of the functional gene [15]. Several other studies have described differences in the inflammatory response induced by different isolates of M. tuberculosis [12], [13], [17], [18], [19].
In the present study, we used a collection of strains covering the global genetic diversity of MTBC to test the hypothesis that inflammatory phenotype would be linked to genotype.
To test for a link between genotype and inflammatory phenotype, we selected 26 isolates representative of the global diversity of human MTBC from a well-characterized clinical strain collection [1], [4], [20] (Figure 1) plus two laboratory adapted strains as references (M. tuberculosis H37Rv and M. bovis BCG Pasteur) and measured their ability to induce production of inflammatory cytokines by human GM-CSF monocyte-derived macrophages (T1-MDMs) [21]. Figure 2 shows cytokine levels from culture supernatants harvested 24 hours after infection with each of the strains for two human donors, and highlights three important observations.
First, there were clear differences in the level of pro-inflammatory cytokines produced by a single donor in response to different strains; ranging from a few hundred picograms to several nanograms. Second, although the absolute amount of cytokine varies between individual donors (results in Figure 2 are plotted against two separate y-axes), the relative hierarchy of low and high responses is the same in the two donors. This is further illustrated in Figure 3, summarizing results from eight donors, with strains ranked according to their ability to induce a cytokine response after median normalization of the dataset. In addition to robust consistency across different human donors, a similar hierarchy was observed when the same strains were used to stimulate bone-marrow-derived macrophages from Balb/c mice (Figure S1). A third observation is that the different pro-inflammatory cytokines – IL-6, IL-12p40/p70 and TNFα – all showed a similar pattern. Again, this is particularly clear when analyzed across the panel of eight donors, showing a highly significant correlation between production of IL-6 and IL-12p40/p70 for example (Figure 3C; Spearman rank correlation coefficient = 0.988, P<0.001).
We next compared the pro-inflammatory responses across the main MTBC lineages. Combining results from the eight individual donors after median normalization revealed higher heterogeneity in IL-6 production in response to Lineage 5, 6 and 1 (Figure 4A). However, there was overall a statistically significant effect of lineage on the cytokine levels observed (Kruskal-Wallis rank test, P<0.001). To look for higher order grouping of the lineages we performed a principal component analysis (PCA) using the genetic variation among the strains. PCA distinguished three major groups: strains belonging to Lineages 5+6 (M. africanum), strains belonging to Lineage 1, and strains belonging to the Lineages 2+3+4 (these three lineages have been referred to as evolutionarily “modern” based on previous work [1], [22]) (Figure 4B). These three major groupings were consistent with the most recent genome-based MTBC phylogeny published earlier this year [2] and shows that the “modern” lineages are more genetically homogenous than “ancient” lineages.
Comparing the level of cytokine induction across these three groups showed that the modern group of strains consistently induced a lower IL-6 response when compared to Lineage 5+6 or Lineage 1 (Figure 4C), though Lineage 1 had an intermediate phenotype reflecting again the higher heterogeneity of the ancient group when compared to the modern strains. Even so, combining the lineages according to the ancient/modern grouping revealed an overall difference in inflammatory phenotype, with strains from the modern lineages always eliciting significantly lower levels of IL-6 (Figure 4C) and other cytokines and chemokines such as IL-12p40/p70, TNFα, IL-15, MIP-1α, CCL5 and others (Figure 5). Therefore for the rest of the study we decided to keep the "ancient"/"modern" dichotomy to highlight the different behavior of the later group when compared with the other lineages. However, relevant figures splitting the data in three groups ("M. africanum", Lineage 1 and "Modern") are made available as supplementary material (Figure S2 A-E).
As a further test of the reproducibility of the differential inflammatory response, experiments were repeated using a second batch of the same strains of MTBC separately cultured and quantified. While a few individual isolates showed evidence of batch variation in inflammatory phenotype, the overall pattern of a lower response to the modern lineage was maintained as shown by a statistically significant correlation test (Spearman correlation test, P<0.05) (Figure S3).
There was a differential increase in the production of pro-inflammatory cytokines as the infection progressed, and the difference between ancient and modern lineages observed at the 24-hour time point had markedly diminished by 72 hours (Figure 6; Mann-Whitney U test, P<0.05). Thus, the reduced response to the modern strains is due to a delay in the kinetics, rather than a complete inhibition of the immune response.
The differential response pattern was not observed in infection experiments using unfractionated peripheral blood mononuclear cell (PBMC) preparations in place of differentiated monocytes. In these experiments, we observed a lower amount of TNFα but with no significant difference between strains from the ancient and modern lineages, and a statistically significant higher amount of IL-6 in comparison to T1-MDMs (Figure 7A; Mann-Whitney U test, P<0.05).
Replacement of GM-CSF by M-CSF during monocyte differentiation generated a “Type 2” macrophage population characterized by a lower level of production of pro-inflammatory cytokines, along with increased expression of IL-10 [21]. Figure 7B illustrates the differences between Type 1 (T1-MDMs) and Type 2 (T2-MDMs) polarization in response to LPS stimulation. Comparative flow cytometry analysis is shown as supplementary material (Figure S4). Infection of T2-MDMs with the different strains of MTBC resulted in lower production of IL-6 and IL-12p40/p70 as compared to T1-MDMs, with a significant difference between lineages only for IL-6 (Figure 7C, D). In contrast to T1-MDMs, a consistent IL-10 response was induced during infection of the T2-MDMs. The IL-10 response was also variable between isolates, but no significant difference was observed when comparing ancient and modern lineages (Figure 7E).
IL-10 is an anti-inflammatory cytokine that has been shown to act as an auto-regulatory inhibitor of pro-inflammatory cytokine production by human monocytes [23], [24]. Notably, IL-10 production has been associated with the anti-inflammatory phenotype of a recent outbreak strain belonging to the modern lineage [15]. IL-10 is produced and effective at very low concentrations. To test whether the differential inflammatory response observed in T1-MDMs might be influenced by variations in production and consumption of IL-10 that were not detected by ELISA, we repeated infection experiments in the presence of anti-IL-10 blocking antibodies. Consistent with the idea that IL-10 is produced and acting at very low levels, IL-10 blockage resulted in a systematic increase in pro-inflammatory cytokines TNFα and IL-6, but not in IL-12p40/p70. However, it did not affect the differential between ancient and modern lineages (Figure 8).
We also matured monocytes from three different donors in the presence of GM-CSF and IL-4 in order to generate monocyte-derived dendritic cells (MD-DCs) [21]. Compared with autologous T1-MDMs, MD-DCs significantly down-regulated CD14 and expressed higher levels of MHC class II molecules and CD86 (Figure S4). The response of MD-DCs to infection with the different strains resembled that of autologous T1-MDMs. Although the absolute amounts of immune mediators were generally lower, the trend towards weaker responses to the modern lineage was preserved for a largely overlapping panel of cytokines and chemokines but also IL-1β, IL-1RA, CXCL8 and GM-CSF (Figure 9).
Using a panel of strains representative of human MTBC genetic variability, we have found that genetically diverse strains of MTBC vary widely in their induction of an early inflammatory response during infection of human macrophages, and that these differences are linked to MTBC lineages. Overall there was a significantly lower response to evolutionarily modern lineages as compared to ancient lineages.
Previous reports have described differences amongst M. tuberculosis isolates in their inflammatory phenotype [10], [12], [13], [15], [18], [19] but the present study is the first to link this to MTBC phylogeny. Consistent with previous reports, we observed a higher inflammatory phenotype for the laboratory strain H37Rv in comparison to the Beijing strain, HN878 [11]. We found that these differences were robust and reproducible using different batches of bacteria to infect monocyte-derived macrophages and autologous dendritic cells across different human donors. We were concerned that an experimental bias could result from inaccuracies in quantifying bacterial preparations. Our quantitation is based on measurement of colony forming units and, while we made every effort to use comparable actively growing cultures, to minimize clumping artifacts and to repeat multiple measurements, errors could arise if there is a variation between strains in the rate of accumulation of dead cells during culture. While it is difficult to rigorously exclude this possibility, we consider it an unlikely explanation for our results since (a) differences in accumulation of dead cells would need to span several orders of magnitude, and (b) in some experimental systems – infection of PBMCs or T2-MDMs – the same preparations showed opposite or no differences in cytokine response. Convergence in responses of T1-MDMs to different strains at later time points is also consistent with equivalence between preparations.
Although we did observe a link between inflammatory phenotype and the various bacterial genotype clusters, each of the phylogenetic lineages included strains that induced high and low levels of inflammatory cytokines. For example, strain N0024 belonging to Lineage 3 consistently elicited a stronger inflammatory response as compared to the other strains from the same lineage. Future comprehensive genetic and biochemical analysis of these strains will be performed with the aim of deciphering the origin of this difference. The variation observed within lineages suggests that differences in inflammatory phenotype cannot be explained simply by the presence or absence of a single molecular determinant, and one model to account for the dispersal of high and low inflammatory strains across the phylogenetic tree is that a diverse range of mutations might influence innate immune recognition and arise independently in different lineages. This model is consistent with the known complexity and diversity of mycobacterial products that have been shown to stimulate or inhibit inflammatory responses [11], [25], [26], [27], [28]. Considering the multiple families of cell-surface and intracellular receptors involved in mycobacterial recognition [29], one might expect that the consequences of molecular changes that affect different ligands would depend on the exact receptor repertoire expressed by different donors [20]. However, the consistent hierarchy in inflammatory response that was observed across independent donors suggests that there is limited human variability in the initial immune response to genetically different mycobacteria. Our demonstration of a robust hierarchy in inflammatory phenotype within the different lineages of MTBC poses a challenge to our understanding at a molecular level of the microbiology and pathogenesis of tuberculosis as multiple mechanisms might converge towards similar phenotypic effects in distinct MTBC lineages. This observation may hold important lessons for development of new vaccines.
In spite of the heterogeneity within MTBC lineages, we observed a statistically significant distribution towards more pro-inflammatory strains amongst members of the ancient lineages, and lower inflammatory responses to strains from the modern lineages. The low inflammatory phenotype of modern strains is in agreement with previous studies of individual Beijing strains [10], [12] and other strains [15] belonging to the modern lineages. Our results are also consistent with the previous suggestion that a low inflammatory response may lead to a reduction in the adaptive response [9].
A possible model to rationalize this finding is that the respective high and low inflammatory responses could reflect different virulence strategies that emerged during the evolution of the ancient and modern lineages. The characteristic latency in TB has been suggested to represent an evolutionary adaptation to low host densities, with reactivation after several decades allowing the pathogen to access new susceptible birth cohorts [1], [30]. By contrast, the low inflammatory response induced by evolutionary modern outbreak strains has been associated with an enhanced ability to cause early progressive disease [10], [15]. Such a strategy may be an advantage in the context of high human population densities, where the number of susceptible hosts is large, and rapid lethality does not threaten to exhaust the pool of new uninfected hosts.
We previously presented an evolutionary scenario for human TB based on population genetic analyses of multilocus sequence data [1] referred to recently as “the most well defined phylogeny of the MTB complex” [31]. According to this scenario, M. tuberculosis originated in Africa and accompanied early modern humans on their Out-of-Africa migrations. In those times, human populations were small, and M. tuberculosis might have benefited from the latency strategy [30]. During the last few hundred years, the three modern lineages of M. tuberculosis experienced strong population expansions as a consequence of the recent human population increases in Europe, India and China [1]. The overall lower inflammatory responses observed in the modern lineages of M. tuberculosis might be a consequence of their access to rapidly increasing numbers of susceptible hosts resulting in selection for faster progression to active disease. In support of this hypothesis, a study in the Gambia showed that members of the modern strain lineages were three times more likely than members of the ancient lineages to cause active TB in recently exposed contacts [32].
To our knowledge, this is the first time that the immune response to a particular infectious agent has been measured in a systematic manner by selecting representative strains belonging to the major human MTBC lineages and grasping the global M. tuberculosis genetic diversity, including notably M. africanum. As we show here, the combination of genotypic, phenotypic and epidemiological studies offers the potential for novel insights into the biology of this pathogen.
Mycobacterial cultures of clinical isolates were obtained from a single colony forming unit. One volume of a stationary phase culture of mycobacteria in Middlebrook 7H9 medium with ADC supplement (BD Biosciences), 0.05% Tween-80 (Sigma-Aldrich) and in some cases sodium pyruvate 40 mM (e.g. M. africanum strains [33]) was diluted with 100 volumes of the same medium in the absence of detergent and incubated for 10 days at 37°C. Gentle culture dispersion was performed manually every 48 h. Mycobacteria were pelleted, supernatants discarded and pellets dispersed by manual shaking for 1 min with equal volumes of 2–3 mm glass beads. Mycobacteria were resuspended in PBS and centrifuged at 260 xg for 10 min to remove clumps. Cleared supernatants mostly composed of single particles [34] were supplemented with 5% glycerol and titrated on 7H11 agar plates complemented with OADC (BD Biosciences) and sodium pyruvate 40 mM before and after freezing and storage.
Peripheral blood mononuclear cells (PBMCs) from healthy anonymous donors were isolated from buffy coats processed by the National Blood Services, Colindale, UK. PBMCs were prepared on a Ficoll-Paque density gradient (Amersham Biosciences AB, Uppsala, Sweden) by centrifugation (800 xg, 30 min at room temperature). Recovered PBMCs were washed twice with RPMI (Gibco, Invitrogen) and resuspended in RPMI/FCS(4%)/methyl-cellulose(2%)/DMSO(9%) for gradual overnight freezing in a NalgeneTM Cryo 1°C container before storage in liquid nitrogen.
Monocytes were selected from fresh or frozen PBMCs by magnetic cell sorting using CD14 microbeads (Miltenyi Biotec, Auburn, CA, USA) according to manufacturer's recommendations. Cell purity checked by flow cytometry was always >95%. Macrophages were differentiated from monocytes after 6 days of culture in the presence of recombinant human GM-CSF or M-CSF (Peprotech Ltd) for T1-MDMs or T2-MDMs respectively and monocyte derived dendritic cells (MD-DCs) in the presence of GM-CSF and IL-4 as previously described [21]. Cells were recovered after 15 min Trypsin/EDTA (2 mM) treatment, resuspended in RPMI and 5% FCS, and evenly distributed at 8×104 to 1×105/well (according to experiment) in tissue culture treated 96 well plates or 12.5×103/well in 384 well plates before mycobacterial infection at a multiplicity of infection of 1∶1 unless specified otherwise. LPS stimulation was performed at a final concentration of 10 ng/ml. IL-10 blocking experiments were performed as described elsewhere [35], antibodies were added prior to infection at a final concentration of 0.1 µg/ml.
Mice were bred in the animal facilities of the National Institute for Medical Research and provided after being sacrificed in line with code of practice for the humane killing of animals under schedule 1 to the animals (scientific procedures) Act 1986. Authors were not involved in the handling and/or sacrificing of live mice. Femurs from dead Balb/c mice were flushed with 1 ml complete medium (RPMI1640, 1 mM sodium pyruvate, 2 mM glutamine, 10 mM HEPES, 0.05 mM β-mercapthoethanol and 10% FCS) using a 25 G needle. Cells were pelleted, red blood cells lysed for 5 mins with 10 ml 0.83% ammonium chloride, filtered through a 70 µM strainer and washed twice with complete medium before incubation in a CO2 incubator at 37°C for 4 days in 90 mm Petri dishes (4×106 in 8 ml complete medium containing 20% L-cell medium). On day 4, 10 ml conditioned medium was added and cells were harvested on day 7 by removing supernatant and adding 5 ml PBS containing 2 mM EDTA to detach the macrophages. Recovered cells were pelleted and resuspended in complete medium and plated out as described for human monocyte derived macrophages.
Cell supernatants were recovered at indicated time points, sterilised twice using 96 well filter plates (0.2 µm, Corning) and stored at −20°C until analysis. IL-6, IL-12p40/p70, TNF-α and IL-10 were measured using either ELISA kits (Peprotech) or Luminex 30-plex kit (Invitrogen) following manufacturer's recommendations.
Data analysis, correlation study, paired t-tests, Mann-Whitney U tests and Kruskal-Wallis rank test were performed using GraphPad Prism software and STATA s.e.m. version 10. Without assuming a pre-defined distribution of the response tested, non-parametric statistical analysis has been used all across the study. Unless otherwise stated, we used the individual measures for each combination of donor and strain for all the statistical analysis (n = 200). Principal component analysis was conducted using STATA s.e.m. version 10 with the polymorphic positions found in Hershberg et al. 2008 for the strains used in the present study.
|
10.1371/journal.ppat.1006075 | Increased Abundance of M Cells in the Gut Epithelium Dramatically Enhances Oral Prion Disease Susceptibility | Many natural prion diseases of humans and animals are considered to be acquired through oral consumption of contaminated food or pasture. Determining the route by which prions establish host infection will identify the important factors that influence oral prion disease susceptibility and to which intervention strategies can be developed. After exposure, the early accumulation and replication of prions within small intestinal Peyer’s patches is essential for the efficient spread of disease to the brain. To replicate within Peyer’s patches, the prions must first cross the gut epithelium. M cells are specialised epithelial cells within the epithelia covering Peyer’s patches that transcytose particulate antigens and microorganisms. M cell-development is dependent upon RANKL-RANK-signalling, and mice in which RANK is deleted only in the gut epithelium completely lack M cells. In the specific absence of M cells in these mice, the accumulation of prions within Peyer’s patches and the spread of disease to the brain was blocked, demonstrating a critical role for M cells in the initial transfer of prions across the gut epithelium in order to establish host infection. Since pathogens, inflammatory stimuli and aging can modify M cell-density in the gut, these factors may also influence oral prion disease susceptibility. Mice were therefore treated with RANKL to enhance M cell density in the gut. We show that prion uptake from the gut lumen was enhanced in RANKL-treated mice, resulting in shortened survival times and increased disease susceptibility, equivalent to a 10-fold higher infectious titre of prions. Together these data demonstrate that M cells are the critical gatekeepers of oral prion infection, whose density in the gut epithelium directly limits or enhances disease susceptibility. Our data suggest that factors which alter M cell-density in the gut epithelium may be important risk factors which influence host susceptibility to orally acquired prion diseases.
| Prion diseases are infectious neurodegenerative disorders that affect humans and animals. Many natural prion diseases are orally acquired through consumption of contaminated food or pasture. An understanding of how prions infect the intestine will help identify factors that influence disease susceptibility and allow the development of new treatments. After oral infection prions first accumulate within the lymphoid tissues that line the intestine (known as Peyer’s patches) before they spread to the brain where they cause neurodegeneration. To do this, the prions must first cross the intestinal epithelium, a single layer of cells that separates the body from the gut contents. M cells are found within the epithelium that covers the Peyer’s patches and are specialised to transport large particles and whole bacteria across the gut epithelium. We show that M cells act as the gatekeepers of oral prion infection. In the absence of M cells, oral prion infection is blocked, whereas an increase in M cells increases the risk of prion infection and shortens the disease duration. Therefore, our data demonstrate that factors such as pathogen infection, inflammation and aging, which alter the abundance of M cells in the intestine, may be important risk factors which influence susceptibility to orally-acquired prion infections.
| Prion diseases (transmissible spongiform encephalopathies) are a unique group of subacute neurodegenerative diseases that affect humans and animals. During prion disease, aggregations of PrPSc, an abnormally folded isoform of cellular PrP (PrPC), accumulate in affected tissues. Prion infectivity co-purifies with PrPSc and constitutes the major, if not sole, component of the infectious agent [1–3]. Many natural prion diseases, including natural sheep scrapie, bovine spongiform encephalopathy (BSE), chronic wasting disease in cervids, and variant Creutzfeldt-Jakob disease in humans (vCJD), are acquired peripherally, such as by oral consumption of prion-contaminated food or pasture. The precise mechanism by which orally-acquired prions are propagated from the gut lumen across the epithelium to establish host infection is uncertain. In the U.K. relatively few vCJD cases have fortunately occurred despite widespread dietary exposure to BSE [4], suggesting that the acquisition of prions from the gut lumen may differ between individuals. Further studies are clearly necessary to precisely characterise the cellular route that prions exploit to establish infection after oral exposure, and how alterations to this cellular route, both intrinsic and extrinsic, can affect disease susceptibility. Treatments which prevent the accumulation and replication of prions in host lymphoid tissues can significantly reduce disease susceptibility [5–9]. Therefore, identification of the cellular route by which prions are first transported across the gut epithelium to achieve host infection will identify an important factor which influences oral prion disease susceptibility and to which intervention strategies can be developed.
Following oral exposure the early accumulation and replication of prions upon follicular dendritic cells (FDC) within the gut associated lymphoid tissues (GALT), such as Peyer’s patches of the small intestine, is essential for efficient neuroinvasion [7, 10–13]. FDC are a unique subset of stromal cells resident within the primary B cell follicles and germinal centres of lymphoid tissues [14]. After amplification upon the surface of FDC [15], the prions then infect neighbouring enteric nerves and spread along these to the CNS (a process termed neuroinvasion) where they ultimately cause neurodegeneration and death of the host [16–19].
The follicle-associated epithelia (FAE) which covers the lumenal surfaces of the Peyer’s patches contains a unique population of epithelial cells, termed M cells. These highly phagocytic epithelial cells are specialized for the trans-epithelial transfer of particulate antigens and microorganisms from the gut lumen (termed transcytosis) [20], an important initial step in the induction of efficient mucosal immune responses against certain pathogenic bacteria [21, 22] and the commensal bacterial flora [23]. A variety of bacterial and viral pathogens including Brucella abortus [24], Salmonella Typhimurium [25], Yersinia enterocolitica [26], norovirus [27, 28] and reovirus [28] appear to exploit the transcytotic activity of M cells to cross the gut epithelium and infect the host. The food-borne botulinum neurotoxin [29] has also been suggested to exert its toxicity after transcytosis by M cells [29]. Independent studies suggest orally administered prions may similarly be transported by M cells into host tissues [9, 30–32] and that this transport may be important to establish host infection [9]. Other studies have also suggested that prions can be transported across the gut epithelium via enterocytes, independently of M cells [16, 33, 34], however to what extent enterocyte-transported prions contribute to the establishment of host infection has not been assessed.
The differentiation of M cells from uncommitted precursors in the intestinal crypts is critically dependent on stimulation from the cytokine known as RANKL (receptor activator of nuclear factor-κB ligand). This cytokine is expressed by subepithelial stromal cells beneath the FAE in Peyer’s patches, and signals via its receptor RANK (receptor activator of nuclear factor-κB) which is expressed by epithelial cells throughout the intestine [35]. Accordingly, M cell-differentiation is blocked in RANKL-deficient mice or following in vivo RANKL-neutralization with anti-RANKL antibody [35]. RANKL stimulation induces a program of gene expression in intestinal epithelial cells which includes the transcription factor SPIB. Expression of SPIB by intestinal epithelial cells is essential for their differentiation and functional maturation into M cells [22, 36, 37]. We have previously reported that the early accumulation of prions upon FDC in Peyer’s patches and subsequent neuroinvasion were blocked in mice in which M cells were transiently depleted by RANKL-neutralization using anti-RANKL antibody [9]. However, since RANKL-RANK signalling has multiple roles in the immune system, a more refined model is required to specifically determine the role of M cells in oral prion disease pathogenesis. In the current study a unique conditional knockout mouse model was used in which RANK expression was specifically deleted only in the intestinal epithelium (RANKΔIEC mice) [23, 38]. In these mice the complete loss of M cells prevents M cell-mediated antigen uptake from the gut lumen, without altering other RANKL-RANK signalling events required for normal immune development and function [23, 38]. Using these mice our data clearly show that M cells are critically required for the initial trans-epithelial transfer of prions across the gut epithelium into Peyer’s patches in order to establish host infection.
Certain pathogenic bacteria [25, 39] or exposure to inflammatory stimuli such as cholera toxin [40] can significantly increase the density of M cells in the intestine. Inflammation or pathogen infection can also influence prion disease pathogenesis by enhancing the uptake, or expanding the distribution, of prions within the host [11, 41–43]. This raised the hypothesis that exposure to inflammatory stimuli that enhance M cell-density might increase oral prion disease susceptibility by enhancing the uptake of prions from the gut lumen. We show that increased M cell-density at the time of oral exposure dramatically enhanced the uptake of prions from the gut lumen, decreased survival times and increased disease susceptibility by approximately 10-fold. Our data provide a significant advance in our understanding of how prions exploit M cells to initially infect Peyer’s patches and how factors that increase the density of M cells in the gut epithelium, such as concurrent pathogen infection, may have the potential to increase susceptibility to orally-acquired prion infection.
Our previous study showed that oral prion infection was blocked after transient M cell-depletion by treatment with anti-RANKL antibody, implying a functional role for M cells in the trafficking of prions from the lumen into GALT in vivo [9]. Although the major phenotype observed in the intestine was a transient loss of mature M cells, RANKL-RANK signalling is also important in immune system and lymphoid tissue development. Therefore, systemic RANKL neutralization by treatment with anti-RANKL antibody could have affected other important cellular processes involved in prion pathogenesis. To exclude these, we used a more refined model of M cell-deficiency, RANKΔIEC mice [23, 38], to further elucidate the role of M cells in the transport of prions from the intestinal lumen into GALT. These mice are specifically deficient in Tnfrsf11a (which encodes RANK) only in Vil1-expressing intestinal epithelial cells. As previously published [23], whole-mount immunostaining for the mature M cell marker glycoprotein 2 (GP2; [21, 22]) revealed an absence of GP2+ M cells in the FAE of the Peyer’s patches of RANKΔIEC mice compared to control (RANKF/F) mice (Fig 1A & 1B). Coincident with the loss of RANK expression in the gut epithelium was a significant reduction in area of the FAE (Fig 1C).
Assessment of the uptake of fluorescent latex microbeads from the gut lumen into Peyer’s patches is a reliable in vivo method to compare the functional ability of M cells to acquire and transcytose particulate antigens. Here, RANKΔIEC mice and RANKF/F control mice (n = 3/group) were orally gavaged with 2x1011 200 nm fluorescent microbeads, and 24 h later the number of microbeads in their Peyer’s patches quantified by fluorescence microscopy. This duration was selected to ensure sufficient time for the beads to transit through the intestine and be transcytosed by M cells in the FAE overlying the Peyer’s patches [13]. Coincident with the absence of mature GP2+ M cells, RANKΔIEC mice had substantially less fluorescent microbeads within the subepithelial dome (SED) regions of their Peyer’s patches when compared to controls (Fig 1D), indicating a dramatic reduction in the ability to sample particulate antigen from the gut lumen.
RANK-dependent GP2+ M cells have been described in the epithelium of the nasal associated lymphoid tissue (NALT) [44, 45]. The abundance of GP2+ M cells in the NALT was unaffected in RANKΔIEC mice (Fig 1E), highlighting the intestinal specificity of the model.
In addition to being transported through M cells, prions have also been observed trafficking into Peyer’s patches through the large LAMP1+ endosomes of FAE enterocytes [16]. Immunohistochemical (IHC) analysis of LAMP1 expression showed that these endosomes were still present in the FAE of RANKΔIEC mice (Fig 1F). If the presence of these endosomes in the FAE was dependent on RANKL-RANK signalling, we reasoned that the abundance of LAMP1+ immunostaining would be decreased in the FAE of RANKΔIEC mice. However, morphometric analysis indicated equivalent areas of LAMP1+ immunostaining in the FAE of RANKΔIEC and RANKF/F mice (Fig 1G). These data suggest that the presence of LAMP1+ endosomes in the FAE was not RANKL-RANK signalling dependent.
Antigens that are transcytosed by M cells are released into their basolateral pockets where they are sampled by lymphocytes and mononuclear phagocytes (MNP; a heterogeneous population of macrophages and classical dendritic cells; DC) [46–48]. The acquisition of prions by MNP such as CD11c+ classical DC may mediate their initial transport to FDC [8, 16, 49], and the subsequent transfer of prions from FDC to the peripheral nervous system [50–52]. IHC and morphometric analysis revealed a significant reduction in the % area of CD11c-specific immunostaining in the SED of the Peyer’s patches from RANKΔIEC mice (Fig 2A & 2B), whereas the % area of CD68-specific immunostaining (indicative of tissue macrophages) was equivalent in RANKΔIEC and RANKF/F mice (Fig 2A & 2C). Analysis of the intestinal lamina propria (LP) showed a similar trend (Fig 2D–2F).
Following replication upon FDC, the prions subsequently infect enteric nerves (both sympathetic and parasympathetic) to reach the CNS where they ultimately cause neurodegenerative disease [16, 18]. Our IHC analysis of the expression of the neuronal synaptic vesicle marker synaptophysin 1 suggested that the magnitude of the enteric innervation in the LP was similar in the intestines of RANKΔIEC and RANKF/F mice (Fig 2G & 2H).
Together these data demonstrate that RANKΔIEC mice represent a refined model in which to study the specific role of M cells in oral prion disease pathogenesis.
The early replication of many prion strains upon FDC within the B cell-follicles of the draining lymphoid tissues is essential for their efficient transmission to the CNS after peripheral exposure [5–7, 15]. FDC in mice characteristically express high levels of CD21/35 (complement receptors 2 & 1, respectively). Our IHC analysis showed that the area of CD21/35-specific immunostaining in Peyer’s patches of 10 wk old RANKΔIEC and RANKF/F mice was similar (Fig 3A & 3B), suggesting that the size of the FDC networks (CD21/35+ cells) in the Peyer’s patches of each mouse strain was equivalent. The replication of prions upon FDC is critically dependent on their expression of PrPC [15, 53, 54]. Morphometric analysis also indicated that the magnitude of the PrPC-expression co-localized upon CD21/35+ FDC in the Peyer’s patches (Fig 3A & 3C) and mesenteric lymph nodes (MLN) (Fig 3D & 3E) of RANKΔIEC mice and RANKF/F mice was similar.
We next determined whether the FDC in the lymphoid tissues of RANKΔIEC mice were capable of accumulating prions to a similar extent as those of control mice. After injection by the intra-peritoneal (i.p.) route high levels of prion accumulation and replication are first detected in the spleen within 35 d post infection (dpi) [53]. The prions are then subsequently disseminated around the host via the blood and lymph to most other secondary lymphoid tissues [55]. Furthermore, by 140 dpi the prions are also detectable within Peyer’s patches. Since the prions do not need to cross the gut epithelium to eventually infect the Peyer’s patches after injection by the i.p. route, RANKΔIEC and RANKF/F were injected with a 1% dose of ME7 scrapie prions via this route and tissues collected at 140 dpi, to determine whether the FDC in the lymphoid tissues of RANKΔIEC mice were capable of accumulating prions. Prion disease-specific accumulations of PrP (referred to as PrPd) were detected by immunostaining for the abnormal aggregates of PrP characteristically present only in affected tissues [6, 9, 11, 13, 53, 56], complimented with paraffin-embedded tissue (PET) blot analysis of adjacent membrane-bound sections to confirm that these aggregates contained relatively proteinase-K (PK)-resistant prion disease-specific PrPSc [57]. Abundant accumulations of PrPSc were evident in association with FDC (CD21/35+ cells) in the Peyer’s patches, MLN and spleens of RANKΔIEC and RANKF/F mice (Fig 4A–4C). These data clearly show that the FDC in the Peyer’s patches, MLN and spleen of RANKΔIEC mice were functionally capable of acquiring and accumulating prions, and that the dissemination of prions between lymphoid tissues was not impaired. Importantly, these data also suggest that the cause of any difference in prion pathogenesis between RANKΔIEC and RANKF/F mice observed after oral exposure would be restricted to effects on M cells in the gut epithelium.
Within weeks after oral exposure, high levels of ME7 scrapie prions first accumulate upon FDC in the Peyer’s patches and subsequently spread to the MLN and spleen [7–9, 11, 13]. The initial replication of prions upon FDC in the Peyer’s patches is essential for the efficient transmission of disease to the CNS [7, 11, 13]. In order to determine the effect of specific M cell-deficiency on oral prion disease pathogenesis, RANKΔIEC mice and RANKF/F (control) mice were orally exposed to a moderate dose of ME7 scrapie prions (50 μl of a 1% brain homogenate from a mouse clinically-affected with ME7 scrapie prions; [7, 9, 11, 13, 58]). At intervals after exposure the accumulation of PrPd and PrPSc in tissues from 4 mice/group were compared by IHC and PET blot analysis, respectively, as above. As anticipated, at 105 dpi, abundant accumulations of PrPd (middle row, brown) and PrPSc (lower row, black) were detected in association with FDC (CD21/35+ cells, upper row, brown) in the Peyer’s patches, MLN and spleen of RANKF/F control mice (Fig 5A). However, no PrPSc accumulations were detected in the same tissues from RANKΔIEC mice (Fig 5A). Mice on a C57BL/6 background typically succumb to a moderate dose of ME7 scrapie prions by ~340 d after oral exposure [9, 13]. However, RANKΔIEC mice (n = 8) remained free of the clinical signs of prion disease up to at least 440 dpi, at which point no PrPd or PrPSc was detected in their Peyer’s patches, MLN, spleen (Fig 5B), spinal cords or brains (Fig 5C) by IHC and PET blot analysis. Together these data clearly show that M cells are essential for the initial uptake of prions from the gut lumen into Peyer’s patches in order to establish host infection, since oral prion disease pathogenesis was blocked in the specific absence of M cells in RANKΔIEC mice.
Certain pathogen infections or inflammatory conditions can enhance M cell-differentiation within the intestine [25, 39, 40]. We therefore reasoned that alterations to M cell-density in the gut epithelium may significantly alter oral prion disease pathogenesis and susceptibility. The density of functionally mature M cells in the intestine can be promoted in mice through exogenous administration of RANKL [22, 35]. Recombinant RANKL was prepared and its ability to stimulate M cell-differentiation was confirmed in in vitro intestinal enteroids derived from RANKΔIEC and RANKF/F mice [23, 36]. As anticipated, RANKL-treatment of enteroids from RANKF/F (control) mice induced robust expression of several M cell-associated genes (Marcksl1, Anxa5, Spib, Ccl9, and Gp2; [22]) without significantly altering expression of genes associated with other intestinal lineages, including Paneth cells (Lyz1, Lyz2) and intestinal stem cells (Lgr5) (S1 Fig). No induction of expression of M cell-specific genes was observed in RANKL-treated enteroids derived from RANKΔIEC mice.
Next, C57BL/6 mice (n = 4/group) were treated daily with RANKL to induce M cell-differentiation and tissues harvested on d 3, coincident with the peak period of induction of M cell gene expression in the gut epithelium [22, 35]. A parallel group of mice were treated with PBS as a control. IHC and morphometric analysis revealed that RANKL-treatment induced a significant increase in the number of GP2-expressing (mature) and SPIB-expressing (differentiating and mature) M cells within the FAE of Peyer’s patches (Fig 6A–6C) and also in the villous epithelium (Fig 6D–6F). This increase in M cells was associated with increased functional ability to acquire particulate antigen from the gut lumen, demonstrated by a significant increase in the number of 200 nm microbeads transcytosed into the SED of Peyer’s patches and villous cores 24 h after their administration by oral gavage (Fig 6G–6I). Although a small increase in the area of LAMP1+ immunostaining was observed in the FAE after RANKL treatment, the abundance of LAMP1+ immunostaining was unchanged in the villous epithelium (Fig 6J–6L).
We also determined whether RANKL-treatment affected other important parameters considered to be required for prion infection. IHC and morphometric analysis suggested there was no significant difference in the area of CD21/35+ (indicative of FDC size) or PrPC+ immunostaining in the Peyer’s patches (S2A–S2C Fig) or MLN (S2D & S2E Fig) of RANKL-treated mice when compared to PBS-treated controls. This implied that RANKL-treatment had no significant effect on FDC status in the Peyer’s patches or MLN.
IHC and morphometric analysis also indicated that the % area of CD11c+ immunostaining in the SED of the Peyer’s patches (Fig 7A & 7B) and the LP (Fig 7D & 7E) did not differ between tissues from PBS- and RANKL-treated mice. Although the % area of CD68+ immunostaining was equivalent in the SED of the Peyer’s patches (Fig 7A & 7C), a significant increase was observed in the LP of RANKL-treated mice (Fig 7D & 7F). No difference in the % area of synaptophsyin 1+ immunostaining was observed in the LP (Fig 7G & 7H), suggesting that RANKL-treatment did not significantly affect the magnitude of the enteric innervation in the intestine.
Together, these data demonstrate that RANKL-treatment promotes M cell-differentiation in the FAE of Peyer’s patches and villous epithelium without significant effects on other key cells (FDC, CD11c+ cells and enteric nerves) considered to play an important role in oral prion disease pathogenesis.
To determine whether increased M cell-density in the intestine altered oral prion disease susceptibility, groups of C57BL/6 mice were treated daily with RANKL (or PBS as a control) for 4 d as above, and between the 3rd and 4th treatments (coincident with the peak period of induction of M-cell gene expression in the gut epithelium [22, 35]) the mice were orally exposed to either a moderate (1%) or limiting (0.1%) dose of ME7 scrapie prions. Exposure of C57BL/6 mice to a 1% dose of prions typically yields a clinical disease incidence of 100% in the recipients, whereas a 0.1% dose has a much lower incidence allowing the effects of RANKL-treatment on both survival time and prion disease susceptibility to be assessed. As anticipated, following oral exposure to a moderate (1%) dose of ME7 scrapie prions, all PBS and RANKL-treated mice developed clinical disease. However, the RANKL-treated mice succumbed to clinical disease approximately 17 d earlier with a shorter mean survival time when compared to PBS-treated control mice (PBS-treated mice, mean 346±25 d, median 343 d, n = 7/7; RANKL-treated mice, mean 329±18 d, median 322 d, n = 8/8; Fig 8A). When mice were orally exposed to a limiting (0.1%) dose of prions only three of eight PBS-treated mice succumbed to clinical disease with individual survival times of 371, 378 and 420 d (Fig 8A). The five remaining PBS-treated mice did not develop clinical prion disease up to 525 dpi. In contrast, RANKL-treatment significantly enhanced prion disease pathogenesis as seven of eight RANKL-treated mice exposed to a limiting dose of prions succumbed to clinical disease with significantly shorter survival times (Fig 8A; RANKL-treated mice, mean 352±22 d, median 350 d, n = 7/8; P<0.0078, Log-rank [Mantel-Cox] test). Only one of the eight RANKL-treated mice exposed to a limiting dose of prions was free of the clinical signs of prion disease up to at least 525 dpi.
The brains of all mice that developed clinical signs of prion disease in each treatment group had the characteristic spongiform pathology (vacuolation), astrogliosis, microgliosis and PrPSc accumulation typically associated with terminal infection with ME7 scrapie prions (Fig 8B). The distribution and severity of the spongiform pathology was also similar in the brains of all the clinically-affected mice (Fig 8C & 8D), indicating that RANKL treatment did not alter the course of CNS prions disease after neuroinvasion had occurred. In contrast, no histopathological signs of prion disease were detected in the brains of any of the clinically-negative mice.
As expected, at the terminal stage of disease high levels of PrPSc were maintained upon FDC in the Peyer’s patches, MLN and spleen of all clinically-affected mice. However, no evidence of PrPSc accumulation within these lymphoid tissues was observed in any of the orally-exposed clinically-negative mice (S3 Fig). These data show that all the clinically-negative mice were free of prions in their lymphoid tissues and brains, and therefore highly unlikely to succumb clinical prion disease after substantially extended survival times, had the observation period been extended beyond 525 dpi.
Our data suggested that RANKL-treatment significantly increased susceptibility to orally-administered prions. Indeed, no significant difference in disease incidence or mean survival time was observed in the RANKL-treated mice exposed to a 0.1% dose of prions when compared to PBS-treated control mice given a 10X higher (1%) dose (PBS/1% vs. RANKL/0.1%, P = 0.205; Log-rank [Mantel-Cox] test; Fig 8A). Together, these data demonstrate that increased M cell-deficiency in the gut epithelium following RANKL-treatment significantly enhances oral prion disease susceptibility by approximately 10-fold. Although certain concurrent pathogen infections or inflammatory stimuli may have multiple effects on the gut epithelium, our data suggest that factors such as these that modify M cell-density in the intestine [25, 39, 40] may represent important risk factors which can significantly influence susceptibility to orally-acquired prion infections.
Prion replication within Peyer’s patches is essential for efficient neuroinvasion after oral exposure [10–13]. We therefore determined whether the decreased survival times and increased prion disease susceptibility in orally-exposed RANKL-treated mice were associated with the earlier accumulation of prions in their lymphoid tissues. Mice were treated with RANKL (or PBS as a control) and orally exposed to a 1% dose of ME7 scrapie prions as above, and culled at intervals afterwards (n = 4/group). Abundant accumulations of PrPSc were clearly evident in association with FDC in the Peyer’s patches, MLN and spleen of RANKL-treated mice by 70 dpi, and were undetectable in the majority of the tissues from the PBS-treated animals at this time (Fig 9A & 9B). To compare prion infectivity levels between the treatment groups, spleen homogenates were prepared and injected intracerebrally (i.c.) into groups of tga20 indicator mice (n = 4/spleen homogenate). As the expression level of PrPC controls the prion disease incubation period, tga20 mice which overexpress PrPC are extremely useful as indicator mice in prion infectivity bioassays as they succumb to disease with much shorter survival times than conventional mice [59]. Significantly high levels of prion infectivity were detected in three of four of the spleens collected from the RANKL-treated mice at 70 dpi, whereas only one of four spleens from the PBS treated spleen contained detectable levels of prion infectivity (P<0.0002, Log-rank [Mantel-Cox] test; Fig 9C). By 105 dpi abundant accumulations of PrPSc were detected at equivalent frequencies in the lymphoid tissues of the PBS- and RANKL-treated animals (Fig 9D).
These data show that an increased density of M cells in the intestinal epithelium at the time of oral exposure enhanced the uptake of prions from the gut lumen, as the RANKL-treated mice accumulated prions within their lymphoid tissues significantly earlier than control mice.
Although a rare occurrence in the steady-state, certain pathogenic microorganisms can stimulate the direct sampling of the gut lumenal contents by classical DC [60–63]. Whether this direct sampling activity by classical DC also contributes to the efficient uptake of orally-administered prions in the steady-state is uncertain [8, 16, 49]. Since RANKL was administered systemically in the current study, it is plausible that this treatment may have stimulated the direct sampling of the lumenal contents by cells other than M cells such as classical DC. An additional experiment was performed to test this hypothesis. As shown above, RANKΔIEC mice are unable to accumulate prions in their Peyer’s patches due to the specific lack of M cells (Fig 5). Since RANK-deficiency in RANKΔIEC mice is restricted only to intestinal epithelial cells [23], we reasoned that if the effects of RANKL-treatment on disease pathogenesis were independent of their effects on M cells, then RANKL-treatment would also facilitate the uptake of prions into the Peyer’s patches of RANKΔIEC mice. To address this issue, RANKΔIEC mice were treated with RANKL and orally exposed to a 1% dose of ME7 scrapie prions as in the previous experiment. At 105 dpi Peyer’s patches and MLN were collected and analysed for the presence of PrPSc as before. As anticipated, abundant accumulations of PrPSc were detected in association with FDC in the Peyer’s patches and MLN of orally-exposed C57BL/6 wild-type (WT) control mice by 105 dpi. However, no PrPSc was detected in tissues from RANKL-treated RANKΔIEC mice (S4 Fig). These data clearly show that RANKL-treatment was unable to restore prion accumulation in the Peyer’s patches and MLN of RANKΔIEC mice, indicating that the major effects of RANKL-treatment on oral prion disease pathogenesis were due to effects on M cell-deficiency in the intestinal epithelium.
RANKL-treatment stimulates M cell-differentiation within the FAE of the Peyer’s patches and also in the villous epithelium (Fig 6; [22, 35, 64]). We therefore considered it plausible that the enhanced prion pathogenesis we observed in RANKL-treated mice was due to the increased uptake of prions by the M cells induced in the villous epithelium. If RANKL-treatment had stimulated the uptake of prions predominantly via villous M cells, we reasoned that this would have facilitated the earlier transport of prions directly to the MLN [65]. An additional experiment was designed to test this hypothesis.
Lymphotoxin-β-deficient (LTβ-/-) mice lack Peyer’s patches and most peripheral lymph nodes, but retain MLN and the spleen [66]. These mice also lack FDC in their remaining lymphoid tissues, as constitutive LT-stimulation is essential for their maintenance [67], and are refractory to oral prion infection [10, 11]. Peyer’s patches-deficient LTβ-/- mice were γ-irradiated and reconstituted with LTβ-expressing (WT) bone marrow (termed WT→LTβ-/- mice, hereinafter) and tissues collected at 2.5 weekly intervals (n = 4 mice/group). Although the formation of FDC networks within the MLN and spleens of WT→LTβ-/- mice is restored by 5 wk after reconstitution (Fig 10A), WT→LTβ-/- mice remain refractory to oral prion disease [11] as Peyer’s patches, not the MLN, are the essential early sites of prion accumulation and neuroinvasion after oral exposure [11, 13].
The reconstitution of LTβ-/- mice with WT bone marrow also induces the differentiation and maturation of isolated lymphoid follicles (ILF) throughout the small intestine [11, 68, 69]. Mature ILF characteristically contain a single organized B cell-follicle, a network of FDC, and an M cell-containing FAE at the lumenal surface [11, 13, 68]. Since we have shown that mature small intestinal ILF are important sites of prion accumulation and neuroinvasion [11, 13], it was necessary to ensure there were no ILF with M cell-containing FAE in the intestines of WT→LTβ-/- mice at the time of RANKL-treatment and prion exposure. Whole-mount immunostaining of three individual 2 cm sections of small intestine from each WT→ LTβ-/- mouse showed that ILF with developed FAE containing GP2+ M cells were not present until 12.5 post-reconstitution (Fig 10B & 10C). These data revealed a window of opportunity between 5–10 wk post-reconstitution during which the small intestines of WT→LTβ-/- mice lacked FAE and M cell-containing GALT, but possessed FDC within their MLN. This FAE-deficient model was therefore used to determine whether RANKL-treatment facilitated the direct delivery of prions from the gut lumen to the MLN.
At 7.5 wk post-reconstitution WT→LTβ-/- mice (n = 3-4/group) were treated with RANKL (or PBS as a control) for 4 d and orally-exposed to prions as before, and prion infectivity levels determined in their MLN 28 d later. Tissues were assayed for prion infectivity at this time after oral exposure to determine whether RANKL-treatment of WT→LTβ-/- mice facilitated the earlier replication of prions within the MLN. Consistent with our previous data showing that Peyer’s patches in the small intestine, not the MLN, are the important early sites of prion accumulation after oral exposure [11, 13], prion infectivity was undetectable in the MLN of the PBS control-treated WT→LTβ-/- mice. Similarly, prion infectivity was also undetectable in the MLN of the RANKL-treated WT→LTβ-/- mice. In each instance all the recipient tga20 indicator mice (n = 4/MLN homogenate tested) were free of clinical disease up to 200 dpi (Fig 10D) and had no histopathological signs of prion disease in their brains (spongiform pathology and PrPd deposition; Fig 10E). These data clearly show that RANKL-treatment did not stimulate the early transport of prions directly to the MLN. This suggests that the enhanced prion disease pathogenesis observed in RANKL-treated mice was due to the increased uptake of prions from the gut lumen by M cells in the FAE of the Peyer’s patches, rather than by villous M cells.
Here we show that the density of M cells in the gut epithelium directly influences oral prion disease pathogenesis and susceptibility. In the specific absence of M cells, the accumulation of prions in Peyer’s patches and subsequent neuroinvasion was blocked, demonstrating that prion translocation across the gut epithelium in association with M cells is essential to establish host infection. Our data also imply that an absence or reduction in M cell-abundance may significantly reduce susceptibility to many naturally acquired prion diseases such as vCJD in humans, CWD in cervids and natural sheep scrapie. For example, in the UK most clinical vCJD cases have predominantly occurred in young adults (median age at death, ~28 years) [4], but epidemiological data indicate that this age-related susceptibility is not simply due to the exposure of young individuals to greater levels of the BSE agent through dietary preference [70]. We have previously shown that the density of functionally mature M cells in the Peyer’s patches of aged mice is substantially reduced [71], suggesting that the reduced susceptibility of aged mice to oral prion infection [72] is at least in part due to the inefficient uptake of prions from the gut lumen by M cells.
We also show that increased M-cell density at the time of oral exposure exacerbated prion disease pathogenesis: the uptake of prions from the gut lumen was enhanced, and as a consequence, survival times were decreased and disease susceptibility was increased approximately 10-fold. The density of M cells in the gut epithelium can be modified by the presence of certain pathogenic bacteria or inflammatory stimuli [25, 39, 40]. Although these stimuli may have multiple effects on the gut epithelium which can influence the integrity of this barrier, data in the current study provide a significant advance in our understanding of how factors that increase the density of M cells in the gut epithelium may increase susceptibility to orally-acquired prion infection. For example, the enteroinvasive bacterium Salmonella Typhimurium can specifically and rapidly transform enterocytes in the FAE of Peyer’s patches into M cells in order to facilitate host infection [25]. Furthermore, an independent study has shown that concurrent infection with S. Typhimurium significantly increased oral prion disease susceptibility [43]. Although this observation was originally attributed to the colitis induced by the pathogen in the large intestine, data in the current study suggest a role for effects on M cells in the small intestine cannot be excluded.
During the BSE epidemic in the UK it is estimated that approximately 500,000 infected cattle were slaughtered for human consumption [73]. Despite the widespread dietary exposure of the UK human population to BSE prions, clinical cases of vCJD have fortunately been rare (Ref. [4]; 178 definite or probable cases, as of 5th December 2016; www.cjd.ed.ac.uk/documents/figs.pdf). This implies that the ability to acquire prions from the gut lumen may differ between individuals. Studies using transgenic mice expressing human PrPC suggest that the transmission of BSE to humans is restricted by a significant species barrier [74]. After interspecies prion exposure, the processing and amplification of prions upon FDC in lymphoid tissues is important for their adaptation to the new host and to achieve neuroinvasion [75, 76]. Thus, it is plausible that factors which increase the density of M cells in the small intestine may enable a greater burden of prions to enter Peyer’s patches, increasing the probability that more will be able to avoid clearance by cells such as macrophages, [11, 77]. This may provide a greater opportunity for prion quasi-species present within the original inoculum with zoonotic potential to be selected and undergo adaptation and amplification upon FDC [78]. These effects may help to reduce the transmission barrier to some orally acquired prion strains.
Enterocytes within the FAE overlying the Peyer’s patches specifically contain large LAMP1+ endosomes [16]. A detailed high resolution IHC-based study has shown that within the first day following oral exposure of mice to prions, PrPSc was detected within these large LAMP1+ endosomes of FAE enterocytes, and to a lesser extent in M cells [16]. These FAE enterocyte-associated endosomes have been proposed as a potential M cell-independent route through which lumenal proteins and prions may also be taken up into Peyer’s patches [16]. In the current study the presence and abundance of the large LAMP1+ endosomes within FAE enterocytes was unaffected in M cell-deficient RANKΔIEC mice. These data clearly show that the specific lack of M cells in the FAE, rather than an absence of the large LAMP1+ endosomes within FAE enterocytes, was responsible for the blocked prion accumulation in Peyer’s patches. Furthermore, the accumulation of prions in the Peyer’s patches, MLN and spleens of orally-exposed M cell-deficient RANKΔIEC mice was undetectable up to at least 440 d after exposure. As abundant prion accumulation is typically evident in these tissues in conventional (WT) mice by 105 d after exposure, this implies that in the absence of M cells, any prions that do enter the Peyer’s patches via alternative routes may be of insufficient magnitude to establish infection. Indeed PrPSc was also undetectable in the lymphoid tissues and CNS of these mice up to at least 440 dpi. Instead the prions that are acquired from the gut lumen by these M cell-independent routes are most likely sequestered and destroyed by cells such as macrophages, which are considered to degrade prions [77], rather than being efficiently transported to FDC where they undergo amplification before neuroinvasion [7, 10, 13, 15]. RANKΔIEC mice show reduced IgA production and delayed germinal centre responses in their Peyer’s patches [23]. This suggests that antigens that are transcytosed by M cells are preferentially targeted to the FDC-containing B-cell follicles to initiate antibody responses. Therefore, M cells, in contrast to FAE enterocytes with large LAMP1+ endosomes, may be considered to facilitate the efficient transfer of prions from the gut lumen to FDC in the B-cell follicles of Peyer’s patches.
A separate IHC-based study also has proposed that the uptake of scrapie-affected brain homogenate across the jejunal epithelium of lambs occurs independently of M cells [34]. However, if prions do efficiently establish infection within Peyer’s patches after their translocation across the gut epithelium by enterocytes, one would not expect the specific absence of M cells in RANKΔIEC mice to block oral prion disease susceptibility. In the above in vivo study [34], large quantities of scrapie-affected brain homogenate were injected directly into the lumen of ligated loops of jejunum. The presence of a large bolus of prions concentrated within the lumen of these ligated loops may have facilitated prion uptake into alternative cellular compartments to those utilized following exposure to physiologically relevant doses via the oral cavity. Although evidence of prions (PrPd) was detected in the underlying LP of these lambs, it was interesting to note that no intraepithelial PrPd accumulations were detected by IHC [34]. Whether the prions were transiently present in enterocytes and/or M cells soon after exposure, but at levels below the reliable detection limit or in a conformation which could not be detected by the IHC protocols used, remains to be determined. By comparison, in the study by Kujala and colleagues discussed above [16], PrPSc was detected within the FAE during the first day after oral exposure using highly sensitive cryo-immunogold electron microscopy. M cells unlike the neighbouring enterocytes have a very narrow cytoplasm due to the presence of the MNP-containing basolateral pocket [20]. Thus it is also plausible that the prion transit time through M cells may be extremely rapid, restricting the ability of IHC to reliably detect low levels of prions or other particles which are being transcytosed through them. Surgical manipulation and manual compression of the intestine can temporarily inhibit intestinal motility and induce intestinal inflammation with activation of resident macrophages, as occurs during postoperative ileus [79, 80]. These factors may have a significant influence on the uptake of prions from the lumen of surgically-ligated intestinal loops.
Using extremely sensitive PrPSc-based detection assays, two independent studies reported the presence of low/trace levels of prions in the blood-stream within minutes of oral exposure [81, 82]. The cellular route through which the prions initially gained access to the blood-stream was not determined in these studies. Urayama and colleagues [82] suggested that the levels of PrPSc that initially contaminated the blood-stream after oral exposure were sufficient to initiate infection in the brain. However, data from several studies show that prion replication upon FDC in Peyer’s patches in the small intestine is essential to establish host infection after oral exposure [7, 8, 10–13]. Furthermore, in the temporary absence of FDC at the time of oral exposure, prion disease susceptibility is blocked [6]. Thus although PrPSc may be detected in the blood-stream soon after oral exposure using highly sensitive assays [81, 82], data elsewhere indicate that the levels of prions that are initially within it are unable to directly establish host infection and achieve neuroinvasion.
After uptake by M cells, CD11c+ classical DC are considered to deliver prions towards FDC, as their transient depletion reduces susceptibly to oral prion disease [8]. A partial reduction in CD11c+ immunostaining was observed in the SED of Peyer’s patches from RANKΔIEC mice, implying a partial reduction in the abundance of these cells. M cells specifically express the chemokine CCL9 [22] which mediates the attraction of certain classical DC populations towards the FAE [83]. Thus, the reduced CD11c+ immunostaining in the SED of RANKΔIEC mice may be a consequence of the absence of attraction of CD11c+ cells towards the basolateral pockets of M cells. This partial reduction in CD11c+ immunostaining in SED region alone could not account for the complete block of prion accumulation observed in RANKΔIEC mice, as our previous data show that the depletion of CD11c+ cells (>85%) prior to oral exposure does not block neuroinvasion [8]. Although the germinal centre response is delayed in RANKΔIEC mice [23], our data suggested that FDC status was unaffected in these mice. Furthermore, the FDC in the Peyer’s patches, MLN and spleen of these mice were capable of accumulating high levels of PrPSc after injection of prions by the i.p. route. We have also previously shown that an absence of germinal centres themselves does not influence peripheral prion disease pathogenesis [84].
The GALT in the small intestine such as the Peyer’s patches, not those in the large intestine, are the major early sites of prion uptake, replication and neuroinvasion after oral exposure [11, 13, 16]. RANKL-RANK signalling is also necessary for the induction of M cell-differentiation within the large intestine, but in contrast to its role in the small intestine, it does not induce their maturation. As a consequence, GP2-expressing functionally mature M cells are scarce in the FAE overlying the large intestinal GALT [64]. Consistent with this, systemic RANKL-treatment also does not increase the abundance of functionally mature M cells in the FAE overlying the caecal patches or in the conventional epithelium of large intestine [64]. These data suggest that the effects of systemic RANKL-treatment on oral prion disease pathogenesis observed in the current study were due to an increased abundance of mature M cells specifically in the small intestine.
In the steady state, functionally mature M cells are confined to the FAE overlying the Peyer’s patches and are extremely rare within the villous epithelium. However, systemic RANKL-treatment, as used here, significantly increases the abundance of mature M cells in the FAE overlying Peyer’s patches and throughout the villous epithelium [22, 35, 64]. Therefore, it is plausible that the effects of systemic RANKL-treatment on oral prion disease pathogenesis were in part due to the enhanced uptake of prions by villous M cells, facilitating their more efficient delivery to the MLN. Using LTβ-/- mice reconstituted with WT bone marrow (WT→LTβ-/- mice), we generated mice that lacked Peyer’s patches or other M cell-containing GALT structures (ILF) in their small intestines, but retained MLN which contained mature FDC. If the major effect of RANKL-treatment on oral prion pathogenesis was due to uptake by villous M cells and enhanced delivery from the LP to the MLN, the accumulation of prions in the MLN would likewise be enhanced in these mice after RANKL-treatment. However, our data clearly show that RANKL-treatment did not enhance the accumulation of prions within the MLN of WT→LTβ-/- mice. This demonstrates that the major effect of RANKL-treatment on oral prion disease pathogenesis and susceptibility was due to the increased uptake of prions across the FAE overlying the Peyer’s patches in the small intestine. The absence of detectable levels of prion infectivity in the MLN at the time of analysis suggests that any low levels of prions that do reach this tissue immediately after oral exposure are either not delivered to FDC in the MLN as efficiently as they are in the Peyer’s patches, or are of insufficient magnitude to establish infection on FDC and are thus most likely degraded by macrophages [11, 77]. Our IHC analysis implied that the abundance of CD68+ macrophages was increased in the LP after RANKL-treatment, suggesting that it is also plausible that any prions that had been acquired by villous M cells were subsequently sequestered and destroyed in the LP by macrophages. Classical DC in the LP of the intestine are considered to deliver lumenal antigens directly to MLN [65]. Here, RANKL-treatment of RANKΔIEC mice did not restore prion accumulation in their Peyer’s patches and MLN following oral exposure, demonstrating that RANKL-treatment did not alter the uptake of prions from the gut lumen by non-epithelial cells, such as classical DC. Our data suggest that direct sampling of the lumenal contents by classical DC in the LP [60–63] is also unlikely to contribute significantly to prion uptake from the gut lumen, as this too would result in the direct delivery of prions to the MLN [65].
In conclusion, we show that the initial uptake and transfer of prions across the gut epithelium in association with M cells is essential to establish host infection. Importantly, we also demonstrate that the density of M cells in the FAE overlying the Peyer’s patches in the small intestine directly controls the efficiency of oral prion infection. In the specific absence of M cells, the uptake and accumulation of prions in Peyer’s patches and their subsequent spread to the MLN and spleen is blocked. In contrast, oral prion disease susceptibility was enhanced approximately 10-fold in mice in which M cell-deficiency in the gut epithelium was increased. Thus, M cells could be considered as the gatekeepers of oral prion infection whose density directly limits or enhances disease susceptibility. Further studies are necessary to determine whether most orally acquired prion strains similarly exploit intestinal M cells to establish host infection after oral exposure, but data from independent in vivo and in vitro studies using mouse-passaged RML scrapie prions [30], Fukuoka-1 prions [31], BSE prions [32] and 263K hamster prions [17] imply a similar requirement. Antigen sampling M cells are also present in the FAE overlying the NALT in the nasal cavity [44, 45], but data from the analysis of prion disease pathogenesis in hamsters implies that the requirement for M cell-mediated uptake may vary depending on the route of exposure [85]. After intra-nasal exposure some transient uptake of 263K prions was observed in M cells within the FAE overlying the NALT, but a greater magnitude of paracellular transport across the epithelia within the nasal cavity was also noted [85]. Although certain concurrent pathogen infections, inflammatory stimuli and aging may have multiple effects on the gut epithelium, our data suggest that factors such as these that can modify M cell-density in the small intestine [25, 39, 40, 71] may represent important risk factors which can significantly influence susceptibility to orally-acquired prion infections. Our data also raise the possibility that the density of M cells in the gut epithelium may similarly influence susceptibility to other important orally-acquired bacterial and viral pathogens which are considered to exploit M cells to infect the host [24–28].
All studies using experimental mice and regulatory licences were approved by both The Roslin Institute’s and University of Edinburgh’s ethics committees. All animal experiments were carried out under the authority of a UK Home Office Project Licence (PPL60/4325) within the terms and conditions of the strict regulations of the UK Home Office ‘Animals (scientific procedures) Act 1986’. Where necessary, anaesthesia appropriate for the procedure was administered, and all efforts were made to minimize harm and suffering. Mice were humanely culled by a UK Home Office Schedule One method.
The following mouse strains were used in this study where indicated: C57BL/6J; Villin-cre (Tg(Vil-cre)997Gum/J strain; The Jackson Laboratory, Bar Harbor, ME); RANKfl/fl, which have loxP sites flanking exons 2 and 3 of Tnfrsf11a (which encodes RANK) [23]; LTβ-/- [86]; tga20, which overexpress PrPC [59]. All mice were bred and maintained on a C57BL/6J background and housed under SPF conditions.
Bone-marrow from the femurs and tibias of donor mice was prepared as single-cell suspensions (3x107–4x107 viable cells/ml) in HBSS (Life Technologies, Paisley, UK). Recipient adult LTβ-/- mice (6–8 weeks old) were γ-irradiated (10 Gy) and 24 h later reconstituted with 100 μl bone-marrow by injection into the tail vein.
Glutathione S-transferase—RANKL fusion protein was prepared as described [35]. To enhance M-cell-density in the gut epithelium mice were treated with RANKL in vivo as previously described [22, 35]: d 0 injected with RANKL by a combination of i.p. and subcutaneous injection (50 μg/ea.); d 1, 50 μg RANKL by subcutaneous injection; d 2, 50 μg RANKL by subcutaneous injection; d 3, 50 μg RANKL by subcutaneous injection. Mice were orally exposed to prions or gavaged with fluorescent microbeads on d 2 after the onset of RANKL treatment.
For oral exposure, mice were fed individual food pellets doused with 50 μl of a 1% (containing approximately 2.5 X 104 i.c. ID50 units) or 0.1% (w/v) dilution of scrapie brain homogenate prepared from mice terminally-affected with ME7 scrapie prions according to our standard protocol [7–9, 11, 13, 72]. During the dosing period mice were individually housed in bedding- and food-free cages. Water was provided ad libitum. A single prion-dosed food pellet was then placed in the cage. The mice were returned to their original cages (with bedding and food ad libitum) as soon as the food pellet was observed to have been completely ingested. The use of bedding- and additional food-free cages ensured easy monitoring of consumption of the prion-contaminated food pellet. For i.p. exposure, mice were injected with 20 μl of a 1% dilution of scrapie brain homogenate. Following prion exposure, mice were coded and assessed weekly for signs of clinical disease and culled at a standard clinical endpoint. The clinical endpoint of disease was determined by rating the severity of clinical signs of prion disease exhibited by the mice. Following clinical assessment, mice were scored as “unaffected”, “possibly affected” and “definitely affected” using standard criteria which typically present in mice clinically-affected with ME7 scrapie prions. Clinical signs following infection with the ME7 scrapie agent may include: weight-loss, starry coat, hunched, jumpy behaviour (at early onset) progressing to limited movement, upright tail, wet genitals, decreased awareness, discharge from eyes/blinking eyes, ataxia of hind legs. The clinical endpoint of disease was defined in one of the following ways: i) the day on which a mouse received a second consecutive “definite” rating; ii) the day on which a mouse received a third “definite” rating within four consecutive weeks; iii) the day on which a mouse was culled in extremis. Survival times were recorded for mice that did not develop clinical signs of disease or were culled when they showed signs of intercurrent disease. Prion diagnosis was confirmed by histopathological assessment of vacuolation in the brain. For the construction of lesion profiles, vacuolar changes were scored in nine distinct grey-matter regions of the brain as described [87].
For bioassay of prion infectivity individual MLN or spleen were prepared as 1% (wt/vol) homogenates in physiological saline. For each tissue homogenate groups of tga20 indicator mice (n = 4/homogenate) were injected i.c. with 20 μl of each homogenate. The prion infectivity titre in each sample was determined from the mean incubation period in the indicator mice, by reference to a dose/incubation period response curve for ME7 scrapie-infected spleen tissue serially titrated in tga20 mice using the relationship: y = 9.4533–0.0595x (where y is log ID50 U/20 μl of homogenate, and x is the incubation period; R2 = 0.9562).
Whole-mount immunostaining was performed as previously described [9]. Peyer’s patches, NALT and pieces of small intestines were fixed with BD Cytofix/Cytoperm (BD Biosciences, Oxford, UK), and subsequently immunostained with rat anti-mouse GP2 mAb (MBL International, Woburn, MA; 5 μg/ml). Following addition of primary Ab, tissues were stained with Alexa Fluor 488-conjugated anti-rat IgG Ab (Life Technologies), rhodamine-conjugated Ulex europaeus agglutinin I (UEA-1; Vector Laboratories Inc., Burlingame, CA; 20 μg/ml) and Alexa Fluor 647-conjugated phalloidin to detect f-actin (Life Technologies; 4 U/ml).
Intestines, MLNs and spleens were also removed and snap-frozen at the temperature of liquid nitrogen. Serial frozen sections (6 μm in thickness) were cut on a cryostat and immunostained with the following antibodies: FDC were visualized by staining with mAb 7G6 to detect CR2/CR1 (CD21/35; BD Biosciences; 1 μg/ml) or mAb 8C12 to detect CR1 (CD35; BD Biosciences; 1.25 μg/ml); cellular PrPC was detected using PrP-specific polyclonal antibody (pAb) 1B3 [88] (1/1000 dilution); B cells were detected using rat anti-mouse B220 mAb (clone RA3-6B2, Life Technologies; 5 μg/ml); MNP were detected using hamster anti-mouse CD11c mAb (clone N418, Bio-Rad, Kidlington, UK; 5 μg/ml) or rat anti-mouse CD68 mAb (clone FA-11, Biolegend, Cambridge, UK; 5 μg/ml); rat anti-mouse CD107a (clone 1D4B; Biolegend; 2.5 μg/ml) to detect LAMP1; nerve synapses were detected using rabbit anti-synaptophysin 1 (Synaptic Systems, Göttingen, Germany; 1/150 dilution). For the detection of SPIB in paraformaldehyde-fixed sections, antigen retrieval was performed with citrate buffer (pH 7.0, 121°C, 5 min.) prior to immunostaining with sheep anti-mouse SPIB polyclonal antibody (R&D Systems, Abingdon, UK; 0.4 μg/ml). Appropriate species and immunoglobulin isotype control Ab were used as controls (S5 Fig). Where appropriate, sections were counter-stained with DAPI (2.86 μM) to detect cell nuclei (Life Technologies).
For the detection of disease-specific PrP (PrPd) in intestines, MLN, spleens and brains, tissues were fixed in periodate-lysine-paraformaldehyde fixative and embedded in paraffin wax. Sections (thickness, 6 μm) were deparaffinised, and pre-treated to enhance the detection of PrPd by hydrated autoclaving (15 min, 121°C, hydration) and subsequent immersion formic acid (98%) for 10 min. Sections were then immunostained with 1B3 PrP-specific pAb (1/1000 dilution). For the detection of astrocytes, brain sections were immunostained with anti-glial fibrillary acidic protein (GFAP; DAKO, Ely, UK; 1/400 dilution). For the detection of microglia, deparaffinised brain sections were first pre-treated with citrate buffer and subsequently immunostained with anti-ionized calcium-binding adaptor molecule 1 (Iba1; Wako Chemicals GmbH, Neuss, Germany; 0.5 μg/ml). For the detection of FDC in intestines, MLN and spleens, deparaffinised sections were first pre-treated with Target Retrieval Solution (DAKO) and subsequently immunostained with anti-CD21/35 mAb. PET immunoblot analysis was used to confirm the PrPd detected by immunohistochemistry was proteinase K-resistant PrPSc [57]. Membranes were subsequently immunostained with 1B3 PrP-specific pAb (1/4000 dilution).
For light microscopy, following the addition of primary antibodies, biotin-conjugated species-specific secondary antibodies (Stratech, Soham, UK) were applied and immunolabelling was revealed using HRP-conjugated to the avidin-biotin complex (ABC kit, Vector Laboratories) and visualized with DAB (Sigma, Dorset, UK). Sections were counterstained with haematoxylin to distinguish cell nuclei. For fluorescent microscopy, following the addition of primary antibody, streptavidin-conjugated or species-specific secondary antibodies coupled to Alexa Fluor 488 (green), Alexa Fluor 594 (red) or Alexa Fluor 647 (blue) dyes (Life Technologies) were used. Sections were counterstained with either DAPI or Alexa Fluor 647-conjugated phalloidin and subsequently mounted in fluorescent mounting medium (DAKO). Images of whole-mount immunostained tissues and cryosections were obtained using a Zeiss LSM710 confocal microscope (Zeiss, Welwyn Garden City, UK).
For morphometric analysis, images were analysed using ImageJ software (http://rsb.info.nih.gov/ij/) as described on coded sections [89]. Background intensity thresholds were first applied using an ImageJ macro which measures pixel intensity across all immunostained and non-stained areas of the images. The obtained pixel intensity threshold value was then applied in all subsequent analyses. Next, the number of pixels of each colour (black, red, green, yellow etc.) were automatically counted. For these analyses, data are presented as the proportion of positively-stained pixels for a given IHC marker per total number of pixels (all colours) in the specific area of interest (eg: SED, FAE, LP etc.). In each instance, typically 3–6 images were analysed per mouse, from tissues from multiple mice per group (n = 4–8 mice/group). Full details of all the sample sizes for each parameter analysed are provided in every figure legend.
Mice were given a single oral gavage of 2x1011 of Fluoresbrite Yellow Green labelled 200 nm microbeads (Polysciences, Eppelheim, Germany) in 200 μl PBS. Mice were culled 24 h later and Peyer’s patches and small intestine segments were snap-frozen at the temperature of liquid nitrogen. Serial frozen sections (6 μm in thickness) were cut on a cryostat and counterstained with DAPI. Images of SED from three Peyer’s Patches (duodenal, jejunal and ileal) and 8 LP areas per mouse (n = 3–4 mice/group) from 3 non-sequential sections (total 21–31 SED, or 24 LP areas per mouse studied) were typically acquired using a Nikon Eclipse E400 fluorescent microscope using Micro Manager (http://www.micro-manager.org). For example, each Peyer’s patch was trimmed until at least one SED region was visible and 20 sections collected. The 1st, 10th and 20th sections were then analysed. Tissue auto-fluorescence was subtracted from displayed images using ImageJ, the size of the area of interest in each section was then measured and the number of beads determined using the cell counter function in ImageJ and the bead density calculated.
Intestinal crypts were dissociated from mouse small intestine using Gentle Cell Dissociation Reagent (Stemcell Tech, Cambridge, UK) and used establish enteroids by cultivation in Matrigel (BD Bioscience) and Intesticult medium (Stemcell Tech) as described [23, 90]. Where indicated, some wells were treated with RANKL (100 ng/ml). Enteroids were cultivated in triplicate and either passaged after 5 d of cultivation [90] or harvested for mRNA expression analysis as described [23].
Total RNA was isolated from the enteroid cultures using RNA-Bee (AMS Biotechnology, Oxfordshire, UK) followed by treatment with DNase I (Ambion, Warrington, UK). First strand cDNA synthesis was performed using 1 μg of total RNA and the First Strand cDNA Synthesis kit (GE Healthcare, Bucks, UK) as described by the manufacturer. PCR was performed using the Platinum-SYBR Green qPCR SuperMix-UDG kit (Life Technologies) and the Stratagene Mx3000P real-time qPCR system (Stratagene, CA, USA). The qPCR primers (S1 Table) were designed using Primer3 software [91]. The cycle threshold values were determined using MxPro software (Stratagene) and normalized relative to Gapdh.
All data are presented as mean ± SD. Unless indicated otherwise, differences between groups were analysed by a Student's t-test. In instances where there was evidence of non-normality (identified by the Kolmogorov-Smirnov, D’Agostino & Pearson omnibus, or Shapiro-Wilk normality tests), data were analysed by a Mann-Whitney U test. Survival rates were analysed using the Log-rank (Mantel-Cox) test. Values of P<0.05 were accepted as significant.
|
10.1371/journal.pcbi.1002408 | Impact of Network Structure and Cellular Response on Spike Time Correlations | Novel experimental techniques reveal the simultaneous activity of larger and larger numbers of neurons. As a result there is increasing interest in the structure of cooperative – or correlated – activity in neural populations, and in the possible impact of such correlations on the neural code. A fundamental theoretical challenge is to understand how the architecture of network connectivity along with the dynamical properties of single cells shape the magnitude and timescale of correlations. We provide a general approach to this problem by extending prior techniques based on linear response theory. We consider networks of general integrate-and-fire cells with arbitrary architecture, and provide explicit expressions for the approximate cross-correlation between constituent cells. These correlations depend strongly on the operating point (input mean and variance) of the neurons, even when connectivity is fixed. Moreover, the approximations admit an expansion in powers of the matrices that describe the network architecture. This expansion can be readily interpreted in terms of paths between different cells. We apply our results to large excitatory-inhibitory networks, and demonstrate first how precise balance – or lack thereof – between the strengths and timescales of excitatory and inhibitory synapses is reflected in the overall correlation structure of the network. We then derive explicit expressions for the average correlation structure in randomly connected networks. These expressions help to identify the important factors that shape coordinated neural activity in such networks.
| Is neural activity more than the sum of its individual parts? What is the impact of cooperative, or correlated, spiking among multiple cells? We can start addressing these questions, as rapid advances in experimental techniques allow simultaneous recordings from ever-increasing populations. However, we still lack a general understanding of the origin and consequences of the joint activity that is revealed. The challenge is compounded by the fact that both the intrinsic dynamics of single cells and the correlations among then vary depending on the overall state of the network. Here, we develop a toolbox that addresses this issue. Specifically, we show how linear response theory allows for the expression of correlations explicitly in terms of the underlying network connectivity and known single-cell properties – and that the predictions of this theory accurately match simulations of a touchstone, nonlinear model in computational neuroscience, the general integrate-and-fire cell. Thus, our theory should help unlock the relationship between network architecture, single-cell dynamics, and correlated activity in diverse neural circuits.
| New multielectrode and imaging techniques are revealing the simultaneous activity of neural ensembles and, in some cases, entire neural populations [1]–[4]. This has thrust upon the computational biology community the challenge of characterizing a potentially complex set of interactions – or correlations – among pairs and groups of neurons.
Beyond important and rich challenges for statistical modeling [5], the emerging data promises new perspectives on the neural encoding of information [6]. The structure of correlations in the activity of neuronal populations is of central importance in understanding the neural code [7]–[13]. However, theoretical [9]–[11], [14]–[16], and empirical studies [17]–[19] do not provide a consistent set of general principles about the impact of correlated activity. This is largely because the presence of correlations can either strongly increase or decrease the fidelity of encoded information depending on both the structure of correlations across a population and how their impact is assessed.
A basic mechanistic question underlies the investigation of the role of collective activity in coding and signal transmission: How do single-cell dynamics, connection architecture, and synaptic dynamics combine to determine patterns of network activity? Systematic answers to this question would allow us to predict how empirical data from one class of stimuli will generalize to other stimulus classes and recording sites. Moreover, a mechanistic understanding of the origin of correlations, and knowledge of the patterns we can expect to see under different assumptions about the underlying networks, will help resolve recent controversies about the strength and pattern of correlations in mammalian cortex [1], [20], [21]. Finally, understanding the origin of correlations will inform the more ambitious aim of inferring properties of network architecture from observed patterns of activity [22]–[24].
Here, we examine the link between network properties and correlated activity. We develop a theoretical framework that accurately predicts the structure of correlated spiking that emerges in a widely used model – recurrent networks of general integrate and fire cells. The theory naturally captures the role of single cell and synaptic dynamics in shaping the magnitude and timescale of spiking correlations. We focus on the exponential integrate and fire model, which has been shown to capture membrane and spike responses of cortical neurons [25]; however, the general approach we take can be applied to a much broader class of neurons, a point we return to in the Discussion.
Our approach is based on an extension of linear response theory to networks [24], [26]. We start with a linear approximation of a neuron's response to an input. This approximation can be obtained explicitly for many neuron models [27]–[29], and is directly related to the spike triggered average [30]. The correlation structure of the network is then estimated using an iterative approach. As in prior work [31]–[33], the resulting expressions admit an expansion in terms of paths through the network.
We apply this theory to networks with precisely balanced inhibition and excitation in the inputs to individual cells. In this state individual cells receive a combination of excitatory and inhibitory inputs with mean values that largely cancel. We show that, when timescales and strengths of excitatory and inhibitory connections are matched, only local interactions between cells contribute to correlations. Moreover, our theory allows us to explain how correlations are altered when precise tuning balance is broken. In particular, we show how strengthening inhibition may synchronize the spiking activity in the network. Finally, we derive results which allow us to gain an intuitive understanding of the factors shaping average correlation structure in randomly connected networks of neurons.
Our goal is to understand how the architecture of a network shapes the statistics of its activity. We show how correlations between spike trains of cells can be approximated using response characteristics of individual cells along with information about synaptic dynamics, and the structure of the network. We start by briefly reviewing linear response theory of neuronal responses [28], [34], [35], and then use it to approximate the correlation structure of a network.
To illustrate the results we consider a network of nonlinear integrate-and-fire (IF) neurons with membrane potentials modeled by(1)Here is the leak reversal potential, and represents the mean synaptic input current from parts of the system not explicitly modeled. A spike-generating current may be included to emulate the rapid onset of action potentials. Unless otherwise specified, we utilize the exponential IF model (EIF), so that [25]. Cells are subject to internally induced fluctuations due to channel noise [36], and externally induced fluctuations due to inputs not explicitly modelled [37]. We model both by independent, Gaussian, white noise processes, [38]. An external signal to cell is represented by .
Upon reaching a threshold , an action potential is generated, and the membrane potential is reset to , where it is held constant for an absolute refractory period . The output of cell is characterized by the times, , at which its membrane potential reaches threshold, resulting in an output spike train . Synaptic interactions are modeled by delayed -functions(2)The matrix contains the synaptic kernels, while the matrix contains the synaptic weights, and hence defines the network architecture. In particular, if is the membrane conductance, is the area under a post-synaptic current evoked in cell by a spike in the presynaptic cell , and along with the membrane and synaptic time constants, determines the area under a post-synaptic potential. represents the absence of a synaptic connection from cell to cell .
Table 1 provides an overview of all parameters and variables.
Neuronal network models are typically described by a complex system of coupled nonlinear stochastic differential equations. Their behavior is therefore difficult to analyze directly. We will use linear response theory [28], [34], [35], [39] to approximate the cross-correlations between the outputs of neurons in a network. We first review the linear approximation to the response of a single cell. We illustrate the approach using current-based IF neurons, and explain how it can be generalized to other models in the Discussion.
The membrane potential of an IF neuron receiving input , with vanishing temporal average, , evolves according to(3)The time-dependent firing rate, , is determined by averaging the resulting spike train, , across different realizations of noise, , for fixed . Using linear response theory, we can approximate the firing rate by(4)where is the (stationary) firing rate when . The linear response kernel, , characterizes the firing rate response to first order in . A rescaling of the function gives the spike-triggered average of the cell, to first order in input strength, and is hence equivalent to the optimal Weiner kernel in the presence of the signal . [39], [40]. In Figure 1, we compare the approximate firing rate obtained from Eq. (4) to that obtained numerically from Monte Carlo simulations.
The linear response kernel depends implicitly on model parameters, but is independent of the input signal, , when is small relative to the noise . In particular, is sensitive to the value of the mean input current, . We emphasize that the presence of the background noise, , in Eq. (3) is essential to the theory, as noise linearizes the transfer function that maps input to output. In addition, when applying linear response methods, there is an implicit assumption that the fluctuations of the input do not have a significant effect on the response properties of the cell.
The linear response kernel can be used to approximate the response of a cell to an external input. However, the situation is more complicated in a network where a neuron can affect its own activity through recurrent connections. To extend the linear response approximation to networks we follow the approach introduced by Lindner et al. [26]. Instead of using the linear response kernel to approximate the firing rate of a cell, we use it to approximate a realization of its output(5)Here represents a realization of the spike train generated by an integrate-and-fire neuron obeying Eq. (3) with .
Our central assumption is that a cell acts approximately as a linear filter of its inputs. Note that Eq. (5) defines a mixed point and continuous process, but averaging in Eq. (5) over realizations of leads to the approximation in Eq. (4). Hence, Eq. (5) is a natural generalization of Eq. (4) with the unperturbed output of the cell represented by the point process, , instead of the firing rate, .
We first use Eq. (5) to describe spontaneously evolving networks where . Equation (1) can then be rewritten as(6)where and represents the temporal average.
Lindner et al. used Eq. (5) as an ansatz to study the response of an all–to–all inhibitory network. They postulated that the spiking output of cell in the network, can be approximated in the frequency domain bywhere are the zero-mean Fourier transforms of the processes , and for all other quantities (see Table 1 for the Fourier transform convention). The term in parentheses is the Fourier transform of the zero-mean synaptic input, , in Eq. (6), and represents a realization of the spiking output of cell in the absence of synaptic fluctuations from the recurrent network (i.e assuming ). In matrix form this ansatz yields a simple self-consistent approximation for the firing activities which can be solved to givewhere the interaction matrix has entries defined by . When averaged against its conjugate transpose, this expression yields an approximation to the full array of cross-spectra in the recurrent network:(7)
We next present a distinct derivation of this approximation which allows for a different interpretation of the ansatz given by Eq. (5). We iteratively build to the approximation in Eq. (7), showing how this expression for the correlation structure in a recurrent network can be obtained by taking into account the paths through the network of increasing length.
We start with realizations of spike trains, , generated by IF neurons obeying Eq. (6) with . This is equivalent to considering neurons isolated from the network, with adjusted DC inputs (due to mean network interactions). Following the approximation given by Eq. (5), we use a frozen realization of all to find a correction to the output of each cell, with set to the mean-adjusted synaptic input,As noted previously, the linear response kernel is sensitive to changes in the mean input current. It is therefore important to include the average synaptic input in the definition of the effective mean input, .
The input from cell to cell is filtered by the synaptic kernel . The linear response of cell to a spike in cell is therefore captured by the interaction kernel , defined above asThe output of cell in response to mean-adjusted input, , from cell can be approximated to first order in input strength using the linear response correction(8)We explain how to approximate the stationary rates, , in the Methods.
The cross-correlation between the processes in Eq. (8) gives a first approximation to the cross-correlation function between the cells,which can be simplified to give(9)where we used . Ostojic et al. obtained an approximation closely related to Eq. (9). [24] They first obtained the cross-correlation between a pair of neurons which either receive a common input or share a monosynaptic connection. This can be done using Eq. (4), without the need to introduce the mixed process given in Eq. (5). Ostojic et al. then implicitly assumed that the correlations not due to one of these two submotifs could be disregarded. The correlation between pairs of cells which were mutually coupled (or were unidirectionally coupled with common input) was approximated by the sum of correlations introduced by each submotif individually.
Equation (9) provides a first approximation to the joint spiking statistics of cells in a recurrent network. However, it captures only the effects of direct synaptic connections, represented by the second and third terms, and common input, represented by the last term in Eq. (9). The impact of larger network structures, such as loops and chains are not captured, although they may significantly impact cross-correlations [41]–[43]. Experimental studies have also shown that local cortical connectivity may not be fully random [44]–[46]. It is therefore important to understand the effects on network architecture on correlations.
We therefore propose an iterative approach which accounts for successively larger connectivity patterns in the network [32], [33]. We again start with , a realization of a single spike train in isolation. Successive approximations to the output of cells in a recurrent network are defined by(10)
To compute the correction to the output of a neuron, in the first iteration we assume that its inputs come from a collection of isolated cells: When , Eq. (10) takes into account only inputs from immediate neighbors, treating each as disconnected from the rest of the network. The corrections in the second iteration are computed using the approximate cell responses obtained from the first iteration. Thus, with , Eq. (10) also accounts for the impact of next nearest neighbors. Successive iterations include the impact of directed chains of increasing length: The isolated output from an independent collection of neurons is filtered through stages to produce the corrected response (See Figure 2.)
Notation is simplified when this iterative construction is recast in matrix form to obtain(11)where and are length column vectors, and represents a -fold matrix convolution of with itself. We define the convolution of matrices in the Methods.
The approximation to the matrix of cross-correlations can be written in terms of the interaction kernels, , and the autocorrelations of the base processes as(12)where , and is the -fold matrix convolution of with itself.
Eq. (12) can be verified by a simple calculation. First, Eq. (11) directly implies thatwhich we may use to find, for each ,(13)Since , Eq. (13) is equivalent to Eq. (12).
If we apply the Fourier transform, to Eq. (12), we find that for each ,(14)where denotes the conjugate transpose of the matrix . As before, the zero-mean Fourier transforms of the processes are defined by , and for all other quantities.
Defining to be the spectral radius of the matrix , when , we can take the limit in Eq. (14) [47], [48], to obtain an approximation to the full array of cross-spectra(15)As noted previously, this generalizes the approach of Lindner et al. [26] (also see [13]). In the limit , directed paths of arbitrary length contribute to the approximation. Equation (15) therefore takes into account the full recurrent structure of the network. Note that Eq. (15) may be valid even when . However, in this case the series in Eq. (14) do not converge, and hence the expansion of the correlations in terms of paths through the network is invalid. We confirmed numerically that for all of the networks and parameters we considered.
Finally, consider the network response to external signals, , with zero mean and finite variance. The response of the neurons in the recurrent network can be approximated iteratively bywhere and . External signals and recurrent synaptic inputs are both linearly filtered to approximate a cell's response, consistent with a generalization of Eq. (4). As in Eq. (12), the approximation to the matrix of correlations iswhere is the covariance matrix of the external signals. We can again take the Fourier transform and the limit , and solve for . If ,(16)When the signals comprising are white (and possibly correlated) corrections must be made to account for the change in spectrum and response properties of the isolated cells [26], [49], [50] (See Methods).
We note that Eq. (11), which is the basis of our iterative approach, provides an approximation to the network's output which is of higher than first order in connection strength. This may seem at odds with a theory that provides a linear correction to a cell's response, cf. Eq. (4). However, Eq. (11) does not capture nonlinear corrections to the response of individual cells, as the output of each cell is determined linearly from its input. It is the input that can contain terms of any order in connection strength stemming from directed paths of different lengths through the network.
We use the theoretical framework developed above to analyze the statistical structure of the spiking activity in a network of IF neurons described by Eq. (1). We first show that the cross-correlation functions between cells in two small networks can be studied in terms of contributions from directed paths through the network. We use a similar approach to understand the structure of correlations in larger all–to–all and random networks. We show that in networks where inhibition and excitation are tuned to exactly balance, only local interactions contribute to correlations. When such balance is broken by a relative elevation of inhibition, the result may be increased synchrony in the network. The theory also allows us to obtain averages of cross-correlation functions conditioned on connectivity between pairs of cells in random networks. Such averages can provide a tractable yet accurate description of the joint statistics of spiking in these networks.
The correlation structure is determined by the response properties of cells together with synaptic dynamics and network architecture. Network interactions are described by the matrix of synaptic filters, , given in Eq. (2), while the response of cell to an input is approximated using its linear response kernel . Synaptic dynamics, architecture, and cell responses are all combined in the matrix , where describes the response of cell to an input from cell (See Eq. (1)). The correlation structure of network activity is approximated in Eq. (15) using the Fourier transforms of the interaction matrix, , and the matrix of unperturbed autocorrelations .
We first consider a pair of simple microcircuits to highlight some of the features of the theory. We start with the three cell model of feed-forward inhibition (FFI) shown in Figure 3A [51]. The interaction matrix, , has the formwhere cells are indexed in the order . To simplify notation, we omit the dependence of and other spectral quantities on .
Note that is nilpotent of degree 3 (that is, ), and the inverse of may be expressed as(17)Substituting Eq. (17) into Eq. (15) (and noting that a similar equation as Eq. (17) holds for ) yields an approximation to the matrix of cross-spectra. For instance,(18)Figure 3B shows that these approximations closely match numerically obtained cross-correlations. is the uncoupled power spectrum for cell .
Equation (18) gives insight into how the joint response of cells in this circuit is shaped by the features of the network. The three terms in Eq. (18) are directly related to the architecture of the microcircuit: Term I represents the correlating effect of the direct input to cell from cell . Term II captures the effect of the common input from cell . Finally, term III represents the interaction of the indirect input from to through with the input from to (See Figure 3C). A change in any single parameter may affect multiple terms. However, the individual contributions of all three terms are apparent.
To illustrate the impact of synaptic properties on the cross-correlation between cells and we varied the inhibitory time constant, (See Figure 3B and C). Such a change is primarily reflected in the shape of the first order term, I: Multiplication by is equivalent to convolution with the inhibitory synaptic filter, . The shape of this filter is determined by (See Eq. (2)), and a shorter time constant leads to a tighter timing dependency between the spikes of the two cells [24], [52]–[55]. In particular, Ostojic et al. made similar observations using a related approximation. In the FFI circuit, the first and second order terms, I and II, are dominant (red and dark orange, Figure 3B). The relative magnitude of the third order term, III (light orange, Figure 3B), is small. The next example shows that even in a simple recurrent circuit, terms of order higher than two may be significant.
More generally, the interaction matrices, , of recurrent networks are not nilpotent. Consider two reciprocally coupled excitatory cells, and (See Figure 4A, left). In this case,so thatEquation (15) gives the following approximation to the matrix of cross-spectra(19)In contrast to the previous example, this approximation does not terminate at finite order in interaction strength. After expanding, the cross-spectrum between cells and is approximated by(20)Directed paths beginning at and ending at (or vice-versa) are of odd length. Hence, this approximation contains only odd powers of the kernels , each corresponding to a directed path from one cell to the other. Likewise, the approximate power spectra contain only even powers of the kernels corresponding to directed paths that connect a cell to itself (See Figure 4A).
The contributions of different sub-motifs to the cross- and auto-correlations are shown in Figures 4C, D when the isolated cells are in a near-threshold excitable state (). The auto-correlations are significantly affected by network interactions. We also note that chains of length two and three (the second and third submotifs in Figure 4A) provide significant contributions. Earlier approximations do not capture such corrections [24].
The operating point of a cell is set by its parameters (, etc.) and the statistics of its input (). A change in operating point can significantly change a cell's response to an input. Using linear response theory, these changes are reflected in the response functions , and the power spectra of the isolated cells, . To highlight the role that the operating point plays in the approximation of the correlation structure given by Eq. (15), we elevated the mean and decreased the variance of background noise by increasing and decreasing in Eq. (1). With the chosen parameters the isolated cells are in a super-threshold, low noise regime and fire nearly periodically (). After the cells are coupled, this oscillatory behavior is reflected in the cross- and auto-correlations where the dominant contributions are due to first and zeroth order terms, respectively (See Figures 4F,G).
The full power of the present approach becomes evident when analyzing the activity of larger networks. We again illustrate the theory using several examples. In networks where inhibition and excitation are tuned to be precisely balanced, the theory shows that only local interactions contribute to correlations. When this balance is broken, terms corresponding to longer paths through the network shape the cross-correlation functions. One consequence is that a relative increase in inhibition can lead to elevated network synchrony. We also show how to obtain tractable and accurate approximation of the average correlation structure in random networks.
We have extended and further developed a general theoretical framework that can be used to describe the correlation structure in a network of spiking cells. The application of linear response theory allows us to find tractable approximations of cross-correlation functions in terms of the network architecture and single cell response properties. The approach was originally used to derive analytical approximations to auto- and cross-spectra in an all–to–all inhibitory network in order to study the population response of the electrosensory lateral line lobe of weakly electric fish [26]. The key approximation relies on the assumption that the activity of cells in the network can be represented by a mixed point and continuous stochastic process, as given in Eq. (9). This approximation may be viewed as a generalization of classic Linear-Poisson models of neural spiking: the crucial difference is the replacement of the stationary firing rate by a realization of an integrate-and-fire spiking process. This allows for the retention of the underlying IF spiking activity while additionally posing that neurons act as perfect linear filters of their inputs. An iterative construction then leads to the expressions for approximate cross-correlations between pairs of cells given by Eq. (15).
The linear response framework of Lindner et al. [26] was extended by Marinazzo et al. [60] to somewhat more complex networks, and compared with other studies in which networks exhibit collective oscillations. In addition, other works [13], [61], [62] used linear response techniques to study information in the collective response of cells in a network. More recently, Ostojic et al. [24] obtained formulas for cross-correlations given in Eq. (9), which correspond to the first step in the iterative construction. Their approach captures corrections due to direct coupling (first order terms) and direct common input (second order terms involving second powers of interaction kernels; see also [49], [63]). Our approach can be viewed as a generalization that also accounts for length two directed chains, along with all higher order corrections. As Figure 4 illustrates, these additional terms can be significant. The present approach also allows us to calculate corrected auto-correlations, in contrast with that of Ostojic et al.
Our work is also closely related to that of Pernice et al. [31], who analyzed the correlation structure in networks of interacting Hawkes processes [58], [59]. Both studies represent correlations between cell pairs in terms of contributions of different connectivity motifs. However, our methods also differ: while their expressions are exact for Hawkes processes, Pernice et al. did not compare their results to those obtained using physiological models, and did not account for the response properties of individual cells (though it is possible that both can be achieved approximately by using appropriate kernels for the Hawkes processes). Moreover, for simplicity Pernice et al. examined only “total” spike count covariances, which are the integrals of the cross-correlation functions. However, as they note, their approach can be extended to obtain the temporal structure of cross-correlations. Similarly, Toyoizumi et al. [64] derive expressions for cross-correlations in networks of interacting point process models in the Generalized Linear Model (GLM) class. These are very similar to Hawkes processes, but feature a static nonlinearity that shapes the spike emission rate.
To illustrate the power of the present linear response theory in analyzing the factors that shape correlations, we considered a number of simple examples for which the approximation given by Eq. (15) is tractable. We showed how the theory can be used both to gain intuition about the network and cell properties that shape correlations, and to quantify their impact. In particular, we explained how only local connections affect correlations in a precisely tuned all–to–all network, and how strengthening inhibition may synchronize spiking activity. In each case, we use comparisons with integrate-and-fire simulations to show that linear response theory makes highly accurate predictions.
It may be surprising that linear response theory can be used to provide corrections to cross-correlations of arbitrary order in network connectivity. The key to why this works lies in the accuracy of the linearization. A more accurate approximation could be obtained by including second and higher order corrections to the approximate response of a single cell, as well as corrections to the joint response. While including such terms is formally necessary to capture all contributions of a given order in network connectivity [32], [33], the success of of linear response theory suggests that they are small for the cases at hand. In short, the present approximation neglects higher-order corrections to the approximate response of individual cells, along with all corrections involving joint responses, but accounts for paths through the network of arbitrary length.
As expected from the preceding discussion, simulations suggest that, for IF neurons, our approximations become less accurate as cells receive progressively stronger inputs. The physical reasons for this loss of accuracy could be related to interactions between the “hard threshold” and incoming synaptic inputs with short timescales. Additionally, while the theory will work for short synaptic timescales, it will improve for slower synaptic dynamics, limiting towards being essentially exact in the limit of arbitrarily long synaptic time constants (note the improvement in the approximation for the FFI circuit for the slower timescale exhibited in Figure 3). Another important factor is background noise, which is known to improve the accuracy of the linear description of single cell responses. We assume the presence of a white noise background, although it is possible to extend the present methods to colored background noise [25], [65].
We found that linear response theory remains applicable in a wide range of dynamical regimes, including relatively low noise, superthreshold regimes where cells exhibit strong oscillatory behavior. Moreover, the theory can yield accurate approximations of strong correlations due to coupling: for the bidirectionally coupled excitatory circuit of Figure 4, the approximate cross-correlations match numerically obtained results even when correlation coefficients are large ( in the excitable regime, in the oscillatory regime). Additional discussion of the limits of applicability of linear response to the computation of correlations in networks can be found in the Supplementary Information. There, we show that the approximation is valid over a range of physiological values in the case of the all-to-all network, and that the theory gives accurate predictions in the presence of low firing rates (see Figures S3, S4 in Text S1).
The limits of linear response approximations of time-dependent firing activity and correlations have been tested in a number of other studies. Ostojic and Brunel [66] examined this accuracy in the relatively simple case of a neuron receiving filtered Gaussian noise in addition to a white background. Chacron et al. [61] noted that linear response approaches applied to networks of perfect integrators begin to display significant errors at larger connection strengths. Marinazzo et al. [60] remarked on the errors induced by network effects in linear response approximations to correlations in a delayed feedback loop. In particular, these errors were attributed to network effects such as synchrony in the excitatory population. The authors noted that such activity can not be correctly modeled by a linear approach.
Although we have demonstrated the theory using networks of integrate–and–fire neurons, the approach is widely applicable. The linear response kernel and power spectrum for a general integrate and fire neuron model can be easily obtained [29]. In addition, it is also possible to obtain the rate, spectrum, and susceptibility for modulation of the mean conductance in the case of conductance-based (rather than current-based) synapses (See [67] and Section 3 in Text S1). As the linear response kernel is directly related to the spike triggered average [24], [30], the proposed theoretical framework should be applicable even to actual neurons whose responses are characterized experimentally.
The possibilities for future applications are numerous. For example, one open question is how well the theory can predict correlations in the presence of adaptive currents [67]. In addition, the description of correlations in terms of architecture and response properties suggests the possibility of addressing the difficult inverse problem of inferring architectural properties from correlations [22]–[24], [64]. Ostojic et al. applied linear response methods to the latter problem. It is our hope that the present approach will prove a valuable tool in moving the computational neuroscience community towards a more complete understanding of the origin and impact of correlated activity in neuronal populations.
We quantify dependencies between the responses of cells in the network using the spike train auto- and cross-correlation functions [39]. For a pair of spike trains, , the cross-correlation function is defined asThe auto-correlation function is the cross-correlation between a spike train and itself, and is the matrix of cross-correlation functions. Denoting by the number of spikes over a time window , the spike count correlation, , over windows of length is defined as,We assume stationarity of the spiking processes (that is, the network has reached a steady state) so that does not depend on . We also use the total correlation coefficient to characterize dependencies between the processes and over arbitrarily long timescales.
The spike count covariance is related to the cross-correlation function by [7], [68]We can interpret the cross-correlation as the conditional probability that cell spikes at time given that cell spiked at time . The conditional firing rate,is the firing rate of cell conditioned on a spike in cell at units of time in the past, and
Define the Fourier transform of a function as We will often make use of the cross-spectrum between the output of cells , given by , which is the Fourier transform of the cross-correlation function of cells . The power spectrum is the cross-spectrum between a cell and itself, and is the Fourier transform of the auto-correlation function.
|
10.1371/journal.ppat.1000237 | The Chromatin Remodeler SPLAYED Regulates Specific Stress Signaling Pathways | Organisms are continuously exposed to a myriad of environmental stresses. Central to an organism's survival is the ability to mount a robust transcriptional response to the imposed stress. An emerging mechanism of transcriptional control involves dynamic changes in chromatin structure. Alterations in chromatin structure are brought about by a number of different mechanisms, including chromatin modifications, which covalently modify histone proteins; incorporation of histone variants; and chromatin remodeling, which utilizes ATP hydrolysis to alter histone-DNA contacts. While considerable insight into the mechanisms of chromatin remodeling has been gained, the biological role of chromatin remodeling complexes beyond their function as regulators of cellular differentiation and development has remained poorly understood. Here, we provide genetic, biochemical, and biological evidence for the critical role of chromatin remodeling in mediating plant defense against specific biotic stresses. We found that the Arabidopsis SWI/SNF class chromatin remodeling ATPase SPLAYED (SYD) is required for the expression of selected genes downstream of the jasmonate (JA) and ethylene (ET) signaling pathways. SYD is also directly recruited to the promoters of several of these genes. Furthermore, we show that SYD is required for resistance against the necrotrophic pathogen Botrytis cinerea but not the biotrophic pathogen Pseudomonas syringae. These findings demonstrate not only that chromatin remodeling is required for selective pathogen resistance, but also that chromatin remodelers such as SYD can regulate specific pathways within biotic stress signaling networks.
| In eukaryotes, genomic DNA is organized into a complex DNA-protein structure termed chromatin. The organization of chromatin serves to compact DNA within the nucleus and plays a central role in regulating transcription by controlling the access of DNA to transcriptional machinery. Chromatin structure can be altered through several mechanisms, one of which is chromatin remodeling, a process that disrupts DNA–protein interactions resulting in altered accessibility of specific DNA regions to regulatory proteins in the transcriptional machinery. In this study, we investigated the biological role of chromatin remodeling in defense responses to biotic stresses using the model plant Arabidopsis. We found that a chromatin remodeling protein, SPLAYED, is required for gene expression within specific biotic stress signaling networks. Consistent with this observation, loss of SPLAYED chromatin-remodeling activity resulted in increased susceptibility to a fungal pathogen, Botrytis cinerea, but not to a bacterial pathogen, Pseudomonas syringae. These results demonstrate that reduced stress tolerance in a chromatin-remodeling mutant plant can be stress specific, and is not simply due to a decrease in overall fitness as a result of non-specific global mis-regulation of gene expression.
| In eukaryotic organisms genomic DNA is packaged into chromatin, which can repress transcription by blocking the access of regulatory proteins to DNA. Dynamic changes in chromatin structure are now recognized as a robust mechanism of transcriptional control [1]–[3]. Changes in chromatin structure are brought about by a number of different mechanisms including: chromatin modifications, which covalently modify histone proteins; incorporation of histone variants; and chromatin remodeling, which utilizes ATP hydrolysis to alter histone-DNA contacts [1], [3]–[5]. ATP-dependent chromatin remodeling complexes are present in all eukaryotic organisms and can be grouped into three main classes: the SWI/SNF ATPases, the imitation switch (ISWI) ATPases, and the chromodomain and helicase-like domain (CHD) ATPases [2],[3].
Significant advances have been made in understanding the mechanism of ATP-dependent chromatin remodeling complex action [1],[4]. However, the biological role of chromatin remodeling complexes remains poorly understood, particularly in multicellular organisms where null mutations tend to be lethal [3],[6]. Studies that have investigated the biological role of chromatin remodeling complexes in multicellular organisms have largely focused on their role as regulators of cellular differentiation and development [2],[3]. In particular, Arabidopsis has served as a valuable model due to the fact that mutants in genes encoding a number of chromatin remodeling complex proteins are viable. One of the most well characterized chromatin remodeling complex proteins in Arabidopsis is the SWI/SNF class chromatin remodeling ATPase SPLAYED (SYD). Loss of SYD activity causes defects in many different developmental pathways including stem cell maintenance, patterning, developmental transitions and growth [3], [7]–[9].
The biological role of altering chromatin structure in response to stress via chromatin modifications and incorporation of histone variants has been documented [10]–[14]. However, the biological role of chromatin remodeling complexes or their specificity remains poorly understood. The role of chromatin remodeling in response to stress has been best studied in yeast where it has been shown that chromatin remodeling complexes are required for stress tolerance and are recruited to specific promoters upon stress [15]–[19]. However, few studies performed in multicellular organisms have investigated the role of chromatin remodeling in mediating stress responses. One study conducted in the human cell culture line SW480 demonstrated that chromatin remodeling complexes are recruited to specific promoters upon oxidative stress, which suggests that chromatin remodeling plays a role in the stress tolerance of multicellular organisms [20]. Additionally, it is unknown in any eukaryotic organism whether reduced stress tolerance in chromatin remodeling mutants is stress specific or indicative of decreased overall fitness due to non-specific global mis-regulation of gene expression.
In this study we examine the role chromatin remodeling plays in biotic stress responses. We found that SYD is required for expression of specific genes within biotic stress signaling networks. This requirement is likely both direct and indirect as SYD is recruited to the promoter of some, but not all, of the genes for which it is required for expression. We show that SYD is required for resistance against the necrotrophic pathogen B. cinerea but not the biotrophic pathogen P. syringae. These findings demonstrate not only that chromatin remodeling is required for selective pathogen resistance, but also that chromatin remodelers, such as SYD can regulate specific pathways within biotic stress signaling networks.
We investigated the role of chromatin remodeling in stress signaling using Arabidopsis, a multicellular organism where viable chromatin remodeling null mutants exist [9]. In a previous study we showed that the SWI/SNF class chromatin remodeling ATPase SYD transcript is upregulated rapidly following mechanical wounding [21]. We also demonstrated that mechanical wounding is a response common to numerous biotic stresses that a plant may encounter [21]. The upregulation of SYD in response to wounding suggests that SYD may be recruited to remodel promoters within stress signaling networks. To begin delineating the placement of SYD in stress signaling we first examined whether SYD is required for expression of other transcripts upregulated rapidly in response to wounding. This demonstrated that SYD is not required for the expression of the rapid wound response genes ETHYLENE RESPONSE FACTOR #18 (ERF#18) or CCR4 ASSOCIATED FACTOR 1-like (CAF1-like) (Figure S1A). We next investigated the role of SYD in the ethylene (ET), jasmonate (JA), and salicylic acid (SA) stress signaling pathways, which respond to abiotic and biotic stresses such as wounding and pathogen infection (Figure 1A) [22],[23]. As shown in Figure 1B, basal expression of the plant defensin PDF1.2a, a marker for intact ET and JA signaling, is lost in syd-2 null mutants [9]. As basal levels of PDF1.2a are generally low, but detectable, we increased cycle number to improve our ability to detect basal differences between syd-2 and wild-type (WT) [24],[25]. In contrast, basal expression of PATHOGENESIS-RELATED1 (PR1), a marker for intact SA signaling, is maintained in syd-2 plants (Figure 1B). These data suggest that SYD is required for ET and JA signaling but not SA signaling.
The loss of basal PDF1.2a but not PR1 expression in non-stressed syd-2 plants suggests that SYD impacts specific stress signaling pathways. To explore the role of SYD under inductive stress treatments we inoculated plants with the necrotrophic pathogen Botrytis cinerea and the virulent biotrophic pathogen Pseudomonas syringae. As resistance to B. cinerea requires ET and JA signaling, whereas resistance to P. syringae is predominantly mediated by SA signaling [22],[26], use of these two pathogens allows further experimental evaluation of requirements for SYD function in defense signaling. We first monitored expression of key genes in the ET/JA pathway in response to B. cinerea treatment. The expression of the transcription factor ETHYLENE RESPONSE FACTOR1 (ERF1), which requires both ET and JA for induction [22], [27]–[29], is similar in WT and syd-2 plants (Figure 2A). In contrast to ERF1, the expression of PDF1.2a requires SYD in response to B. cinerea (Figure 2A). In addition, we examined the expression of PR1 in plants treated with B. cinerea and determined that this gene is expressed at similar levels in WT and syd-2 plants (Figure 2A). We next assayed the expression of a suite of genes involved in SA biosynthesis and signaling in response to P. syringae and found that SYD is not required for their expression (Figure 2B and Figure S2). Additionally, expression of PR1, but not upstream genes (PAD4, ICS1, NPR1, and WRKY70), is enhanced in syd-2 plants. The apparent lack of detectable enhancement in PR1 expression levels by RT-PCR (Figure 1B) is likely due to signal saturation inherent to ETBr staining. However, this pattern of transcriptional alteration is similar to what is observed in myc2/jin1 mutants, suggesting that MYC2 expression may be reduced in syd (Figure 1A) [22],[30]. Furthermore, as SWI/SNF class chromatin remodeling ATPase's are primarily considered activators of transcription it is highly unlikely that SYD is acting directly to repress PR1 expression in WT plants [4]. Taken together these data demonstrate that SYD is required within specific stress signaling pathways in response to pathogen infection.
The finding that SYD is required for gene expression within specific stress signaling pathways suggests that loss of SYD function may reduce tolerance to specific biotic stresses. To test this hypothesis we first examined the resistance of syd mutants to B. cinerea. For this experiment we tested two independent syd null alleles and compared them to their respective WT background. As shown in Figure 3A and 3B, syd mutant plants are more susceptible to B. cinerea infection. The increased susceptibility of the syd mutants to B. cinerea is likely due to altered ET and/or JA signaling impacting defence mechanisms. It should also be noted that the phytoalexin camalexin plays a role in B. cinerea resistance [26]. However no significant difference in camalexin levels were detected between syd mutant and WT plants after elicitor treatment (data not shown). This suggests that SYD affects B. cinerea resistance through ET/JA signaling independent of camalexin production. To determine if reduced resistance was specific to B. cinerea we inoculated syd mutants with virulent P. syringae, for which resistance is predominantly mediated via SA signaling [26]. In contrast to B. cinerea, syd mutants and WT plants show similar resistance to P. syringae (Figure 3C and 3D). These results demonstrate that chromatin remodeling via SYD is required for stress specific disease resistance.
SYD was originally implicated in stress responses by the observation that SYD transcripts accumulate upon wounding. [21]. We therefore examined which aspects of the ET and JA signaling pathways are impacted in syd mutants following wounding. Furthermore, we wished to directly compare gene expression levels with SYD recruitment to specific promoters via chromatin immunoprecipitation (ChIP) assays. Wounding is therefore advantageous as it enables better synchronization of the stress stimulus and is a feasible treatment for the large amount of tissue required for ChIP. We first monitored the expression of ALLENE OXIDE SYNTHASE, which is involved in JA biosynthesis, and found its transcription to be similar in WT and syd-2 plants before and after wounding (Figure S1B). Measurement of JA levels reveals that basal and wound-induced JA biosynthesis is intact in syd-2 plants (Figure 4A). In agreement with B. cinerea treatment (Figure 2A), expression of ERF1 is similar in WT and syd-2 in response to wounding (Figure 4B). Furthermore, an additional ethylene response factor (ERF2), which when overexpressed results in enhanced PDF1.2a levels [31], was similar in WT and syd-2 (Figure S1B). These data collectively indicate that SYD activity is required downstream of ET and JA biosynthesis and ERF1&2 expression.
Downstream of ERF1 the expression of PDF1.2a is severely reduced, to similar levels, before and after wounding in syd-2 mutants (Figure 5A). SYD is also required downstream of JA biosynthesis for the expression of the bHLH Leu zipper transcription factor MYC2 (Figure 5B). The reduced expression of MYC2 suggests that the increased level of PR1 may indeed be due to decreased MYC2 levels in syd plants. Consistent with the reduced level of MYC2 transcripts, the expression of VEGETATIVE STORAGE PROTEIN 2 (VSP2), a gene in the JA signaling pathway which requires MYC2 for expression [22], [32]–[35], is severely reduced in syd-2 mutants (Figure 5C). Taken together these data show that while chromatin remodeling via SYD is not required for expression of ET and JA biosynthesis genes, SYD activity is required for expression of PDF1.2a, MYC2 and VSP2.
The finding that MYC2 transcript levels are reduced in syd-2 plants (Figure 5B), even though syd-2 is more susceptible to B. cinerea (Figure 3A and 3B), appears to conflict with published models of defense signaling where MYC2 acts as a negative regulator of PDF1.2a expression and resistance to necrotrophic pathogens (Figure 1A) [22],[34]. However, the apparent discrepancy can be reconciled. The slight increase in resistance against B. cinerea of myc2/jin1 mutants is thought to be due to derepression of pathogen defense genes such as PDF1.2a [34]. In syd-2 mutants derepression of PDF1.2a does not occur even though the level of MYC2 is reduced, suggesting that the requirement of SYD for PDF1.2a expression precedes repression of PDF1.2a by MYC2. It is also possible that the level of MYC2 transcript reduction in syd-2 is not great enough to have a measurable biological impact. Additionally, increases in resistance to B. cinerea exhibited by the myc2/jin1 mutants, assayed by qualitative disease symptom rating, appeared to be subtle [34], suggesting that the myc2/jin1 effect could be masked in syd mutants. To better quantify the impact of MYC2 on resistance to B. cinerea, we measured lesion development on leaves of myc2/jin1 mutant plants following infection with multiple B. cinerea isolates. We found no significant quantitative difference in lesion formation between WT and myc2/jin1 mutants (Figure 6A and Figure S3A and S3C). We also measured defense-associated secondary metabolites, including camalexin and glucosinolates, in mock- and B. cinerea- treated WT and myc2/jin1 plants. Glucosinolates are associated with Arabidopsis defense against insect herbivores and pathogens and some are regulated by JA signaling [36]. Of the five measured metabolites known to be regulated by JA, camalexin and indole-3-yl-methyl were unaffected by the myc2/jin1 mutation in comparison to WT, 3-methylsulfinyl and 4-methylsulfinyl decreased only in mock treated myc2/jin1 while 4-methoxy-indole-3-yl-methyl was present at higher concentrations in only B.cinerea treated myc2/jin1 (Figure 6B and Figure S3B, S3D and S3E and Table S1). Together these data suggest that MYC2 has neither directionally consistent nor major impacts on all molecular JA responses. It is therefore not surprising that syd mutants are more susceptible to B. cinerea even though MYC2 levels are reduced in syd-2.
The requirement of SYD for the expression of select ET and JA responsive genes suggests that SYD may be directly recruited to remodel their promoter regions, thereby enabling transcriptional induction. To test this hypothesis in vivo we performed ChIP followed by quantitative polymerase chain reaction (ChIP-qPCR) assays using the SYD specific antibody, which was previously used in ChIP experiments to show that SYD binds the WUSCHEL promoter to regulate stem cell fate [8]. Additionally, ChIP-qPCR was performed using IgG (negative control) and RNA polymerase II (POLII) (positive control for actively transcribed regions) antibodies. The background level of SYD binding to non-specific genomic loci (dashed line in Figure 7A and Figure S4) was determined by ChIP-qPCR performed on the promoters of two seed specific genes, OLEOSIN1 and AT2S3, which are subject to repressive histone H3 lysine 27 trimethylation and are not expressed in Arabidopsis leaf tissue [37]. Additionally, ChIP-qPCR was performed on syd-2 tissue wounded for 12 h to further ensure that the results are SYD specific (Figure S5). Under the experimental conditions tested, SYD does not bind the promoter of either PDF1.2a or ERF1 above the background level of detection (Figure 7A and Figure S4). To further verify the lack of SYD binding to PDF1.2a a second region of the PDF1.a promoter was assayed and SYD binding was not detected (Figure S4). As shown in Figure 7A, SYD binds the promoter of MYC2 before and after wounding. Finally, SYD is recruited to the promoter of VSP2 following wounding (Figure 7A).
Based on our findings we propose a model (Figure 7B) that summarizes the roles of SYD in response to wounding. Although SYD is required for the expression of PDF1.2a we were unable to detect SYD enrichment at the promoter of PDF1.2a, suggesting that SYD may act indirectly through an unknown factor(s) to enable transcription of PDF1.2a. Additionally, SYD is bound to the MYC2 promoter, which is consistent with the reduced expression of MYC2 in syd-2. The direct recruitment of SYD to the MYC2 promoter may also help explain the reduced transcript levels of VSP2, in non-wounded syd-2, even though SYD binding to the VSP2 promoter region was only detected following wounding. Altogether these data suggest that the altered expression of ET and JA responsive genes in syd-2 is likely a result of the loss of SYD acting both directly and/or indirectly on their promoters to regulate transcription.
Our results show that ATP-dependent chromatin remodeling is required for expression of specific genes within stress signaling networks. Additionally, this requirement is likely both direct and indirect as the chromatin remodeling ATPase SYD binds several, but not all, of the stress responsive promoters examined in vivo. Loss of chromatin remodeling activity also results in increased susceptibility to B. cinerea but not P. syringae. These results provide biological evidence that chromatin remodeling complexes, which are evolutionarily conserved within eukaryotes, are required for stress tolerance not only in yeast but also multicellular organisms. Furthermore, the requirement of ATP-dependent chromatin remodeling complexes is pathogen-specific and not a result of a general reduction in fitness.
Arabidopsis thaliana plants were grown in a 16 h light/8 h dark photoperiod at 22°C; except plants for pathogen treatments, which were grown in a 12 h light/12 h dark photoperiod. Wounding was performed as previously described [21]. All experiments were performed on 4 to 5-wk-old plants, which exhibited no disease symptoms or insect herbivory prior to treatment. Detached leaf assays were performed using the B. cinerea isolates DN, Grape, B05.10 and 83-2 [38]. Arabidopsis leaves were inoculated with 5 µl of spores at a concentration of 50,000 spores/ml [38],[39]. For P. syringae bacterial growth assays Arabidopsis leaves were inoculated with 2×104 CFU/ml P. syringae pv. tomato (Pst) DC3000 by hand injection.
Total RNA from rosette leaves was isolated by TRIzol extraction (Life Technologies, Grand Island, NY) and treated with DNAaseI to control for DNA contamination. RNA was reverse transcribed using Superscript III (Invitrogen, Carlsbad, California). PCR for RT-PCR were conducted in 25 µl reactions containing 20 ng cDNA, 1.5 mM MgCl2, 0.2 mM each dNTP, 0.05 µM each primer, and 1 U Choice-Taq Blue (Denville Scientific, Metuchen, NJ) and amplified for 29 cycles except for PDF1.2a in Figure 1B and ERF1 in Figure 2A, which were amplified for 34 cycles. Quantitative RT-PCR was conducted in 50 µl reactions containing 10 ng cDNA, 1× iQ SYBR Green supermix (Bio-Rad Laboratories, Hercules, CA), and 200 or 250 nM each primer. Amplification and data analysis were carried out as previously described [21]. The internal controls At4g34270 and At4g26410 previously described were used for transcript normalization [40]. Primers are listed in Table S2.
Extraction of JAs (MeJA and JA) were carried out as previously described [41] and further analyzed by GC-MS using a Hewlett and Packard 6890 series gas chromatograph coupled to an Agilent Technologies 5973 network mass selective detector operated in electronic ionization (EI) mode.
Camalexin and glucosinolates were measured 72 h after mock or B. cinerea inoculation as previously described [42]. Briefly, individual leaves were collected into deep 96-well plates containing 0.5 ml 90% methanol in each well. Following tissue disruption and centrifugation, 150 µl of leaf extract was removed for camalexin measurement. De-sulfo glucosinolates were extracted from an additional 150 µl of the same sample by passing the methanolic extract over a column of DEAE Sephadex A-25 (Sigma-Aldrich) and, after methanol and water washes, incubating the samples overnight with an excess of sulfatase before eluting with 150 µl H20. Extractions were performed largely as previously described, but using centrifugation rather than vacuum to remove liquid from the Sephadex columns [43]. Separation of 50 µl of aqueous extracts was performed on a 5-µm column (Lichrocart 250-4 RP18e, Hewlett-Packard, Waldbronn, Germany) attached to a Hewlett-Packard 1100 series HPLC, using the following series of solvent gradients: 6-min 1.5% to 5.0% (v/v) acetonitrile, 2-min 5% to 7% (v/v) acetonitrile, 7-min 7% to 25% (v/v) acetonitrile, 2-min gradient from 25% to 92% (v/v) acetonitrile, 6 min at 92% (v/v) acetonitrile, 1-min 92% to 1.5% (v/v) acetonitrile, and a final 5 min at 1.5% (v/v) acetonitrile. Compounds were detected at 229 nm using a diode array detector, identified by comparison with retention time and absorption spectra of purified references, and quantified using response factors as previously published (Table S1) [44],[45].
ChIP-qPCR assays were performed as previously described [8] with the following modifications. Each ChIP was conducted using 500 mg of Ler rosette leaf tissue. DNA was sonicated to a size range of 0.3–1.5 kb. For the IgG control ChIP 2 µg of IgG from rabbit serum (Sigma, St. Louis, MO) was used. Following reverse cross-linking of the immunoprecipitation reactions the samples were treated with RNase A solution (CalBiochem, La Jolla, CA) and Proteinase K (Sigma, St. Louis, MO). qPCR of the ChIP eluates was performed with iQ SYBR Green supermix according to manufacturer. ChIP-qPCR results were calculated based on the ΔΔCt method using the SuperArray ChIP-qPCR Data Analysis Template (Frederick, MD) according to the SuperArray manual, as described [46]. Briefly, ChIP DNA fractions were first normalized to input DNA (ΔCt) to account for chromatin sample preparation differences. Input normalized SYD and POLII ChIP fractions were then adjusted for the normalized non-specific background (IgG) giving the ΔΔCt value. Fold differences relative to the IgG reference were then calculated by raising 2 to the ΔΔCt power. The primers used in this study are listed in Table S2.
To determine statistical significance of treatment effects comparing WT versus syd t-tests were performed using Sigma Stat v3.5 (San Jose, CA). For comparison of WT versus myc2/jin1 factorial ANOVA performed within SAS (Cary, NC) was used to analyze the effects of genotype and treatment on measured phenotypes, with significance of differences determined via t-tests of pre-selected comparisons.
PR1: At2g14610, PDF1.2a: At5g44420, UBQ10: At4g05320, ERF1: At3g23240, PAD4: At3g52430, ICS1: At1g74710, NPR1: At1g64280, WRKY70: At3g56400, ERF#18: At1g74930, CAF1-like: At3g44260, AOS: At5g42650, ERF2: At5g47220, MYC2: At1g32640, VSP2: At5t24770, Oleo1: At4g25140, AT2S3: At4g27160
|
10.1371/journal.ppat.1006467 | Septins restrict inflammation and protect zebrafish larvae from Shigella infection | Shigella flexneri, a Gram-negative enteroinvasive pathogen, causes inflammatory destruction of the human intestinal epithelium. Infection by S. flexneri has been well-studied in vitro and is a paradigm for bacterial interactions with the host immune system. Recent work has revealed that components of the cytoskeleton have important functions in innate immunity and inflammation control. Septins, highly conserved cytoskeletal proteins, have emerged as key players in innate immunity to bacterial infection, yet septin function in vivo is poorly understood. Here, we use S. flexneri infection of zebrafish (Danio rerio) larvae to study in vivo the role of septins in inflammation and infection control. We found that depletion of Sept15 or Sept7b, zebrafish orthologs of human SEPT7, significantly increased host susceptibility to bacterial infection. Live-cell imaging of Sept15-depleted larvae revealed increasing bacterial burdens and a failure of neutrophils to control infection. Strikingly, Sept15-depleted larvae present significantly increased activity of Caspase-1 and more cell death upon S. flexneri infection. Dampening of the inflammatory response with anakinra, an antagonist of interleukin-1 receptor (IL-1R), counteracts Sept15 deficiency in vivo by protecting zebrafish from hyper-inflammation and S. flexneri infection. These findings highlight a new role for septins in host defence against bacterial infection, and suggest that septin dysfunction may be an underlying factor in cases of hyper-inflammation.
| Shigella are human-adapted Escherichia coli and cause bacillary dysentery via inflammatory destruction of the gut epithelium. In this study, we use a zebrafish (Danio rerio) model of Shigella infection to discover new roles for the cytoskeleton in inflammation and infection control. Septins, a poorly understood component of the cytoskeleton, are important in numerous biological processes including cell division and host-pathogen interactions. Here, we show that zebrafish septins can restrict inflammation and Shigella infection in vivo. In the absence of septins, larvae infected with Shigella exhibit increased mortality and bacterial burdens associated with increased Caspase-1 activity and neutrophil death. Pharmacological suppression of Il-1β signaling rescues septin-deficiency in vivo by reducing neutrophil death and preventing larval mortality. These findings reveal a new link between septins and inflammation, and highlight the cytoskeleton as a structural determinant of host defence.
| Septins, a poorly understood component of the cytoskeleton, are highly-conserved guanosine triphosphate (GTP) binding proteins organized into 4 groups based on sequence homology (the SEPT2, SEPT3, SEPT6, and SEPT7 groups). Septins from different groups assemble into hetero-oligomeric complexes which can form non-polar filaments and ring-like structures [1]. By acting as scaffolds for protein recruitment and diffusion barriers for cellular compartmentalization, septins have key roles in numerous biological processes, including cell division and host-pathogen interactions [1, 2]. Studies using human epithelial cells have revealed important roles for septins in cell-autonomous immunity, showing that septins assemble into cage-like structures to prevent the dissemination of cytosolic bacteria polymerizing actin tails [3–5]. Septin cages have also been observed in vivo using bacterial infection of zebrafish (Danio rerio) larvae [6], yet roles for septins in innate immunity in vivo remain largely unexplored.
The inflammasome is an intracellular platform that assembles in response to infection to recruit and activate Caspase-1 [7]. Caspase-1 activation enables the processing and secretion of the proinflammatory cytokine interleukin 1β (IL-1β) to control infection. How the inflammasome is triggered and assembled is the subject of intense investigation [8, 9], and the mechanisms underlying inflammation regulation are poorly understood [5]. New work has shown that components of the cytoskeleton play important roles in innate immunity and are required for inflammation control [5]. Actin and other proteins involved in actin polymerization regulate the NLRP3 (NACHT, LRR and PYD domains-containing protein 3) inflammasome by interacting with Caspase-1 and other inflammasome components [10–12]. A separate study showed that actin depolymerization, as a consequence of mutations in WD repeat-containing protein (WDR1), can trigger disease by activation of the pyrin inflammasome [13, 14]. Other components of the cytoskeleton, including microtubules and the intermediate filament protein vimentin, promote NLRP3 activity by helping to recruit ASC (apoptosis-associated speck-like protein containing a caspase-recruitment domain) and stabilize NLRP3 inflammasomes, respectively [15, 16]. The role of the septin cytoskeleton in inflammation control has not yet been tested.
Shigella, a Gram-negative enteroinvasive pathogen, causes nearly 165 million illness episodes and over 1 million deaths annually [17]. Similar to other Gram-negative pathogens in hospital patients, cases of drug-resistant Shigella strains are rising [18]. To explore the innate immune response to Shigella, several infection models have been valuable, helping to discover key roles for NOD-like receptors (NLRs) [19], neutrophil extracellular traps (NETs) [20], bacterial autophagy [21], and inflammasomes [22] in host defence. Remarkably, the major pathogenic events that lead to shigellosis in humans (i.e., macrophage cell death, invasion and multiplication within epithelial cells, cell-to-cell spread, inflammatory destruction of the host epithelium), are faithfully reproduced in a zebrafish model of S. flexneri infection [6]. Exploiting the optical accessibility of zebrafish larvae, we now have the possibility to spatio-temporally examine the biogenesis, architecture, coordination, and resolution of the innate immune response to S. flexneri in vivo.
In this study, we use a S. flexneri-zebrafish infection model to discover new roles for septins in host defence. We show that zebrafish septins restrict inflammation and are required for neutrophil-mediated immunity. To rescue septin-deficiency in vivo, we used therapeutic inhibition of Il-1β signaling and prevent neutrophil death and larval mortality. These results demonstrate a previously unknown role for septins in inflammation and infection control, and highlight the cytoskeleton as a target for suppression of inflammation.
The hindbrain ventricle (HBV) of zebrafish larvae is uniquely suited to analyze host-pathogen interactions because it is optically accessible and enables analysis of a directed leukocyte response to a compartmentalized infection (Fig 1A). We therefore characterized the HBV of zebrafish larvae as an infection site for S. flexneri M90T. Larvae aged 3 days post fertilization (dpf) were microinjected with a low (≤ 3 x 103 CFU) or high (≥ 1 x 104 CFU) dose of bacteria and their survival was assessed over time (Fig 1B). Larvae infected with a low dose of S. flexneri presented 100% survival, whereas a high dose of S. flexneri resulted in the death of ~40% of larvae within 48 hours post infection (hpi). We measured the bacterial load of infected zebrafish larvae over time by plating homogenates of viable larvae, excluding those that had already succumbed to infection. Larvae infected with a low dose of S. flexneri controlled bacterial replication, whereas larvae receiving a high dose of S. flexneri were associated with an increasing bacterial burden (Fig 1C).
To visualize the course of infection, larvae were infected with GFP-S. flexneri M90T and imaged by fluorescence microscopy. In agreement with bacterial enumerations, larvae that received a low dose inoculum showed limited proliferation of GFP-S. flexneri (Fig 1D). In contrast, larvae that received a high dose inoculum showed increasing bacterial burdens at 24 and 48 hpi (Fig 1D). Irrespective of the dose used, S. flexneri remained in the HBV and forebrain and did not cause systemic infection. Histological analyses of transverse sections of infected larvae confirmed the aggregation of S. flexneri on walls of the HBV (S1A Fig). In humans, Shigella infection and pathogenesis is strictly dependent upon the type III secretion system (T3SS) [23]. To test the role of the T3SS in our infection model, larvae were injected with T3SS-deficient (T3SS-) S. flexneri (ΔmxiD strain). The survival of zebrafish larvae infected with T3SS- S. flexneri at low or high dose was ~100% (S1B and S1C Fig), demonstrating that the T3SS contributes to Shigella virulence in vivo.
We next used the zebrafish HBV model of infection to study the control of S. flexneri by leukocytes. We outcrossed Tg(mpeg1:Gal4-FF)gl25/Tg(UAS-E1b:nfsB.mCherry)c264 (herein referred to as mpeg1:G/U:mCherry) with Tg(mpx:GFP)i114 (herein referred to as mpx:GFP) to generate double transgenic zebrafish larvae with red macrophages and green neutrophils. Larvae were infected with Crimson-S. flexneri M90T and leukocyte behavior recorded by confocal microscopy. Using a low dose of S. flexneri, we observed rapid aggregation of bacteria on walls of the HBV and by 12 hpi most bacteria had been cleared (S1D Fig, see also S1 Video). Here, macrophages were the first leukocytes to arrive (from 20 mpi) and engulf bacteria, however, as we have previously shown using caudal vein injections, S. flexneri were able to proliferate within macrophages and cause their death [6]. In contrast, neutrophils are massively recruited within hours and become the predominant leukocyte, and actively participate in the control of S. flexneri by engulfing both aggregates of extracellular bacteria and debris from macrophages unable to control infection. We thus infected Tg(lyz:dsRed)nz50 (herein referred to as lyz:dsRed) zebrafish embryos, a transgenic line in which dsRed is expressed specifically in neutrophils [24]. In the case of a low dose, neutrophils are recruited hours following infection and control S. flexneri proliferation (Fig 1E, see also S2 Video). In contrast, a high dose of S. flexneri results in uncontrolled bacterial proliferation and concomitant neutrophil cell death (Fig 1F, see also S3 Video). Collectively, these results demonstrate that infection of the zebrafish HBV is a valuable system to study in vivo the control of Shigella infection by neutrophils.
Structural analysis of a human septin complex revealed that SEPT7 is essential for septin filament assembly and function [25]. Zebrafish have orthologs for members of all 4 human septin groups (S1 Table), including Sept15 and Sept7b which share 88.7% and 92.5% identity with human SEPT7, respectively [26]. Confocal microscopy of zebrafish larvae labeled with human anti-SEPT7 antibody shows that Sept15 and/or Sept7b are present in epithelial cells, macrophages, and neutrophils (S2A Fig). To investigate the role of septins in host defence in vivo, zebrafish larvae were injected with control or Sept15 morpholino oligonucleotide and infected with S. flexneri (Fig 2A). As compared to infected control morphants, infected Sept15 morphants present significantly reduced survival and higher bacterial loads (Fig 2B and 2C). Using fluorescent microscopy we observed that in the absence of Sept15 zebrafish larvae failed to clear the infection, showing increasing fluorescence of GFP-Shigella over time (Fig 2D, see also S4 and S5 Videos). In contrast to the ~40% mortality observed for Sept15 morphants infected with wild type S. flexneri, Sept15 morphants infected with avirulent T3SS- S. flexneri present ~100% survival (S2B and S2C Fig). Together, these results show that susceptibility of Sept15 morphants to S. flexneri infection is dependent on the T3SS, and suggest that septins have an important role in the control of S. flexneri infection in vivo. To test if these results are specific to Sept15, we performed experiments using a morpholino oligonucleotide against Sept7b (S2D Fig). Similar to results obtained for Sept15 morphants, Sept7b morphants infected with S. flexneri present significantly reduced survival and higher bacterial loads as compared to infected control morphants (S2E and S2F Fig). To test if the impact of septin depletion is specific to infection of the HBV, Sept15 morphants were systemically infected with S. flexneri via the caudal vein (S2G and S2H Fig). In the case of caudal vein infection, Sept15 morphants present ~40% mortality as compared to control morphants which showed 100% survival.
Neutrophils are crucial to control Shigella infection in vivo [6]. To characterize the ability of Sept15-depleted neutrophils to clear GFP-S. flexneri, we analyzed S. flexneri-neutrophil interactions at the level of the single cell using high-resolution confocal microscopy (Fig 2E and 2F, see also S6 and S7 Videos). In both control and Sept15 morphants, neutrophils are massively recruited to the infection site where they engulf bacteria. Time-lapse movies confirmed that neutrophils from control morphants reliably clear a low dose of GFP-Shigella. In contrast, neutrophils from Sept15 morphants are unable to clear the same dose of Shigella, and are killed upon bacterial challenge.
To investigate the fate of neutrophils during infection of Sept15-depleted zebrafish, we used live cell imaging and monitored neutrophils in lyz:dsRed control or Sept15 morphants infected with GFP-S. flexneri. We quantified the total number of neutrophils at the whole animal level in control or Sept15 morphants at 3 dpf. Whereas neutrophil numbers in control morphants infected for 6 h with a low dose of S. flexneri are not significantly different from PBS-injected larvae, larvae infected for 6 h with a high dose of S. flexneri are neutropenic (Fig 3A). Sept15 morphants develop fewer neutrophils than control morphants, and when infected for 6 h with a low or high dose of S. flexneri, neutrophils were reduced even further (Fig 3B). The infection-mediated decrease in neutrophils is dependent on the T3SS, as infection with T3SS- S. flexneri has no effect on neutrophil number in either control or Sept15 morphants (S3A and S3B Fig). To test if increased mortality in Sept15 morphants is a result of neutropenia, we co-injected control or Sept15 morphants with a morpholino oligonucleotide against Irf8 (a gene involved in leukocyte differentiation [27]) to skew the myeloid cell balance towards neutrophils, and infected larvae with S. flexneri (Fig 3C and 3D). Increasing the number of neutrophils is unable to rescue Sept15 morphants from mortality or increasing bacterial burdens, suggesting that susceptibility of Sept15 morphants to S. flexneri is not because of a reduction in neutrophils per se (Fig 3E and 3F). Depletion of macrophages by Irf8 knockdown [27] may also contribute to the susceptibility of Sept15 morphants. Indeed, the ablation of macrophages by exposure of the transgenic line Tg(mpeg1:G/U:mCherry) to metronidazole showed that macrophages provide some protection against high dose S. flexneri infection (S3C and S3D Fig), likely because macrophages are implicated in the initial phagocytosis of Shigella and facilitate neutrophil scavenging crucial for host defence [6].
S. flexneri is well known to induce inflammation in vitro and in vivo [6, 28, 29]. To identify sources of inflammation in septin-depleted zebrafish larvae, we followed the spatio-temporal dynamics of interleukin 1 beta (il-1b) induction during S. flexneri infection. For this we outcrossed Tg(il-1b:GFP-F)zf550 (herein referred to as il-1b:GFP-F) a transgenic line which expresses farnesylated GFP under control of the il-1b promoter [30], with lyz:dsRed for live cell analysis by confocal microscopy (Fig 4A, see also S8 and S9 Videos). In both control and Sept15 morphants infected with Crimson-S. flexneri, we observed il-1b:GFP-F expression in neutrophils, macrophages, and epithelial cells surrounding the infection site, indicating that leukocytes and other cell types can be a source of il-1b during S. flexneri infection. To understand why Sept15 morphants succumb to S. flexneri infection, we tested Caspase-1 activity (as a readout of Il-1β processing and maturation [31]) in Shigella-infected larvae using FAM-YVAD-FMK, a fluorochrome-labeled inhibitor of Caspase-1 (FLICA) that binds specifically to active Caspase-1 enzyme. Strikingly, Caspase-1 activity is significantly increased in Sept15 morphants compared to control morphants (Fig 4B). Caspase-1-mediated signaling pathways are closely-linked to host cell death [32]. To test whether increased mortality in Sept15-depleted larvae correlates with increased host cell death, we quantified dying cells in the HBV of control and Sept15 morphants infected for 6 h with S. flexneri using acridine orange (AO), a nucleic acid-binding dye which marks dying cells (Fig 4C and 4D). In agreement with increased Caspase-1 activity, we detected a significant increase in numbers of AO-positive cells (1.7 ± 0.3 fold) in S. flexneri-infected Sept15 morphants as compared to control morphants. Together, these results suggest that hyper-inflammation is an underlying factor in the susceptibility of Sept15-deficient larvae to S. flexneri infection.
Previous work has shown that overexpression of leukotriene A4 hydrolase (lta4h) generates inflammation due to induction of tumor necrosis factor alpha (tnf-a), making zebrafish more susceptible to infection by Mycobacterium marinum [33]. To distinguish between tnf-a and il-1b inflammatory pathways in host defence against S. flexneri, we overexpressed lta4h in zebrafish larvae (S4 Fig). Overexpression of lta4h significantly increased transcript levels of tnf-a without affecting levels of il-1b (S4A and S4B Fig). However, the upregulation of tnf-a failed to increase susceptibility to a low dose of S. flexneri infection (S4C and S4D Fig), strongly suggesting that increased susceptibility of Sept15 morphants to S. flexneri infection is dependent on the activation of an il-1b signaling cascade.
Anakinra is an antagonist of IL-1 receptor (IL-1R) used to prevent inflammatory shock, sepsis, and auto-inflammatory syndromes in humans [34, 35]. Anakinra is an analogue of human IL-1RA (interleukin 1 receptor antagonist), an endogenous inhibitor of IL-1 that binds competitively to the IL-1 receptor. Although an endogenous Il-1β receptor antagonist has not been reported in fish, anakinra presents comparable homology to both human and zebrafish IL-1β (31.0% and 29.7% respectively). We tested the ability of anakinra to reduce inflammation and increase protection in our S. flexneri-zebrafish infection model. Treatment with anakinra rescued the survival of neutrophils and larvae infected with a high dose of S. flexneri (Fig 5A and 5B), without significantly affecting the enumerations of bacterial burden quantified from viable larvae (Fig 5C).
We next tested the protective effect of anakinra in Sept15 morphants. In the absence of infection, neutrophil numbers do not differ between control and anakinra-treated morphants (S5A Fig). Remarkably, upon Shigella infection, anakinra prevented neutrophil death and significantly reduced the mortality of infected Sept15 morphants (Fig 5D and 5E), without significantly affecting enumerations of bacterial burden quantified from viable larvae (Fig 5F). Collectively, these results show that reduction of inflammation by therapy can promote neutrophil and zebrafish survival during S. flexneri infection, and rescue septin-deficiency in vivo.
The zebrafish is a powerful non-mammalian vertebrate model to study the innate immune response to bacterial infection [36, 37]. We have previously used Shigella infection of the zebrafish caudal vein to study bacterial autophagy in vivo [6]. Here, using Shigella infection of the zebrafish HBV, we reveal that septins have a crucial role in restricting inflammation in vivo. Strikingly, anakinra is able to counteract septin-deficiency by preventing neutrophil death and reducing zebrafish mortality upon S. flexneri infection. These findings reveal a novel role for septins in inflammation control and host defence.
The zebrafish HBV has been used to model infection by other bacterial pathogens including Listeria monocytogenes [38], Salmonella Typhimurium [39], Pseudomonas aeruginosa [40, 41], and M. marinum [42]. In the case of L. monocytogenes, bacteria in the HBV disseminate 2–3 dpi, spreading infection to the trunk and tail muscle. Although S. flexneri is well known for invasion and inflammatory destruction of the human intestinal epithelium, and similarly to L. monocytogenes has the ability to form actin tails and spread from cell-to-cell [43], we did not observe S. flexneri dissemination outside of the zebrafish HBV or forebrain ventricle. This allowed us to analyze S. flexneri-neutrophil interactions in a compartmentalized environment, where we observed that recruited neutrophils efficiently engulf and eliminate a low dose of S. flexneri. This neutrophil behavior is in stark contrast to HBV infections of non-pathogenic E. coli, where neutrophils poorly engulf fluid-borne bacteria [44]. These observations are likely a result of S. flexneri virulence factors which promote bacterial recognition and engulfment by neutrophils.
The rabbit ileal loop model is commonly used to study the host response to Shigella infection [28]. Recently, a mouse model of shigellosis by intraperitoneal infection has been described [29]. In both animal models, S. flexneri induces the expression of proinflammatory cytokines, including IL-1β and TNF-α, as observed in humans suffering from shigellosis [45]. However, mammalian models remain poorly suited to spatio-temporally examine the innate immune response to Shigella in vivo. By contrast, the natural translucency of zebrafish larvae enables non-invasive in vivo imaging at high resolution throughout the organism. We show that il-1b:GFP-F larvae can be used to visualize the spatio-temporal dynamics of il-1b during S. flexneri infection. In-depth investigation of infection by Shigella and other bacteria that induce inflammatory signals, including L. monocytogenes and S. Typhimurium, will help to describe more precisely the coordination between septin assembly and inflammation.
When applied as a model of vertebrate development, the zebrafish has been key in linking Sept9a and Sept9b to growth defects in vivo [46]. In support of a highly conserved role for septins amongst vertebrates, the depletion of Sept15 induces cell differentiation and division defects in the pancreatic endocrine cells of zebrafish larvae [47]. More recently, the zebrafish has been used to highlight the central role of Sept15 in actin-based myofibril and cardiac function [48]. Septins are known components of the ciliary diffusion barrier in humans, and zebrafish Sept6 and Sept15 morphants present phenotypes resembling human ciliopathies, highlighting translatability of the zebrafish as a model for the study of septin biology in vivo [26, 49]. Here, we report defects in innate immunity that derive from Sept15 depletion, including inflammation and neutropenia, and show that inflammation increases the susceptibility of neutrophils to S. flexneri infection. The mechanisms underlying cell death by Shigella in epithelial cells [50] and macrophages (including apoptosis [51], necrosis [52], pyroptosis [53], and pyronecrosis [54]), have been the subject of intense investigation. The zebrafish can represent a unique experimental system to investigate Shigella-neutrophil interactions and dissect the molecular features underlying Shigella-mediated host cell death in vivo. Moreover, it is envisioned that insights into neutrophil biology arising from our S. flexneri-zebrafish model can enable novel therapeutic approaches towards diseases with an important neutrophil component.
The dysregulation of IL-1β is associated with a wide variety of inflammatory diseases [55]. Intervention into this pro-inflammatory pathway, either by blocking IL-1R or by preventing the processing / secretion of IL-1β, is critical for treatment [34]. For example, anakinra has been used to reduce IL-1β levels in a mouse model of chronic granulomatous disease (CGD), an immunodeficiency characterized by defective production of ROS [56]. Anakinra has also been effective in treatment of human patients with Schnitzler syndrome (an autoimmune disorder) or with mutations in cold-induced autoinflammatory syndrome 1 gene (CIAS1) [57]. Results obtained from our S. flexneri-zebrafish HBV infection model show that septins play a key role in the restriction of inflammation and neutrophil clearance of S. flexneri. Other studies performed in zebrafish have identified a role for the inflammasome in leukocyte clearance of L. monocytogenes and S. Typhimurium [58, 59]. What is the precise role of septins in inflammation? Septins are a unique component of the cytoskeleton that associate with cellular membranes, actin filaments, and microtubules [1]. Previous work has described a role for the actin cytoskeleton in inflammation control, by regulating the NLRP3 and pyrin inflammasomes [10–14]. We hypothesize that septins interact with components of the inflammasome and regulate assembly of this multiprotein complex. Although a precise role for septins in the assembly and activity of the inflammasome awaits investigation, these results add weight to previous studies linking inflammation and the cytoskeleton, and suggest that targeting the cytoskeleton can represent an important anti-inflammatory strategy. It is increasingly recognized that interactions between inflammation and the cytoskeleton play important roles in determining disease outcome. It will now be of great interest to further study the link between septins and inflammation, and pursue components of the cytoskeleton as novel molecular targets for inhibition of inflammation.
Animal experiments were performed according to the Animals (Scientific Procedures) Act 1986 and approved by the Home Office (Project license: PPL 70/7446).
Wild type AB were purchased from the Zebrafish International Resource Center (Eugene, OR). Tg(mpeg1:Gal4-FF)gl25/Tg(UAS-E1b:nfsB.mCherry)c264, Tg(lyz:dsRed)nz50, Tg(mpx:GFP)i114, Tg(mpeg1:YFP)w200 and Tg(il-1b:GFP-F)zf550 transgenic zebrafish lines are described previously [24, 30, 60–62]. Eggs were obtained by natural spawning and reared in Petri dishes containing 0.5x E2 water supplemented with 0.3 μg/ml methylene blue (embryo medium) [63]. For microscopy, embryo medium was supplemented with 0.003% 1-phenyl-2-thiourea (Sigma-Aldrich) from 1 dpf to prevent melanization. Both embryos and infected larvae were reared at 28.5°C. All timings in the text refer to the developmental stage at the reference temperature of 28.5°C [64]. Larvae were anesthetized with 200 μg/ml tricaine (Sigma-Aldrich) in embryo medium for injections and during in vivo imaging.
Bacterial strains used in this study were wildtype invasive S. flexneri serotype 5a M90T expressing green fluorescent protein (GFP), mCherry, or Crimson (GFP-S. flexneri, mCherry-S. flexneri, or Crimson-S. flexneri respectively) and T3SS− non-invasive variant (ΔmxiD) expressing mCherry [6]. S. flexneri were cultured overnight in trypticase soy broth, diluted 80x in fresh trypticase soy broth, and cultured until A600nm = 0.6. For injection of zebrafish larvae, bacteria were recovered by centrifugation, washed and reconstituted at the desired concentration in PBS with 0.1% phenol red. At 3 dpf, zebrafish larvae were microinjected in the HBV with up to 1 nl bacterial suspension as described previously [65]. A low dose was defined as 0.5–3.0 x 103 CFU; a high dose was defined as 1.0–2.2 x 104 CFU. Inoculums were checked a posteriori by injecting into PBS and plating onto Luria Broth (LB) agar. Larvae were maintained in individual wells of 24-well culture dishes containing embryo medium.
At indicated time points, larvae were sacrificed with tricaine, lysed in PBS with 0.4% Triton X-100 and homogenized. Serial dilutions of homogenates were plated onto LB agar supplemented with the appropriate antibiotic and CFU enumerated after 24 h incubation at 37°C; only fluorescent colonies were scored. Only viable larvae were used for CFU enumerations.
Antisense morpholino oligonucleotides were obtained from GeneTools (www.gene-tools.com). Morpholino sequence 5’-ACTCACCTTAAACAGGAAAGCAAGC-3’ was designed to target zebrafish Sept15 (ENSDARG00000102889). Morpholino sequence 5’-GAAACATCTTCACTTCGTACCTGAA-3’ was designed to target zebrafish Sept7b (ENSDARG00000052673). A standard morpholino sequence with no known target in the zebrafish genome was used as a control [6]. To increase neutrophil numbers, embryos were injected with Irf8 splice blocking morpholino as previously described [27]. Morpholinos were diluted to the desired final concentrations (0.5 mM for Sept7b and Sept15 morpholinos, 1mM for Irf8 morpholino) in 0.1% phenol red solution (Sigma-Aldrich) and 0.8 nl/embryo injected. For Leukotriene A4 hydrolase (lta4h) overexpression experiments, 1.2 nl of 200 ng/μl RNA was injected, as previously described [33]. Morpholino and RNA injections were performed on 1–8 cell stage embryos.
Whole-animal in vivo imaging was performed on anaesthetized zebrafish larvae immobilized in 1% low melting point agarose in 60 mm Petri dishes as previously described [65]. Transmission and fluorescence microscopy was done using a Leica M205FA fluorescent stereomicroscope. Imaging was performed with a 10x (NA 0.5) dry objective. Multiple-field Z-stacks were acquired every 15 min for experiments involving neutrophil recruitment to HBV infection. For high resolution confocal microscopy, infected larvae were positioned in 35 mm glass-bottom dishes and immobilized in 1% low melting agarose as described in [65]. Confocal microscopy was performed using Zeiss LSM 710, Leica SPE, or Leica SP8 microscopes and 10x, 20x, 40x oil, or 63x oil immersion objectives. For time-course acquisitions, larvae were maintained at 28.5°C. AVI/MOV files were processed and annotated using ImageJ/FIJI software.
Zebrafish larvae were infected with S. flexneri in the HBV at 3 dpf. At 6 hpi, embryos were fixed with 4% paraformaldehyde overnight at 4°C. Embryos were washed 3 times in PBS and mounted in 1% agarose. The agarose was dehydrated in a series of ethanol from 70 to 100% and then in 100% xylene and embedded in paraffin. Transversal sections of the head were stained with hematoxylin and eosin (H&E). H&E-stained tissues were imaged with an Axio Lab.A1 microscope (Carl Zeiss MicroImaging GmbH, Germany) and images acquired using an Axio Cam ERc5s colour camera. Images were processed using AxioVision (Carl Zeiss MicroImaging GmbH, Germany).
Total RNA from 5 snap-frozen larvae was extracted using RNAqueous Kit (Ambion). cDNA was obtained using QuantiTect reverse transcription kit (Qiagen). Primers for il-1b and tnf-a were previously described [66]. For each experiment, quantitative PCR was performed in technical duplicate using a Rotor-GeneQ (Qiagen) thermocycler and SYBR green reaction power mix (Applied Biosystems). To normalize cDNA amounts, we used the housekeeping gene ef1a1l1 [6] and the 2-ΔΔCT method [67].
Caspase-1 activity was determined using a FAM-FLICA Caspase-1 Assay Kit (ImmunoChemistry Technologies) as described previously [68]. Briefly, a 150x stock solution of FAM-YVAD-FMK was prepared according to the manufacturer’s guidelines. The stock was diluted to 1x in embryo medium and 5–10 larvae per experimental group bathed in the staining solution from 4.5 hpi to 6 hpi at 28.5°C. For cell death detection, larvae were bathed at 6 hpi in 2 μg/ml Acridine Orange in embryo medium for 30 min. Larvae were washed 3 times in embryo medium prior to imaging by confocal microscopy.
Wholemount immunostaining of zebrafish was performed using a standard protocol [65]. To detect septins in neutrophils and macrophages, Tg(mpx:GFP)i114 and Tg(mpeg1:YFP)w200 larvae were labelled with human anti-SEPT7 antibody (IBL), respectively.
To extract zebrafish proteins, 5–8 larvae were lysed in lysis buffer (1 M Tris, 5 M NaCl, 0.5 M EDTA, 0.01% Triton X-100) and homogenized with pestles. After centrifugation at 4°C for 15 min, supernatants were run on 8% acrylamide gels. Extracts were blotted with anti-SEPT7 (IBL) or anti-GAPDH (GeneTex) as a loading control.
Macrophages were ablated by exposure of Tg(mpeg1:Gal4-FF)gl25/Tg(UAS-E1b:nfsB.mCherry)c264 larvae to metronidazole (10 mM, Sigma-Aldrich) in embryo medium supplemented with 1% DMSO (Sigma-Aldrich). Metronidazole was administered at 2 dpf for 24 h and larvae washed 3 times prior to infection. For anakinra experiments, the embryo medium was supplemented with anakinra (10 μM, Kineret) from 1 dpf and refreshed daily until completion of the assay at 48 hpi.
Data are represented as mean ± SEM. Statistical significance was determined using Log Rank test (survival curves), unpaired two-tail Student’s t test (on log10 values of CFU counts, and log2 gene expression data), or ANOVA with Bonferroni posttest as specified in the figure legends (neutrophil counts, Caspase-1 activity) using Prism software (GraphPad Software Inc). Data were considered significant when P<0.05 (*), P<0.01 (**), or P<0.001 (***).
|
10.1371/journal.pcbi.1000722 | Investigating Homology between Proteins using Energetic Profiles | Accumulated experimental observations demonstrate that protein stability is often preserved upon conservative point mutation. In contrast, less is known about the effects of large sequence or structure changes on the stability of a particular fold. Almost completely unknown is the degree to which stability of different regions of a protein is generally preserved throughout evolution. In this work, these questions are addressed through thermodynamic analysis of a large representative sample of protein fold space based on remote, yet accepted, homology. More than 3,000 proteins were computationally analyzed using the structural-thermodynamic algorithm COREX/BEST. Estimated position-specific stability (i.e., local Gibbs free energy of folding) and its component enthalpy and entropy were quantitatively compared between all proteins in the sample according to all-vs.-all pairwise structural alignment. It was discovered that the local stabilities of homologous pairs were significantly more correlated than those of non-homologous pairs, indicating that local stability was indeed generally conserved throughout evolution. However, the position-specific enthalpy and entropy underlying stability were less correlated, suggesting that the overall regional stability of a protein was more important than the thermodynamic mechanism utilized to achieve that stability. Finally, two different types of statistically exceptional evolutionary structure-thermodynamic relationships were noted. First, many homologous proteins contained regions of similar thermodynamics despite localized structure change, suggesting a thermodynamic mechanism enabling evolutionary fold change. Second, some homologous proteins with extremely similar structures nonetheless exhibited different local stabilities, a phenomenon previously observed experimentally in this laboratory. These two observations, in conjunction with the principal conclusion that homologous proteins generally conserved local stability, may provide guidance for a future thermodynamically informed classification of protein homology.
| Protein structure and function are fundamentally determined by thermodynamics. However, for technical as well as historical reasons, current evolutionary classification schemes and bioinformatics tools do not fully utilize thermodynamic information to describe or analyze proteins. In this work, we address this deficiency by computationally estimating the position-specific thermodynamic quantities of stability (ΔG), enthalpy (ΔH), and entropy (TΔS) for a large and diverse representative sample of protein structures. The sample was drawn from an expertly curated database, such that accepted evolutionary relationships existed for all protein pairs. Importantly, trivial relationships between pairs highly similar in amino acid sequence were explicitly excluded. We found that all position-specific thermodynamic quantities ΔG, ΔH, and TΔS were more similar between proteins that were evolutionarily related (i.e., homologous), and were less similar between proteins that were not evolutionarily related (i.e., non-homologous), with stability being particularly similar between homologous proteins. However, interesting statistically significant exceptions to these trends were observed, exceptions that could indicate novel processes of functional adaptation or evolutionary fold change, mediated by thermodynamics, for the proteins involved. Taken together, these results expand our understanding of the role of thermodynamics in protein evolution and suggest an organizational framework for a future thermodynamically-informed classification of protein homology.
| Protein structure and function are ultimately determined by thermodynamics. For example, Anfinsen's seminal work [1] demonstrated that the native state of a protein exists at a minimum in Gibbs free energy of stability under physiological conditions. Binding and catalysis are also governed by free energy: the sign and magnitude of the free energy change of each functional reaction controls the reaction's direction and equilibrium extent, respectively [2],[3].
Gibbs free energy (ΔG) results from the summed, often opposing, contributions of enthalpy (ΔH) and entropy (TΔS): ΔG = ΔH−TΔS. Generally, in the case of proteins, changes in free energy are small as compared to the underlying enthalpic or entropic changes [4]. Reactions can be dominated by either enthalpy or entropy, but it is most often the case that a sometimes delicate balance between enthalpy and entropy controls protein structure and function.
Unfortunately for the goal of thermodynamic characterization of protein folds, each of these quantities can be challenging to accurately predict. While enthalpy can be rationalized in terms of information derived from atomic coordinates (i.e. from the number and types of bonds seen in the structure) [5], entropy is harder to estimate, frequently requiring knowledge not apparent from a single structure, such as information about the conformational degeneracy of the protein [6]–[8]. Equally as challenging is the task of developing a robust analysis that reports the position-specific (i.e. local) stability within the protein, rather than reporting either: 1) the energetic contribution of a residue (which would be highly sequence-dependent) or 2) the stability of a protein as a whole (i.e. global stability).
Due in part to the inherent difficulty of accurately computing global and local enthalpy, entropy, and free energy, all protein structure classification strategies of which we are aware do not incorporate thermodynamic information. It is our hypothesis that this theoretical omission limits the complete understanding of protein fold space. There may also be practical consequences to such an omission. For example, it is possible that thermodynamic information, as a protein observable independent of sequence or structure[9], could improve computational tools for sequence alignment, fold recognition [10], or homology detection, thereby clarifying discrepancies in existing classification schemes that are based on only sequence and structure.
Thermodynamic information may also yield new understanding, not available from current schemes, about evolutionary sequence, structure, and functional relationships [11]. One particularly important and as yet unanswered question is the degree to which protein stability and its components (i.e. enthalpy and entropy) are conserved during fold evolution: does the concept of “thermodynamic homology” meaningfully exist beyond conservative point mutations?
As a step towards integration of thermodynamic information into existing protein classification schemes, the local (or position-specific) free energy of stability (ΔG), enthalpy (ΔH), and entropy (TΔS) are here computed for a large representative database of protein domains using the previously described COREX/BEST algorithm [12]–[14]. Importantly, the diverse proteins studied have accepted evolutionary relationships [15] and are expertly curated [16] such that any homologs are remote (i.e. “twilight zone” [17] pairwise sequence identity or less on average). Thus, by experimental design, trivial comparisons between the thermodynamics of closely related proteins are explicitly excluded from this analysis. The central aim of this work is to assess the degree of thermodynamic conservation among remotely homologous protein domains.
Three findings relating thermodynamics to protein sequence and structure are reported. First, in accordance with previous work [18], it is confirmed that homologous proteins exhibit correlated thermodynamic information. Second, enthalpy and entropy are less correlated than stability, suggesting that homologous sequence differences result in enthalpic and entropic changes that largely balance to preserve the local stability of an evolved protein as compared to an ancestral one. Third, based on manual inspection of structural and thermodynamic alignments of homologous and non-homologous pairs of proteins, an organizational framework is postulated to guide the future integration of COREX/BEST thermodynamic information into theories of protein fold evolution.
Structural coordinates for all protein domains of length less than or equal to 150 residues were obtained from the ASTRAL 1.69 database [16] of 40% maximum sequence identity representatives. Those domains defined as SCOP [15] class “e” (membrane protein domains) were discarded, as the COREX/BEST algorithm was parameterized for globular proteins and thus was not expected to accurately estimate the thermodynamic characteristics of membrane proteins. To focus on single domains, those included in SCOP class “f” (multidomain proteins) were also discarded.
Coordinate files were preprocessed and standardized to minimize run-time errors during subsequent analysis; these minor edits included modification of selenomethionine residues to methionine, removal of multiple atom occupancies other than “A”, removal of multiple NMR models other than “1”, and modification of non-standard amino acids to alanine. In total, 3,688 domains from 666 unique SCOP families, 463 SCOP superfamilies, and 292 SCOP folds were represented within the five SCOP classes: all-α, all-β, α+β, α/β and small proteins. These statistics demonstrated a reasonable and diverse sampling of single domain protein structure space, yet included thousands of homologous protein pairs (as defined by SCOP) at less than approximately “twilight-zone” (i.e. <25%) sequence identity.
The COREX/BEST algorithm [12]–[14] constructs a protein conformational ensemble using its high-resolution structure as a template. COREX/BEST requires as input the three-dimensional structural coordinates of a protein and employs a sliding window to generate a large number of conformational microstates varying from fully folded to fully unfolded. Output is a thermodynamic (i.e. energetic) model of the protein's native state ensemble. The algorithm has been tested by both retrospective validation and blind prediction [12], [14], [19]–[23], and thus has been demonstrated to reasonably represent the ensemble. For this work, a COREX/BEST analysis was performed on each member of the preprocessed ASTRAL database described above using standard parameters: window size, 12; minimum window size, 4; temperature T, 25.0°C; and entropy weighting, W, 0.5.
The strength of COREX/BEST is the ability to capture local, also known as “position-specific”, thermodynamic quantities. The important distinguishing feature of these position-specific quantities is that they reflect the ensemble-averaged thermodynamic contributions of many residues in the three-dimensional neighborhood of one residue, rather than reflecting the independent contribution of only that particular residue [24]. Thus, local thermodynamic quantities, although reported at individual residue positions, greatly depend on the rest of the protein, in the sense that surrounding residues may influence the probability of a particular residue being folded, making it more likely, for example, for blocks of folded residues to be found together. In other words, this ensemble-based formalism separates the energetic contribution of the residue from the position itself. It is possible, and preferable, for these quantities to be obtained from experiment, for example local stability as measured by NMR-detected hydrogen exchange[25] or local enthalpy as measured by the temperature dependence of local stability [26]–[28]. Indeed, comparisons with such experiments have shown that COREX/BEST thermodynamic quantities plausibly reproduce the measured values [12],[14]. However, large scale studies such as the present one are currently difficult, if not impossible, to execute experimentally.
Computation of position-specific thermodynamic quantities from a COREX/BEST ensemble has been described in detail [12],[24],[29]. Briefly, for each partially folded microstate i of the ensemble, a Gibbs free energy of global stability ΔGi is computed from a previously validated and calibrated energy function composed of solvent-exposed surface area and conformational entropy terms [12]. From these stabilities, the probability Pi of each microstate i can be estimated by(1)
In Equation (1), Ki = exp(−ΔGi/RT) is the statistical weight of each microstate, R is the gas constant and Q is the partition function for the system. Given the probabilities of each microstate, a so-called “residue stability constant”, κf,j, can be defined for every residue j of the protein [12]:(2)
In Equation (2), the numerator is the summed probability of states in the ensemble in which a particular residue j is in a folded conformation and the denominator is the corresponding sum for states in which residue j is in an unfolded conformation. The residue stability constant directly gives the local thermodynamic stability ΔG at each residue position j, equivalent to the difference in energy between the Boltzmann-weighted subensembles of states in which residue j is folded (f) and unfolded (nf) [24],[29]:(3)
Similarly, local enthalpy (ΔH) and entropy (TΔS) were computed as a function of residue position j in each protein from the COREX/BEST ensembles as differences between the folded and unfolded subensembles for each respective thermodynamic descriptor [24]:(4)(5)
In Equations (4) – (5), subscript “ap” refers to energetic contributions arising from apolar solvent accessible surface area, “pol” refers to contributions from polar surface area, and “conf” refers to conformational entropy. The specific values of T and W are given above. Note that the total entropy of the calculation, Equation (5), reflects contributions from both solvent and conformational terms, while the enthalpy, Equation (4), reflects contributions from only solvent. Thus, this statistical thermodynamic treatment can distinguish between the two main classes of entropy. Under the native state conditions simulated in this work, the total entropy appears largely dominated by solvent contributions (Text S1, Figure S1.).
At least two different strategies could be envisioned to compare local thermodynamic quantities of two proteins: direct alignment of thermodynamic quantities or alignment of quantities according to residue equivalencies obtained from another source. Although the former strategy is under development [18],[30], for expediency we chose here to implement the latter strategy by aligning thermodynamic quantities according to structure alignment. Pairwise structure alignment was performed for the proteins in the dataset in an all-vs.-all manner using the DALI-Lite package [31] with default parameters. More than 6 million nonredundant pairwise comparisons were attempted; approximately 95% of these comparisons were successful and were retained for further analysis.
Given two sets of N equivalenced thermodynamic descriptors, a Pearson correlation coefficient r [32] was computed using the equation:(6)where, x and y represent sets, one set from each protein, of thermodynamic descriptors (ΔG, ΔH, or TΔS from Equations (3) – (5), the corresponding correlation coefficients are denoted rΔG, rΔH, rTΔS, respectively, in the text). The horizontal bar indicates an average.
A perfect positive correspondence was given by r = +1, no correspondence by r = 0, and a perfect negative correspondence by r = −1. Structural alignments of less than an arbitrary length cutoff of 20 residues were ignored, to reduce artifactual correlations due to the sensitivity of the Pearson r to outlier data points. Thermodynamic descriptors of the first or last four residues in every protein were also ignored, due to end effects in the COREX/BEST calculation caused by the minimum window size.
The Spearman rank-order correlation method [32], perhaps less widely used but more statistically rigorous than the Pearson r, was implemented as an additional test of the robustness of the results. It was observed in essentially all pairwise thermodynamic comparisons, regardless of homology, that the Spearman and Pearson r values were highly correlated (Pearson r = 0.92, Pearson p<10−6, Spearman r = 0.92, Spearman p<10−6, 9,241,311 points, data not shown), with significant individual Spearman p-values of p<0.05 occurring at Pearson r values of approximately |r|>0.25. As this threshold value of significance represented more than 45% of all 9,241,311 data points, it was decided to report the data in terms of the more widely used Pearson r. However, it is emphasized that the qualitative results and conclusions drawn were unchanged whether the Pearson or Spearman methods were used.
A relatively small, but not necessarily exhaustive, number (<50) of homologous protein comparisons involving conformational changes (data not shown) were discovered through manual inspection and discarded, since the conformational change usually dominated the thermodynamics. Although biologically interesting and deserving of future investigation, these changes were not the principal objects of the present study. Mode estimations for probability distributions of correlation coefficients and other quantities were computed using the method of Bickel and Fruewirth [33]. The results reported below were additionally filtered to only include relatively well-determined X-ray crystallographic structures (resolution of ≤2.5 Å). However, all conclusions were unchanged when NMR structures and structures with resolution >2.5 Å were also included (data not shown).
The statistical significance of individual structural and thermodynamic alignments was assessed through construction of two simple null models. In Null Model 1, the probability of chance occurrence at a particular level of structural or thermodynamic similarity was empirically estimated from the frequency of observed length-matched DALI-alignments at or above the particular similarity level. In this model, separate background distributions were used for homologs and non-homologs. In Null Model 2, the probability of chance occurrence at a particular level of structural or thermodynamic similarity was estimated from the frequency of observed length-matched gapless alignments between randomly selected pairs of non-homologous protein fragments. In this model, a minimum alpha-carbon RMSD structure superposition [34]–[36] of the fragment pair as well as the Pearson r-value between thermodynamic descriptors was computed. 30,000 pairs of fragments were chosen for each gapless alignment length L, where 10≤L≤100. In effect, the two null models occupied extremes of background distributions: Null Model 1 accounted for the interdependence of thermodynamic and structural similarity, while Null Model 2 weakened this interdependence. In both models, p-values were conservatively estimated, rounding up to the next lesser power of 10.
Figure 1 illustrates the methods used to compare position-specific thermodynamic descriptors of homologous (and non-homologous) protein pairs. A structural superposition of two homologous SH2-family domains, human Xlp protein SAP and mouse Eat2, is displayed in Figure 1A. The equivalenced residue pairings from this structure superposition were employed in Figure 1B to align the thermodynamic descriptors (e.g. local stability, ΔG) of the two proteins. A Pearson correlation of the aligned thermodynamic descriptors (Figure 1C) quantified the similarity between the two sets of descriptors. Analogous correlations were performed using the enthalpic (ΔH) and entropic (TΔS) values (data not shown). This process was repeated for all non-redundant pairwise comparisons in the structure and sequence diverse protein set, as described in Materials and Methods.
Because every protein in the set held a known position in the SCOP hierarchy, many comparisons could be sub-classified into either homologous (identical SCOP family) or likely non-homologous (different SCOP class) relationships. A clear pattern emerged when the correlations were tabulated for these two subsets: regardless of the thermodynamic descriptor used (i.e., ΔG, ΔH, TΔS), homologous proteins exhibited significantly more highly correlated descriptors than did non-homologous proteins (Figure 2). The general absence of sequence similarity between protein pairs suggested the importance of the structural context of the position (as opposed to the identity of the amino acid at that position) in determining the energetics at each position. In quantitative terms, the mode of the homologous pairs' distribution of stability correlations was 0.61, as compared to 0.29 for the non-homologous pairs (Figure 2A and Table 1). Similarly, the modes for the enthalpy correlation distributions were 0.39 and 0.06 for homologs and non-homologs, respectively (Figure 2B). Modes for the entropy distributions were 0.50 and 0.19 for homologs and non-homologs, respectively (Figure 2C).
Closer inspection of the correlation distributions suggested a second pattern: within homologous proteins, enthalpy and entropy generally did not exhibit correlations as great as those for stability (modes of 0.39, 0.50, and 0.61 respectively, Table 1; differences between these homolog distributions were all highly significant, exhibiting p<10−6 as assessed by chi-square tests with 19 d.o.f). This trend was more fully revealed by plotting individual enthalpy and entropy correlations as a function of the stability correlation for the same homologous protein pair (Figure 3A).
Examination of selected thermodynamic descriptor alignments demonstrated that the source of the differences in correlation coefficients was due to greater variation in position-specific enthalpy and entropy values as compared to the variation in stability values (Figure 3B). In particular, continuous regions of approximately 10 – 20 residues appeared to encompass much of the variation (Figure 3B, boxes). Within these variable regions, changes in enthalpy between the two proteins appeared to be somewhat balanced by changes in entropy such that the overall difference in stability was minimized (Figure 3B, boxes, discussed in detail below).
A clear “gradient” was observed relating structural similarity to thermodynamic correlation: as structural similarity and likelihood of homology decreased, thermodynamic similarity also decreased (Table 1). In other words, proteins of similar structure exhibited similar thermodynamic stability. Such an overall gradient was not surprising, given that it would be expected that in the limit of two identical structures, two identical COREX/BEST ensembles, and thus identical thermodynamics, would result.
However, the correlation distributions of Figure 2 showed a non-negligible degree of overlap between homologs and non-homologs. For example, approximately 10 percent of non-homologous pairs exhibited stability correlation coefficients larger than the homolog mode of 0.61, and the same percentage of homologous pairs even exhibited zero or negative correlation. There are at least two explanations for the significant overlap between the distribution of correlations for homologous and non-homologous proteins. The first is that the overlap is real and reflects actual differences between structural and thermodynamic representations of proteins. The second is that the cases of high correlation between non-homologs are a statistical artifact stemming from an enrichment of poorly described data in certain sequence stretches. To address this issue, we adopted a two-fold strategy designed to probe both for biases in the thermodynamics of the different positions associated with the correlations, as well as biases in the amino acid compositions in those positions.
First, in an effort to ensure that the overlap regions were not enriched with residue positions that occupied a particular region of thermodynamic parameter space, we performed principal components analysis (PCA) on the thermodynamic parameter space of the sequence segments that had the highest frequency of occurrence (top 10%) in the overlap regions and compared the eigenvalues to those obtained for the overall dataset, as well as for the datasets corresponding to the regions of no overlap[9]. The results (Text S1, Figure S2) revealed no bias in the overlap region, indicating that the high correlations were not driven by sequences enriched in a certain type of energetic environment.
To further investigate possible sampling bias as a source of the overlap in the distributions, we investigated the thermodynamic information content of those sequence segments that most frequently aligned with non-homologous proteins. Previously, propensities of amino acids in different thermodynamic environments were used as the basis for a fold recognition algorithm, demonstrating that the thermodynamic architecture outlined in this study represented a general framework within which to understand protein organization [10],[24],[29]. Among several noteworthy results from those studies was the ability to match all helical (or all beta) sequences to their folds (as described by a thermodynamic signature) using propensity information derived exclusively from all beta (or all helical) proteins[29], a result that demonstrated the universality of the thermodynamic representation of proteins as well as its independence from structural descriptors.
To ensure that frequently paired non-homologous sequences (i.e. those sequence stretches that most frequently paired with non-homologs) contained the same thermodynamic information as the overall set, we performed fold recognition experiments using thermodynamic propensities derived exclusively from those sequences. The comparable fold recognition success (Text S1, Figure S3) clearly demonstrated that the thermodynamic information content was identical across the distribution of sequences. In short, the similarity in both the range of thermodynamic parameter space occupied, as well as the distribution of amino acids within this parameter space between sequences that frequently correlate with non-homologs and those that do not, suggested that the overlap regions in the distributions shown in Figure 2 are not statistical artifacts. Instead, the results may provide insight into the relationship between structure, energy, and the evolution of this diverse library of folds. This point is discussed in more detail below.
As expected, inspection of the proteins contained in the overlap regions in Figure 2 revealed interesting exceptions to the overall structural-thermodynamic gradient, exceptions that required a more nuanced interpretation of the gradient. More generally, these exceptions suggested an organizational framework for the integration of thermodynamic information into existing fold classification schemes (as described below). The exceptions could be broadly ordered into at least three distinct classes: 1) non-homologous proteins that contained regions of coincident structural and thermodynamic similarity, 2) homologous proteins containing regions of thermodynamic similarity and structural dissimilarity, and 3) homologous proteins containing regions of structural similarity and thermodynamic dissimilarity.
To facilitate quantitative description of these exceptional cases, two empirical probability models of thermodynamic similarity were constructed to assess how often these cases might be expected due to chance, as described in Materials and Methods and displayed in Figure 4. These models could be regarded as occupying extremes in structural and thermodynamic similarity space and consequently resulted in different probability estimates. The first model (Null Model 1) accounted for the interdependence of structural and thermodynamic similarity at each alignment length. P-values for homologs and non-homologs were determined separately at each length by comparing the specific combination of structural and thermodynamic similarities with the frequency of obtaining such a combination across all comparisons. The density of points is summarized in Figure 4A for different alignment lengths. We note that the comparisons in Null Model 1 are DALI-aligned structures and thus represent comparisons between sequence stretches that have been selected for high structural similarity. To determine the probability of obtaining a particular thermodynamic correlation across any sequence comparison in the database, a second null model (Null Model 2) was adopted. According to Null Model 2, length-matched gapless alignments of randomly paired protein fragments were examined, a step taken to reduce the interdependence of structural and thermodynamic similarity. The Null Model 2 exhibited an inverse dependence of structural and thermodynamic similarity on length, in particular revealing that alignments of less than 20 residues had a substantial probability of high positive or negative thermodynamic correlation (Figure 4B). Because the background distribution of Null Model 2 covered a larger amount of structural/thermodynamic similarity space, p-values estimated from Null Model 2 were generally more significant, as compared to Null Model 1. Projections of these two-dimensional null model distributions into the single dimension of thermodynamic stability similarity, for alignments of approximately 70 residues in length, are displayed in Figure 4C.
As Figure 4C reveals, the probability density of stability correlation coefficients for random alignments of approximately 70 residue stretches (Null Model 2) is centered on zero, with approximately 80% of the comparisons falling below correlations of 0.5. As expected, the probability density functions of structurally aligned sequences for both non-homologs and homologs are shifted to higher correlations, with the shift for homologs being more dramatic. The significance of this result is discussed in more detail below. For now we simply note that these distributions can be used to identify statistically significant exceptions to homologous structural and thermodynamic similarity and to investigate the possible biological and evolutionary relevance of such examples.
Several examples of non-homologous proteins that nonetheless exhibited correlated position-specific stability are displayed in Figure 5. These examples were representative of approximately 10% of non-homologs with high thermodynamic correlation (defined as those above the homolog mode stability correlation value of 0.61, about 10% of the total non-homologs), in that they contained structurally and thermodynamically similar regions within otherwise dissimilar proteins. Some specific types repeatedly observed were β-α-β units (Figure 5A), non-local β-hairpins forming a sheet (Figure 5B), antiparallel helices (Figure 5C), and amphipathic single helices (Figure 5D).
Additional statistically significant exceptions to the structural-thermodynamic gradient, involving homologous proteins, are displayed in Figures 6 and 7. Figure 6 shows three instances of homologous pairs exhibiting conserved local stability despite secondary structure variation. This phenomenon has been previously identified as a possible thermodynamic mechanism for evolutionary fold change[9], and the examples seen here, occurring in a variety of secondary structural contexts, suggest its generality.
However, a novel hypothesis is that these regions of thermodynamically conserved structure change possibly coincide with regions of functional importance; this hypothesis is illustrated with several examples. Figure 6A shows the structure superposition and aligned stability profiles of two immunoglobulin C1-set domains. Highlighted are two boxed regions where stability is conserved despite sequence and structure variation; one region contains functional residues involved in binding of the murine cytomegalovirus m144 protein, alpha 3 domain to the β2m subunit [37]. Figure 6B highlights a strand to helix conversion between aspartate and glutamate racemases, located in a region known to mediate the different dimerization modes of the two enzymes [38],[39]. Similarly, Figure 6C highlights a region of structure change important for dimerization in each of two biotin carboxylase C-terminal domain-like proteins.
In contrast, Figure 7 shows three statistically significant examples of homologous protein pairs whose native state structures were quite similar (RMSD ≈1 Å) and yet exhibited low or modest thermodynamic stability correlations (rΔG≤0.5). One similar example of thermodynamic dissimilarity in the context of high structural similarity has recently been experimentally confirmed using point mutations of Escherichia coli adenylate kinase[40]. As suggested by the relatively small area of negative correlations between homologs in Figures 2A–C, structure similarity in the absence of thermodynamic similarity did not occur very often between homologous proteins in the database (only 8% of homologous pairs with an RMSD <2.5 Å exhibited a negative correlation coefficient).
Despite its relatively low frequency of occurrence, this class of exceptions to the structural-thermodynamic gradient also may have functional relevance, as illustrated by several examples. Displayed in Figure 7A are the superposition and aligned stability profiles of two extremely structurally similar thioredoxins from E. coli and human, with an RMSD of 1.2 Å over 122 CA atoms. However, the stability profiles are only weakly correlated (r = 0.45), largely due to stability differences in the middle half of the proteins' alignment. The region of largest difference (approximately alignment positions 60 – 80) encompasses the conserved Cys 73 residue, not found in the E. coli protein, which facilitates a unique and functionally important dimer form of human thioredoxin [41]. Figure 7B shows the comparison between two MurCD N-terminal domains from Haemophilus influenzae and the thermophile Thermotoga maritima; the low correlation between stability profiles clearly results from the greater predicted stability of the thermophile. Similarly, the stabilized N-terminal region of the zeta-class GST N-terminal domain shown in Figure 7C reduces the correlation with its delta-class homolog's stability profile. The predicted increase in stability is possibly related to the region's unique active site residues and associated novel functionality noted for the zeta-class [42].
Position-specific thermodynamic attributes of proteins, such as local stability, enthalpy, and entropy, are preserved to a large degree in remote (i.e., twilight-zone sequence identity and below) homologs. One implication of this result is that thermodynamics reinforces structure and sequence similarity, suggesting that thermodynamic attributes are likewise evolutionarily conserved properties. Upon closer inspection, however, several important features of the current analysis emerge regarding the relationship between the conservation of structure and energy. As noted above, Figure 4C reveals the shifting of the probability density function for non-homologs and homologs when comparisons are made with DALI-aligned structures, relative to random alignments. The shift observed for the non-homologs relative to the random sequence comparison is expected. In anecdotal terms, this result indicates that a particular stretch of structural elements (e.g., a helix-loop-helix) will have more similar energetics than two stretches of randomly selected structure. Perhaps surprisingly, the energetic correlation for homologs is improved over the non-homologs for a given sequence length (even though homologs with substantial sequence similarity were specifically not included in the analysis). This latter result is important because the difference between the improvement between homologs and nonhomologs provides a quantitative measure of the impact of the “structural context” of the specific sequences being compared. This is noteworthy because it undermines the notion that thermodynamic identity is defined by the RMSD of the structural units being compared. To the contrary, the results suggest a great deal of energetic heterogeneity for a particular structural motif. In other words, not all helix-loop-helix motifs of a given length and structural similarity, for example, will be thermodynamically equal. In fact, over the entire database, the results not only reveal significant instances of energetic heterogeneity for a specific structural motif, but more importantly, energetic similarity between different structures. It is our hypothesis, which we are currently testing, that it is precisely this context dependence of the energies of structural elements that determines how different folds can evolve from parental folds and why minimal sequence changes can dramatically change a protein fold[43]–[45].
Another implication of the conservation of local stability in remotely homologous proteins suggests that some aspect of protein behavior vital to the robustness of the organism is contingent on maintaining the regional stability. There are at least two possible reasons for such conservation. First, it is possible that a specific balance of regional stability within a protein may bias (or preclude) certain folding pathways, thus rendering the stability hierarchy in the protein critical to maintaining folding fidelity [46],[47]. Second, and perhaps more prevalent, is that the locally unfolded state plays an important functional role. Indeed, locally unfolded states have been shown to be functionally important in numerous native state ensembles, mediating catalysis [48],[49], allostery [50],[51], and signaling transduction [22],[52].
Intriguingly, exceptions to the trend of thermodynamic conservation exist, just as they are already known to exist for structure or sequence (i.e. homologous sequences are able to adopt unrecognizably different structures [44],[53] and homologous structures can result from unrecognizably different sequences [54],[55], respectively). As suggested by the examples given in Figures 6 and 7, these exceptions to thermodynamic conservation may be evolutionarily or functionally important, despite their low frequency of occurrence.
One interesting type of exception found here is that position-specific enthalpy and entropy are less conserved than stability. This observation suggests that in regions where this phenomenon occurs, the overall stability is more important than the thermodynamic mechanism utilized to achieve that stability. It is tempting to speculate that amino acid mutation driven changes in local entropy and enthalpy balance in conservation of local stability, as seen in Figure 3B. However, such “enthalpy-entropy compensation”, long reported in proteins as well as other chemical systems, has a controversial history [56]–[59], with the apparent compensation being due (in many cases) to errors in enthalpy and entropy that are effectively amplified when the free energy is determined from the difference between them. Thus, although it is possible that such balance is somehow a mathematical artifact [60], there is currently no evidence for such an artifact in the current analysis.
Other types of exceptions, in the form of stability differences, may arise from differences of structure, organism temperature, or functionality. It is also possible that thermodynamic similarities between putative non-homologs, now treated as “exceptions” (i.e. thermodynamic analogy), may reveal heretofore unknown evolutionary relationships.
We propose in this regard that thermodynamics can mediate mechanisms for evolutionary fold change[9]. In other words, local native state structure change between two homologous proteins is possible without major disruption of local stability and, possibly, enthalpy or entropy. Conversely, functional or temperature adaptation can be achieved by changing the thermodynamics of excited state conformational fluctuations without disrupting the ground (native) state structure[40]. These two complementary processes may be thought of as ways to affect function by “sculpting” (i.e., changing the size, shape, and energetic properties of) a protein's native state ensemble. Future experimental work will be directed at ways to intelligently employ these processes in protein design and engineering.
Finally, we note that considerable debate has emerged regarding whether protein fold space is continuous or discontinuous [61]–[63], with a major limiting factor in its resolution being the absence of a metric that can quantitatively compare different structures within a unified framework. One potential benefit of the unique thermodynamic representation of protein fold space used here is that it provides a quantitative connection between protein stability and fold specificity, in effect providing a vehicle for directly addressing this question. Indeed, this discovery of conserved position-specific thermodynamics not only furthers our understanding of the role of energetics in protein structure, function, and evolution, but also suggests an organizational framework for a possible thermodynamically-informed classification of protein homology.
|
10.1371/journal.pntd.0006760 | Transcriptomic and functional analyses of the piRNA pathway in the Chagas disease vector Rhodnius prolixus | The piRNA pathway is a surveillance system that guarantees oogenesis and adult fertility in a range of animal species. The pathway is centered on PIWI clade Argonaute proteins and the associated small non-coding RNAs termed piRNAs. In this study, we set to investigate the evolutionary conservation of the piRNA pathway in the hemimetabolous insect Rhodnius prolixus. Our transcriptome profiling reveals that core components of the pathway are expressed during previtellogenic stages of oogenesis. Rhodnius’ genome harbors four putative piwi orthologs. We show that Rp-piwi2, Rp-piwi3 and Rp-ago3, but not Rp-piwi1 transcripts are produced in the germline tissues and maternally deposited in the mature eggs. Consistent with a role in Rhodnius oogenesis, parental RNAi against the Rp-piwi2, Rp-piwi3 and Rp-ago3 results in severe egg laying and female adult fertility defects. Furthermore, we show that the reduction of the Rp-piwi2 levels by parental RNAi disrupts oogenesis by causing a dramatic loss of trophocytes, egg chamber degeneration and oogenesis arrest. Intriguingly, the putative Rp-Piwi2 protein features a polyglutamine tract at its N-terminal region, which is conserved in PIWI proteins encoded in the genome of other Triatomine species. Together with R. prolixus, these hematophagous insects are primary vectors of the Chagas disease. Thus, our data shed more light on the evolution of the piRNA pathway and provide a framework for the development of new control strategies for Chagas disease insect vectors.
| Rhodnius prolixus together with other blood-feeding bugs of the Triatominae family are primary vectors of the protozoan Trypanosoma cruzi, the causative agent of the Chagas disease. It has been estimated that 7–8 million people are affected by this life-threatening illness worldwide, which makes the Chagas disease one of the most neglected tropical diseases. In this study, we describe the transcriptome of previtellogenic stages of Rhodnius oogenesis. Furthermore, by using a combination of molecular biology techniques and functional analyses we show that central components of the piRNA pathway are conserved in this species. The piRNA pathway guarantees genomic stability in the germ cells of organisms as distant as flies and mice. In accordance, we find that the knock-down of the piwi genes, which form the backbone of the pathway, results in partial or complete female adult sterility in Rhodnius. Our data will help improve the annotation of the Rhodnius genome and provide a framework for the development of novel techniques aiming at the eradication of Rhodnius prolixus and other Triatomine species from the infested areas. The achievement of this goal will ultimately prevent the transmission of trypanosomes to humans and reduce or eliminate the diffusion of the Chagas disease.
| PIWI-clade Argonaute proteins have been implicated in a range of cellular and developmental events by regulating gene expression and imposing transposon silencing [1–3]. These proteins appear to be particularly critical for the maintenance of genomic stability during gametogenesis in a range of animal species including flies, worms and mice. The activity of the PIWIs is generally associated with the biogenesis and function of a specific class of 23-30nt small non-coding RNAs termed Piwi-interacting RNAs or piRNAs [4,5]. The details of the piRNA pathway have been mostly elucidated using the ovary of the fruit fly Drosophila melanogaster as model system. In this species, two branches of the pathway were shown to act in the germline and in the somatic follicle cells respectively. Drosophila harbors three PIWI proteins: Piwi, the founding member of this protein family, is mostly nuclear and acts in both arms of the piRNA pathway [4,6], while Aubergine (Aub) and Argonaute3 (Ago3) are expressed exclusively in the germ cells [7]. PIWI proteins belong to the Argonaute family and are characterized by typical PAZ, MID and Piwi domains. The PAZ and the MID domains interact with the mature piRNAs, which provide target specificity to the RNAse H slicing activity harbored in the Piwi domain.
In Drosophila, piRNA precursor transcripts (i.e. pre-piRNAs) are mostly transcribed from genomic regions densely populated by transposon remnants and known as piRNA clusters[4]. A second source of piRNAs is provided by the transcripts generated by active transposable elements dispersed in the genome. The RDC complex, which is formed by the Cutoff (Cuff), Rhino (Rhi) and Deadlock (Del) proteins, together with the transcription factor Moonshiner (Moon) and components of the THEO/Trex complex orchestrate that transcription of the piRNA clusters in the germline tissues and the transport of the pre-piRNAs from the nucleus to the cytoplasm [8–13]. The nuclei of Drosophila germ cells are surrounded by a membraneless organelle called the nuage, which hosts several enzymatic activities including the PIWI proteins Aub and Ago3, the DEAD-box helicase Vasa (Vas), and the tudor domain proteins Tudor, Krimper, Tejas and Papi [14]. These proteins act at different levels to process the pre-piRNAs and produce the mature 23-30nt long piRNAs. A critical role in the biogenesis of the piRNAs is exerted by Aub and Ago3, which engage in feedforward amplification mechanisms termed the ping-pong cycle [4,15]. The ping-pong cycle amplifies the piRNA population and couples piRNA biogenesis with the degradation and, consequently, with the downregulation of transposon transcripts. Recent studies revealed that the Vas protein, a well conserved germline-specific DEAD-box helicase, provides an essential scaffold to anchor the Aub protein and to promote piRNA production [16]. Finally, antisense piRNAs are bound by Piwi, that translocates into the nucleus and employs the piRNAs as guide to locate and silence active transposable elements [17]. The second branch of the pathway acts in the somatic cells of the ovary. Slicing of the precursor transcripts from somatic piRNA clusters generates antisense piRNAs, which guide Piwi to silence transposons of the zam, gypsy and idefix families. piRNA biogenesis requires the Zucchini (Zuc) endonuclease, the helicase Armitage (Armi) and the Tudor domain proteins YB, SoYB, BoYB, Vreteno and T2RD2, which accumulate in a cytoplasmic organelle known as the YB-body [18,19]. Both in the somatic as well as in the germline tissues of the fly ovary, the piRNA pathway protects the cells from the deleterious effects of massive transposon mobilization [2,20,21].
In Drosophila, mutations in PIWI proteins result in complete female adult sterility. Consistent with Aub and Ago3 being restricted to germline tissues, the absence of these factors culminates in a severe loss of stem and germ cells, a failure to assemble the chromatin in the oocyte nucleus (i.e. the karyosome phenotype) and the disruption of the dorsal-ventral polarity of the egg chamber and the future embryo [22–25]. Mutations in Piwi instead affect both the development of the follicular epithelium and of the germline. Also, this protein appears to act both in concert with or independent of the piRNAs [21].
A variable number of piwi genes originating from duplication events has been reported in various insect species. Among the Hemipteran insects, Rhodnius prolixus’ genome harbors 3 piwi genes and one ortholog of ago3, while 8 piwis and 2 copies of the ago3 gene were observed in the aphid Acyrthosiphon pisum [26]. Thus, in various animals piwi undergoes gene amplification, and the different copies often display stage- and tissue-specific expression patterns suggesting functional specialization. For instance, the mosquito Aedes aegypti harbors 7 piwi orthologs, whereby the Piwi5 and Ago3 proteins have been connected to the biogenesis of viral piRNAs [27]. piRNAs and piwi orthologs have been also identified in mosquitos of the Anopheles genus [28,29]. piwi and vas have been extensively used to investigate the segregation of germ cell determinants in different species [30]. However, functional studies addressing the role of Piwi genes in insects other than Drosophila are still scarce.
The blood-feeding insect Rhodnius prolixus is a primary vector of Trypanosoma cruzi, the etiologic agent of the Chagas disease [31]. The Chagas disease is life-threatening illness that currently affects 7–8 million people worldwide. Despite its medical relevance, the molecular events that drive oogenesis and guarantee adult fertility in Rhodnius are largely unknown. In this study, we employed transcriptome profiling to unveil the genetic and molecular basis underlying Rhodnius oogenesis. Our results reveal that central components of the piRNA pathway are conserved in this species and are expressed early during oogenesis. Furthermore, we show that Rp-piwi2, Rp-piwi3 and Rp-ago3, but not Rp-piwi1, are expressed in Rhodnius ovaries, accumulate in germline tissues and are necessary for female adult fertility.
Rhodnius females were dissected 10 days after the feeding regimen and ovaries were immediately placed in cold Phosphate Buffered Saline (PBS). Ovaries were fixed and immunostained as previously described [32]. The anti-γH2Ax (Millipore) and DAPI were diluted 1:1000 in PBS + Tween20 0.3% supplemented with 1% BSA. Ovaries were mounted in 70% glycerol and analyzed on a Leica Confocal Microscope.
The evolutionary history of PIWI proteins in Rhodnius and Drosophila was inferred applying a Maximum Likelihood method [33]. The analysis included a total of seven amino acid sequences, which were aligned by the Multiple Sequence Alignment with Log Expectation (MUSCLE, version 3.8.31) method [34], employing standard parameters. The evolutionary history was inferred by Molecular Evolutionary Genetics Analysis version 6.0 (MEGA6), and visualized using interactive Tree of Life (iTOL, v2). The tree was validated by 1000 bootstraps replications. Values higher than 90% were indicated in nodes. The amino acid alignments performed to highlight the Rp-Piwi2 PolyQ stretch in Triatominae species included the following sequences: JAI55027.1 (R. neglectus), JAP02788.1 (T. dimidiata), JAC16725.1 (T. infestans) available in NCBI, and Rp-Piwi2 of R. prolixus. The aminoacid sequences of proteins from R. prolixus and D. melanogaster were obtained from VectorBase (https://www.vectorbase.org/) and FlyBase (http://flybase.org/) respectively.
For the RT-PCR assays, total RNA was extracted from previtellogenic stages and, separately, from choriogenic stages dissected from 10/15 adult females. For qRT-PCR assays, total RNA was extracted from previtellogenic stages of wildtype, pRNAi and control ovaries 2 weeks after blood feeding. Tissues were ground in Trizol Reagent (Invitrogen) and processed as per manufacturer instructions. Total RNA was treated with Turbo DNA-free (Ambion) to remove genomic DNA traces. The resulting DNA-free total RNA was subjected to in vitro Reverse Transcription (RT) with Superscript III (Invitrogen). 1.0μg of DNA-free total RNA was used for each reaction and assays were conducted in biological triplicates. The oligonucleotides used in RT-PCR and qRT-PCR assays, are listed in the S1 Table.
For in vitro transcription, T7 promoter sequences in the appropriate orientation were added to the DNA templates through a PCR-based system using the set of oligonucleotides listed in S1 Table. The same DNA templates were adopted both for in situ hybridization assays and for dsRNA production. For parental RNAi assays, sense and antisense ssRNAs for each gene were produced using the Megascript kit (Ambion). Approximately equal amounts of sense and antisense RNAs for each target were mixed in annealing buffer, precipitated and resuspended in water to a final concentration of approximately 1.5μg/μL. Two microliters of each dsRNA were injected in the abdomen of adult females three days prior blood feeding. A total of 10 adult females were injected with each dsRNA preparation. Ovaries were dissected two weeks after the feeding regimen. The DNA templates used to generate in situ hybridization probes were obtained by PCR using oligonucleotides carrying T7 promoter sequences at the 5'-end. The templates were subjected to in vitro transcription with the DIG RNA labeling kit (Roche). In situ hybridization conditions have been described elsewhere [35]. Oligonucleotides used in this study are listed in the S1 Table.
Ovaries of 10 blood-fed Rhodnius females were dissected in cold PBS and previtellogenic stages of oogenesis were manually separated from vitellogenic stages and chorionated eggs. Total RNA from two biological replicates was isolated with Trizol reagent (Invitrogen) and subjected to paired-end RNA library preparation as per manufacturer's instructions (Illumina, Truseq paired-end RNA library prep. kit). The libraries were sequenced on Illumina HiSeq platforms at the Lactad Facility (University of Campinas, Brazil). RNAseq datasets are available at NIH SRA (SRP158580).
We employed the GEMtools pipeline (https://github.com/gemtools/gemtools) of the GEM mapper [36] to align 128,411,588 paired-end reads sequenced from two replicates (76,963,784 vs. 51,447,804 reads in each of the libraries vit1 and vit2) of the ovary samples to the Rhodnius genome assembly RproC1, using the subsequently quantified transcriptome annotation RProC1.1 as a guide. Overall ~90% (89.1 vs. 89.6%) of these reads mapped, and ~80% (80.9% vs. 79.2%) of the in total sequenced reads were considered informative for the quantification at a proportion of multi-mappings of <1%. These mappings exhibited a fidelity of on average ~3 mismatches and indels (3.2 respectively 3.5) with the RproC1 reference genome sequence and were used for subsequent quantification of the RproC1.1 transcriptome as annotated by the Vectorbase community (PMID: 22135296) and obtained from the Ensembl Metazoa database (v88). The EnsMart annotation of this database to maps 1,029 of the 1,467 most abundant mRNAs bidirectionally (i.e., orthology type "one-to-one") to protein-coding loci of the Flybase RefSeq annotation [37], further 208 Flybase proteins can be rescued through "one-to-many" and "many-to-many" ortholog mappings. After evaluating the number of occurrences for each term in comparison to their occurrence in the entire Flybase reference annotation, a p-value for the statistical overrepresentation is computed according to the model implemented in the DAVID tool [38]. Based on the distribution of p-values, we control the rate of false-discoveries to be not higher than 0.05 and group the remaining terms by a fuzzy clustering procedure on their co-occurrence in gene products, calling clusters with at least 5 members. The RproC1 version of the Rhodnius genome is available at the following URL: https://www.ebi.ac.uk/ena/data/view/GCA_000181055.2.
Adult Rhodnius females develop two ovaries, each composed of groups of seven ovarioles (Fig 1A and 1B) [39–41]. Germline stem cells are present in nymphal stages and the adult females inherit a discrete number of oocyte arrested in meiosis I and aligned at the anterior region of a lancet-like structure termed the tropharium. During oogenesis, each oocyte is surrounded by a layer of somatic follicle cells to form the mature egg chamber. Different from the meroistic polytrophic ovary found in Drosophila and other species, in the meroistic telotrophic ovary of Rhodnius the egg chambers do not harbor nurse cells. Instead, the nurse cells or trophocytes populate the tropharium, where they form a syncytium around a central region termed the trophic core (Fig 1B). The tropharium can be divided in three regions with typical cell populations (Fig 1A and 1B). Actively dividing germ cells are observed only in the Zone1 at the very anterior region [42]. These cells migrate to the Zone2, where they lose their proliferation ability, begin the endoreduplication program and turn into trophocytes (Ts), which are functionally analogous to the Drosophila nurse cells. In the Zone3, that anticipates the previtellogenic egg chambers, the Ts display large nuclei with prominent nucleoli. In this region, the cell also start to degenerate and are progressively lost and replaced by new Ts migrating from Zone2. Nutrients and possibly RNAs produced by the trophocytes accumulate in the trophic core and are subsequently transported to the growing oocytes through specialized cytoplasmic bridges termed trophic cords. The previtellogenic phase of oogenesis starts in the tropharium and ends when the egg chambers reach a diameter of 0.5mm [43]. During vitellogenesis, the egg chambers grow up to 1mm in length, the trophic cords are severed and choriogenesis begins. Mature eggs produced by Rhodnius adult females are protected by a compact and resistant chorion, which regulates the fertilization process and prevents dehydration (Fig 1A). To investigate the molecular mechanisms that coordinate and drive Rhodnius oogenesis, we performed transcriptomic profiling in ovaries of blood-fed females. Our study focused on the previtellogenic phase of oogenesis. Total RNA extracted from these tissues was used to prepare and sequence paired-end RNAseq libraries and in total 12.84 Gigabases were sequenced in two replicates (vit1 and vit2). Sequence reads were then mapped to the Rhodnius genome RproC1 (Methods) as obtained from the Ensembl Metazoa database (v83) [41]. On average ~86% of informative mappings to the genome (55,455,772 in vit1, and 36,324,980 in vit2) superimposed in the correct orientation to the RProC1.3 transcriptome annotation. These mappings provide a deep interrogation of the annotated R. prolixus transcripts, with ~84% of the 14,840 transcripts detected by >10 read mappings, ~67% by >100 mappings, ~48% by >1,000 mappings, and ~10% by 10,000 mappings. Besides six rRNA loci (RPRC015844, RPRC015846, RPRC016406, RPRC016579, RPRC016706 and RPRC016876), these highly expressed loci comprise the RNAseP (RPRC016972) and two SRP genes (RPRC017200 and RPRC017302). Interestingly, we also found a putative ortholog of the Drosophila squid gene to be highly expressed in Rhodnius ovaries. The Squid protein controls the localization and translation of the gurken mRNA, which encodes a TGFa-like morphogen involved in the axial polarization of the egg and future embryo. Of the remaining 7,172 highly expressed protein-coding (pseudo-) genes, we investigated the 1,467 genes that exhibited > 10,000 reads for common functional patterns. Since only 36 of these were annotated with known protein functions in the RproC1.3 genome version, we employed for our functional study 9,188 orthology mappings to the Drosophila melanogaster proteome. Five major groups obtained by clustering 626 functional terms are obtained from these most abundant mRNAs (Fig 1C). These have been annotated in Flybase orthologs, collected from different databases and integrated into the DAVID functional annotation platform [44,45]. Our analysis reveals that genes with high expression levels in Rhodnius ovaries are orthologs of Drosophila proteins annotated with functions related to the ribosome (group 1) and translation (group 3), to mRNA processing and splicing (group 2), to proteasome activity (group 4), and also to helicases that pave the way for transcription of genes and ATP metabolism (group 5). In agreement with our samples being depleted of vitellogenic and choriogenic egg chambers, the functional classes related to vitellogenin biogenesis and uptake as well as chorion synthesis displayed low expression levels in our datasets. Our results demonstrate that cells in the previtellogenic phase of Rhodnius oogenesis invest the major part of their energy in the biogenesis (i.e., at the level of transcription, mRNA processing and translation) and turnover of the existing proteome by elevated proteome activity.
We employed our transcriptome profiling to determine the extent of evolutionary and functional conservation of the piRNA pathway in R. prolixus. Using Blast tools, we interrogated the Vectorbase platform to identify genes with homology to the Drosophila factors involved in the biogenesis and function of the piRNAs. For each putative ortholog, we then computed the expression levels as per RNAseq, as average RPKM between two biological replicates (Fig 2). We immediately noticed that the heat shock protein 83 (hsp83) and uap56 gene, which encode a nuclear-cytoplasmic RNA export factor, are expressed at higher levels (>500 RPKM) than other piRNA pathway components in Rhodnius ovaries (Fig 2A). The majority of the putative piRNA pathway genes however could be grouped in two classes: intermediate and low expression levels (Fig 2B and 2C). The first group is composed of 14 genes, whose steady state expression levels ranged between 50 and 250 RPKM (Fig 2B). These genes encode putative orthologs of several cytoplasmic factors involved in the biogenesis of the piRNAs in Drosophila, including Vas, Tudor, Maelstrom, and two putative orthologs of the Zuc endonuclease belong to this class. The remaining 15 genes displayed average RPKM lower than 50 (Fig 2C). Among them, armitage, the dSetDB1 methyl-transferase encoding eggless gene and the putative orthologs of krimper, papi and tejas, which encode Tudor domain proteins. Surprisingly, spn-E, a critical Helicase for the production of piRNAs in Drosophila, as well as the piwi ortholog Rp-piwi1 seem to be either expressed at very modest levels or not expressed in Rhodnius ovaries (average RPKM<5) (Fig 2C). The results of the RNAseq analysis were validated by qRT-PCR assays with oligonucleotides specific to a selected group of genes (S2 Fig). In the fruit fly, the expression of the piRNA clusters in germline tissues is regulated by the RDC complex, which consists of the HP1 variant Rhino (Rhi), the Rai1-like factor Cutoff (Cuff) and the Deadlock (Del) protein. Blast alignments of the Rhi aminoacid sequence with proteins encoded in the Rhodnius genome did not return a clear match. Since this protein is a member of the Heterochromatin Protein (HP) family, several putative Rhodnius HP proteins share comparable aminoacid sequence similarity with Rhi. Conversely, the Del protein appears to evolve rapidly and is restricted to the Drosophilids. In Drosophila, the cuff and the CG9125 genes code for proteins with aminoacid sequence similarity to the yeast transcription co-factor Rai1. Interestingly, blast search analyses retrieve one single gene in Rhodnius encoding a putative protein displaying 20.7% and 28.1% aminoacid sequence identity with Cuff and with the protein encoded by CG9125 respectively. The Rhodnius Rai1-like gene, which we named Rp-rai1l, is not annotated in the current version of the Rhodnius genome and lies within the first intron of the gene RPRC008241 (supercontig KQ034693) (S1 Fig). Our transcriptomic analysis reveals that Rp-rai1l displays intermediate expression levels during Rhodnius oogenesis (Fig 2B), although its role, if any, in the piRNA pathway needs to be elucidated. For several genes that have been linked to the piRNA pathway in Drosophila, our blast search did not return homologous sequences in the Rhodnius genome. For instance, the Rhodnius genome does not appear to host homologs of the Drosophila moon and squash genes, which are expressed in the germline tissues, as well as of YB, SoYB, BoYB and T2RD2 that act in the somatic branch of the piRNA pathway in the fruit fly.
We then focused our study on the Rhodnius orthologs of the vas and piwi genes, which are central components of the piRNA pathway in Drosophila. Vas is a DEAD-box RNA helicase related to the translation factor eIF-4A and has been extensively used as a marker of germline tissues in distantly related organisms [46]. Importantly, studies in Drosophila showed that Vas elicits piRNA production in concert with Aub and Ago3 [16]. We found that the Rhodnius genome harbors a putative homolog of the vas gene (Vectorbase ID RPRC009661), which encodes a protein 75% identical to DmVas. Our RNAseq profiling reveals that Rp-vas is expressed at intermediate levels during Rhodnius oogenesis (Fig 2B). Interestingly, while the D. melanogaster vas hosts the vig and solo genes in its intronic sequences, this arrangement is absent in Rhodnius, where the Rp-vas intron 1 harbors an Open Reading Frame (ORF) encoding a transposase enzyme from a mariner-like element (Fig 2D). Transcripts of this ORF are readily detected in our ovarian transcriptome datasets. Previous studies reported that Rhodnius displays an amplification of the piwi genes, whereby three putative orthologs of piwi, namely Rp-piwi1, Rp-piwi2 and Rp-piwi3, in addition to the Rp-ago3 ortholog of the Drosophila ago3 gene are present in the genome [26]. The analysis of the normalized RNAseq reads shows that Rp-piwi2, Rp-piwi3 and Rp-ago3 are expressed at intermediate or low levels in Rhodnius oogenesis (Fig 2B–2D), while the Rp-piwi1 transcripts are barely detectable. Based on the absence of intronic sequences, it has been proposed that Rp-piwi1 is a pseudogene and might not be required for Rhodnius development. Our transcriptomic analyses seem to support this hypothesis, although we cannot rule out the Rp-piwi1 might be expressed in tissues other than the ovary. According to the current genome annotation, Rp-piwi2 contains a small intron of approximately 100bp. Our RNAseq analysis however reveals that this sequence is included in the mature transcript and the resulting ORF encodes a putative protein of 882aa. Furthermore, the 5' and 3' untranslated regions of the Rp-vas, Rp-piwi2, Rp-piwi3 and Rp-ago3 extend beyond the limits annotated in the current version of the genome. Thus, our datasets not only provide information on the steady-state expression levels for all the genes and loci expressed in previtellogenic stages of Rhodnius oogenesis, but will also contribute to improve gene annotation and discovery.
The Rp-piwi1, Rp-piwi2 and Rp-piwi3 (VectorBase IDs RPRC00252, RPRC002460 and RPR001891) encode putative proteins with 38.7%, 36.1% and 43.3% aminoacid sequence identity with D. melanogaster Piwi respectively (Fig 3A and S2 Fig). While the piwi orthologs appear to have originated from duplication of an ancestral piwi gene, the Rp-ago3 gene is homologous to the Drosophila ago3 gene (Fig 3A and S2 Fig). Accordingly, the Rp-ago3 locus (Vectorbase ID RPRC013054) encodes a putative protein 43.3% identical to DmAgo3. We then analyzed the degree of sequence identity between the individual Piwi, MID and Paz domains across the PIWI proteins in Drosophila and Rhodnius (Fig 3A and S2 Fig). All the domains appear to be well conserved in all the Rhodnius orthologs including the putative PIWI protein encoded by the Rp-piwi1 gene. Interestingly, the N-terminal region of the putative Rp-Piwi2 protein features a stretch of 18 Glutamine residues (i.e. polyQ tract) (Fig 3A and S2 Fig). This characteristic has not been reported for any of the PIWI proteins so far analyzed in a range of animal species. We therefore wondered whether PolyQ stretches are present in PIWI related proteins from other insects. Using NCBI Blast search analyses, we found that the polyQ tract is present in Piwi-like proteins encoded in the genome of Rhodnius neglectus, Triatoma infestans and Triatoma dimidiata (Fig 3B), but it is absent in PIWI proteins of Drosophila and other animals (Fig 3A and S2 Fig). Thus, the acquisition of a PolyQ sequence is likely a recent evolutionary event and is restricted to certain Triatomine species.
Next, we wondered whether the Rp-vas and the Rp-piwi genes display stage- or tissue-specific expression patterns during Rhodnius oogenesis. To answer this question, we performed in situ hybridization assays in fixed ovaries using antisense probes corresponding to specific sequences within the ORF of Rp-vas, Rp-piwi1, Rp-piwi2, Rp-piwi3 and Rp-ago3 (Fig 4). This approach revealed that the expression of Rp-vas is restricted to the germline tissues given that the Rp-vas probe generates a signal in the tropharium and in the oocyte, but not in the somatic follicle cells (Fig 4A). Thus, RPRC009661/Rp-vas encodes a bona fide ortholog of DmVas, which allows distinguish the germ cell lineage from the somatic cell population. Previous studies reported the expression of a putative vas ortholog in the somatic follicle cells of the Rhodnius ovary [47]. Our in situ hybridization protocols did not allow preserve the vitellogenic egg chambers, thus we could not determine whether RPRC009661/Rp-vas transcripts are produced in the follicular epithelium in late stages of oogenesis. In accordance with the RNAseq data, the Rp-piwi1 probe did not produce any signals above the background levels (Fig 4B). The Rp-piwi2 transcripts instead are clearly detected in the tropharium and in the developing oocytes (Fig 4C). Interestingly, the Rp-piwi2 RNAs seem to unevenly accumulate in more mature oocytes, where they are enriched in the anterior region (Fig 4C). In addition, the Rp-piwi2 transcripts are detected also in the somatic follicle cells, thus suggesting that the expression of this gene is not restricted to the germline tissues (Fig 4C'). Similar to Rp-vas, the Rp-piwi3 and Rp-ago3 transcripts are detected in the region of the tropharium hosting the polyploid Ts and in the ooplasm of newly formed and mature egg chambers, but not in the follicle cells (Fig 4D and 4E). As control assay, we generated a sense probe corresponding to a region of the Rp-ago3 ORF (Fig 4F). This probe did not produce specific signals above the background levels.
During Rhodnius oogenesis, RNAs and nutrients produced by the nurse cells hosted in the tropharium are transported to the growing oocytes through the trophic cords. In order to determine whether piwi and vas transcripts are maternally stored in the mature eggs, we performed RT-PCR assays in ovaries with oligonucleotides specific for the piwi and vas genes by separating the previtellogenic stages of oogenesis from the chorionated mature eggs. In both stages of Rhodnius oogenesis, a clear amplification product of the expected molecular weight was detected for the Rp-piwi2, Rp-piwi3, Rp-ago3 and Rp-vas genes, while no amplification signal was produced for Rp-piwi1 (Fig 4G and 4H). These observations point to a direct role for Rp-piwi2, Rp-piwi3, Rp-ago3 and Rp-vas in Rhodnius germline development and possibly early embryogenesis.
Parental RNAi (pRNAi) was previously shown to induce the efficient reduction of gene expression in Rhodnius, where genetic tools are still lacking [48]. In order to understand the function of the Rp-piwi genes in Rhodnius, we carried on pRNAi assays by injecting dsRNA molecules targeting portions of their coding regions in the abdomen of adult females (Fig 5A). Injected females were blood-fed, their eggs were collected daily over a period of three weeks and let to develop until the first-instar nymphs emerged. After the 3-weeks period, the females were dissected and the eggs retained in the abdomen were also counted. This approach allowed us to investigate the oviposition and fertility of the injected females. Eggs were divided into three bins: 1) total number of eggs, which is given by the sum of the eggs retained in the abdomen and those that were actually oviposited, 2) oviposited eggs, 3) eggs that hatched to produce first-instar nymphs. Each group of control-injected females produced on average a total of 262 eggs of which 103 were oviposited and 67 developed into first-instar nymphs. Compared to these control animals, the Rp-piwi1 pRNAi females produced on average a slightly lower number of total eggs (~86%) and oviposited eggs (~83%). However, the hatching rates were higher than the control (~111%). It is noteworthy that in our pRNAi assays, the Rp-piwi1-KD females consistently produced a slightly higher number of first-instar nymphs than the control females, although this gene is apparently not expressed in ovaries. In contrast, oogenesis seemed to be partially impaired by pRNAi-mediated downregulation of the Rp-piwi3 and Rp-ago3 genes. Compared to the control, females from these assays respectively produced ~63% and ~48% total eggs, ~40% and ~39% deposited eggs and ~33% and ~35% first-instar nymphs. The most striking result however was obtained upon injection of Rp-piwi2 dsRNA molecules in adult females. The total number of eggs and the number of oviposited eggs produced on average by these animals was ~19% and ~13% of the control values, respectively. More importantly, the first-instar nymphs generated by Rp-piwi2 pRNAi females were only ~2% of the control values.
In order to investigate the specificity of our dsRNAs for the respective cognate Rp-piwi gene, we performed qRT-PCR assays in previtellogenic stages of pRNAi ovaries with oligonucleotides specific to Rp-piwi1, Rp-piwi2, Rp-piwi3 and Rp-ago3 (Fig 5B). As internal control for this assay, we used oligonucleotides specific to Rp-rp49 (RPRC014419), a putative Rhodnius ortholog of the Drosophila rp49 gene. These assays revealed that each dsRNA specifically downregulates the expression of the cognate gene by ~30% for Rp-piwi3, ~40% for Rp-ago3 and >50% for Rp-piwi2. As expected, the injection of Rp-piwi1 dsRNA did not produce significant changes in the expression levels of Rp-piwi1. Our results strongly point to a critical role for Rp-piwi2, Rp-piwi3 and Rp-ago3 in Rhodnius oogenesis and female adult fertility.
We then sought to determine the cellular basis of the reduced fertility observed in Rp-piwi2, Rp-piwi3 and Rp-ago3 KD females. To this aim, we immunostained ovaries from pRNAi-injected adult females with DAPI to visualize the nuclei of the germ cells and of the follicle cells (Fig 5). We immediately noticed that Rp-piwi2 pRNAi tropharia displayed abundant DAPI-positive particles in Zone1 and Zone2, which are rarely observed in the tropharia of control ovaries (Fig 5C and 5D). In some cases, the anterior tip of the tropharium corresponding to Zone1 appeared severely atrophic (Fig 5E). The γH2Ax histone variant was shown to accumulate at sites of DNA damage, induced for instance by meiotic recombination events or by transposable element mobilization [23,49]. We therefore monitored the occurrence of DSBs in control and Rp-piwi2-KD ovaries with the antibodies specific to γH2Ax (Fig 5F, 5F’, 5G and 5G’). This assay revealed that the DAPI particles observed in Zone1 of Rp-piwi2-KD tropharia as well as some nurse cell nuclei are enriched in this histone variant. In contrast, control tropharia did not display any signal beyond the background levels. In Zone 3 of control tropharia, the Ts nuclei are arranged at the periphery of the tropharium and display apparent nucleoli (Fig 5H). Instead, the Zone3 of the pRNAi-treated ovaries clearly displays a lower number of Ts nuclei and abundant nuclear debris in the trophic core (Fig 5I). In Rhodnius, each egg chamber is formed by an oocyte surrounded by a follicular epithelium (Fig 5J). pRNAi for Rp-piwi2 seems to strongly impair the progression through oogenesis and the growth of the egg chambers. We frequently observed smaller and apparently collapsed egg chambers during vitellogenesis (Fig 5K). These atrophic egg chambers were still connected posteriorly to younger egg chambers emerging from the tropharium and anteriorly to more mature choriogenic egg chambers through bridges of stalk cells. Despite the clear impact of pRNAi on Rhodnius fertility, the analysis of Rp-piwi3 and Rp-ago3 pRNAi ovaries did not display obvious abnormalities by DAPI staining and anti-γH2Ax immunostainings and additional molecular tools will be necessary to dissect their function during germline development. Our results however, demonstrate that Rp-piwi2 in fundamental for Ts survival and egg chamber development during Rhodnius oogenesis.
PIWI proteins complexed with piRNAs coordinate a defense system that represses mobile genetic elements and protects the genome of animal germ cells. In this study, we show that central components of the piRNA pathway, first described in Drosophila are conserved in the hemimetabolous insect Rhodnius prolixus, which is 350mya distant from the fruit fly. Rhodnius harbors four putative piwi genes, and we show that Rp-piwi2, Rp-piwi3 and Rp-ago3, but not Rp-piwi1, are expressed in ovaries. In order to investigate their expression patterns during oogenesis, we first identified RPRC009661 as a vas ortholog in Rhodnius and we showed that it is a germline-specific gene. The Rp-piwi3 and Rp-ago3 transcripts display a germline-specific expression patterns similar to Rp-vas and appear enriched in the growing oocytes. Interestingly, Rp-piwi2 transcripts can be detected both in the somatic as well as in the germ cells and seem to display an asymmetric localization pattern during oocyte development. This gene is expressed in the tropharium and its transcripts evenly accumulate in the budding egg chamber. In the neighboring and more mature egg chamber, however, Rp-piwi2 transcripts are enriched at the anterior pole of the oocytes. Our in situ hybridization assays suggest that Rp-piwi2 transcripts might diffuse from the ooplasm of the budding egg chamber into the neighboring more mature oocyte. Alternatively, Rp-piwi2 expression might occur in the invading follicle cells that form the boundary between the budding egg chambers and the transcript deposited in the adjacent oocytes. It will be of great interest to determine whether the Rp-piwi2 expression pattern impacts the axial polarization of the Rhodnius eggs and embryos. Our functional studies using pRNAi against the Rp-piwi2 gene resulted in oogenesis arrest and complete female adult sterility. In wild type ovaries, the Zone 1 of the tropharium hosts mitotically active trophocytes, which replenish the population of polyploid Ts in Zone 2 and 3. Reduction of the Rp-piwi2 levels by pRNAi causes a severe loss of dividing cells in Zone 1 and of polyploid Ts in Zone 2 and 3. The accumulation of γH2Ax-positive nuclear debris in the tropharia of injected females strongly suggests that Ts degenerate in Rp-piwi2 KD ovaries. The loss of Ts in turn likely results in dumping phenotypes, which explain the oogenesis arrest and the frequently collapsed egg chambers observed in these females. It is tempting to speculate that the DNA damage and the loss of Ts observed in the Rp-piwi2 KD ovaries might be caused by the deregulation of transposable elements. The percentage of transposable elements in the Rhodnius genome is approximately 6% and two thirds of the transposons in this species belong to the mariner family [41,50,51]. Rp-piwi2 might be required to silence these elements in the germline and, possibly, in the somatic tissues. The cloning and characterization of the piRNA population will be necessary to shed more light on the function of Rp-piwi2 and the piRNA pathway in this species. Remarkably, we found that the putative Rp-Piwi2 protein features a 18aa Polyglutamine (PolyQ) tract at its N-terminal region. PolyQ repeats have been identified in various proteins of organisms as distant as plants and vertebrates and are often found in transcription factors. Interestingly, the PolyQ stretch appears to be conserved in putative PIWI proteins of the closely related species Rhodnius neglectus, Triatoma infestans and Triatoma dimidiata, while it is not present in the PIWI proteins of other organisms including Drosophila. Thus, the PolyQ tract is likely a recent acquisition in the evolution of the PIWI proteins and, based on the available sequenced genomes, appears to be restricted to blood-feeding insects of the Triatomine family.
Albeit to a lesser extent, Rp-piwi3 and Rp-ago3 KDs also affect egg production and female adult fertility in Rhodnius. The expression of Rp-piwi1 gene instead is negligible in Rhodnius ovaries and Rp-piwi1 dsRNA injection in adult females does not negatively affect oogenesis and fertility.
In addition to the Rp-piwi genes our transcriptomic analysis revealed that several components of the piRNA pathway are conserved and expressed in the ovary of this species. We did not find evidence of an RDC complex in Rhodnius except for a putative protein (i.e. Rp-Rai1l) with similarity to Cuff. If piRNA clusters exist in this species, it is likely that their regulation relies on a set of proteins different from the one described in Drosophila. However, we provide evidence that several factors involved in the transport and processing of the pre-piRNAs, including Uap56, Krimper and Maelstrom among others, are expressed during Rhodnius oogenesis. Yet, some critical germline factors, like the Helicase SpnE, are expressed at very low levels. Similarly, the somatic branch of the piRNA pathway might rely on the activity of the Rp-piwi2 gene and the zuc, armi and vret orthologs, while YB, BoYB, SoYB and T2RD2, which associate with the YB bodies and catalyze the production of mature piRNAs in the Drosophila follicle cells, are not present in the Rhodnius genome. These genes have been reported to be absent also from the genome of other insect species, including the Honeybee Apis mellifera and Tribolium castaneum [52]. Thus, both branches of the piRNA pathway are partially conserved in insects and it will be a challenge for the future to fully understand the differences between Drosophila and Rhodnius.
Rhodnius prolixus together with other Triatomine species are major vectors of the protozoan Trypanosoma cruzi, the causal agent of the Chagas disease. In this study, we shed light on the ovarian transcriptome of Rhodnius and unveiled the degree of evolutionary and functional conservation of the piRNA pathway in this species. Furthermore, we show that piwi genes are essential for oogenesis and adult fertility in Rhodnius and likely exert similar functions in other Triatomine species. Sterile Insect Techniques (SIT) have been extensively used to reduce natural populations of insects of medical or economic importance in many countries [53].Thus, our results provide a framework for the development of novel strategies to control the natural populations of Triatomine insect vectors and reduce the spread of the Chagas disease.
|
10.1371/journal.pcbi.1004355 | Reconciling Estimates of Cell Proliferation from Stable Isotope Labeling Experiments | Stable isotope labeling is the state of the art technique for in vivo quantification of lymphocyte kinetics in humans. It has been central to a number of seminal studies, particularly in the context of HIV-1 and leukemia. However, there is a significant discrepancy between lymphocyte proliferation rates estimated in different studies. Notably, deuterated 2H2-glucose (D2-glucose) labeling studies consistently yield higher estimates of proliferation than deuterated water (D2O) labeling studies. This hampers our understanding of immune function and undermines our confidence in this important technique. Whether these differences are caused by fundamental biochemical differences between the two compounds and/or by methodological differences in the studies is unknown. D2-glucose and D2O labeling experiments have never been performed by the same group under the same experimental conditions; consequently a direct comparison of these two techniques has not been possible. We sought to address this problem. We performed both in vitro and murine in vivo labeling experiments using identical protocols with both D2-glucose and D2O. This showed that intrinsic differences between the two compounds do not cause differences in the proliferation rate estimates, but that estimates made using D2-glucose in vivo were susceptible to difficulties in normalization due to highly variable blood glucose enrichment. Analysis of three published human studies made using D2-glucose and D2O confirmed this problem, particularly in the case of short term D2-glucose labeling. Correcting for these inaccuracies in normalization decreased proliferation rate estimates made using D2-glucose and slightly increased estimates made using D2O; thus bringing the estimates from the two methods significantly closer and highlighting the importance of reliable normalization when using this technique.
| Stable isotope labeling is used to quantify the rate at which living cells proliferate and die in humans. It has been central to a number of seminal studies, particularly in viral infections such as HIV-1, and leukemia. However, different labels (deuterated water or deuterated glucose) yield different estimates for the rate of cell proliferation and loss; this hampers our understanding and weakens our confidence in this important technique. We performed in vitro and in vivo experiments as well as a new analysis of existing data to directly compare the two labels. This reveals that a major source of the discrepancy lies in the difficulty of assessing deuterated glucose availability. We reconcile published studies and provide recommendations to avoid this problem in the future.
| Quantification of lymphocyte kinetics is vital for our understanding of immune cell dynamics in health and disease. The development [1,2] of stable isotope labeling techniques, using either deuterium labeled glucose (D2-glucose) or deuterium labeled water (D2O), has enabled the safe quantification of lymphocyte turnover in humans in vivo. This has provided unprecedented insight into immunity in healthy individuals, as well as in various conditions, including aging, viral infection, diabetes and leukemia [3–14].
Despite this breakthrough, estimates of lymphocyte kinetics (i.e. proliferation and loss) are known to differ up to 10-fold between stable isotope labeling studies, with D2-glucose labeling consistently yielding higher proliferation and loss rates than D2O labeling [6,8,15–17]. We have previously shown that short labeling periods can yield higher estimates of proliferation and loss than long labeling periods [18–20]. In the case of proliferation rate estimates this is because, with long labeling periods, the label in rapidly turning over cell subpopulations can saturate leading to an underestimate of the proliferation rate [18,19]. In the case of loss, this is because the loss rate estimated pertains only to the labeled cell population and the composition of the labeled cell population will change as the duration of label administration changes [20], with a stronger bias towards rapidly dividing cells when labelling periods are shorter. Since D2-glucose labeling protocols usually involve a much shorter labeling period than D2O labeling protocols, the relationship between the duration of label administration and the estimate of proliferation and loss explains some of the difference between kinetics obtained using the two techniques. However, even after correcting for the length of the labeling period, our analysis here shows that significant discrepancies remain. This raises the concern that there may be fundamental differences between D2O labeling and D2-glucose labeling that affect the estimates of lymphocyte kinetics they produce.
D2-glucose and D2O labeling experiments differ in several aspects, both in the exact experimental procedures as well as in the way the experimental data are translated into the biological parameters of interest. D2-glucose is typically administered for short periods of time (hours or days). The pool of blood glucose is small and has a high rate of turnover, so rapid up- and de-labeling can be achieved [15,16]. Although any part of the glucose molecule could be labeled, most studies have used 6,6-D2-glucose. During de novo purine and pyrimidine synthesis, the two non-exchangeable deuteriums on the 6-position are carried over into the 5-position carbon of the pentose moiety of DNA precursors (as C1 is lost and C6 is redesignated C5) [2]. Deuterium enrichment is measured (by mass spectrometry) in the pentose moiety of the DNA of the cell population of interest. When D2-glucose labeling experiments are conducted in vitro, enrichment levels in DNA never reach media enrichments but plateau at about 60–75% [2]. This has been attributed to intracellular dilution of labeled nucleotide triphosphates (NTP) by pre-formed NTP, by other pentose precursors, and by unlabeled salvage pathway synthesis. Since the short labeling periods associated with D2-glucose preclude the use of a completely replaced population to derive product-based estimates of in vivo precursor enrichments, these plateau tissue culture values have been used to correct for (an assumed similar level of) intracellular dilution in vivo, by applying a scaling factor in the range 0.6–0.75 [19]. Within this manuscript we refer to this intracellular dilution factor as bg. In addition to adjusting for intracellular dilution, it is necessary to correct for the label availability, which is estimated by measuring the label enrichment in glucose in blood plasma at multiple time points.
By contrast D2O is usually administered for several weeks. Intermediary metabolism introduces the deuterium in place of hydrogen in de novo synthesized deoxyribose. To determine the level of incorporation, mass spectrometric analysis is performed on the deoxyribose moiety of purine nucleotides [1]. The observed deuterium incorporation in the DNA from the cell population(s) of interest is normalized to the maximal level of deuterium incorporation that deoxyribose can attain, which is typically determined in the same individual in a population with rapid turnover, such as granulocytes, monocytes or thymocytes [8]. This maximum enrichment attainable is determined by a scaling factor and the level of D2O in the body. The scaling factor, which is analogous to the intracellular dilution factor bg for D2-glucose, has been variously referred to as the amplification factor or c [8]. Here, to emphasize analogy to bg, we refer to it as bw. Deoxyribose contains seven non-exchangable hydrogen atoms, any of which might potentially be replaced by deuterium [1,21]. Consequently, the maximum level of enrichment seen in deoxyribose exceeds that seen in plasma and the scaling factor, bw, is greater than 1; typically bw is in the range 3.5–5.2 [8]. Body water turns over relatively slowly, so enrichment in the body fluids reaches its maximum and is washed-out from the body more slowly than D2-glucose. Consequently, there is still considerable de novo DNA labeling long after the label has been withdrawn. A correction for the level of D2O present in the body fluids is made by taking the enrichment of blood plasma or urine samples into account [8].
In addition to the methodological differences between D2O and D2-glucose labeling protocols there are differences in the way the compounds are synthesized into deoxyribose [1,19], in the distribution of water and glucose throughout the body, in the transport of water and glucose into cells, and in the feedback mechanisms that control water and glucose levels. Potentially, some or all of these differences could cause the compounds to give different estimates of cell kinetics. To date a direct comparison of the two techniques has not been performed. We sought to address this problem.
We performed stable isotope labeling experiments in vitro and in mice using D2-glucose and D2O, while keeping every other aspect of the study identical. This showed that, for T cells in vitro and murine splenocytes and PBMC in vivo, biochemical differences between the compounds do not lead to differences in parameter estimates. Instead, we found that proliferation rate estimates made using D2-glucose in mice were susceptible to difficulty in estimating rapidly fluctuating blood glucose enrichment levels. We therefore hypothesized that the measurement of deuterium enrichment in blood glucose may be an unreliable estimate of the precursor enrichment. This hypothesis was supported by a new analysis of two published D2-glucose studies [6,13]. Analogous problems, albeit of a much smaller magnitude, were also found in a published D2O study [8]. We show that adjusting for these presumed inaccuracies in normalization decreases the proliferation rates estimated in the D2-glucose studies whilst simultaneously increasing the proliferation rates from the D2O labeling study, thus bringing these estimates significantly closer.
Long labeling periods can lead to an underestimate of cell proliferation rates due to saturation of label [18]. We therefore investigated whether adjusting for the length of the labeling period resolves the discrepancy between estimates of lymphocyte proliferation obtained using D2-glucose and D2O labeling. We focused on three published studies in healthy humans where detailed data were available: a 9 week D2O labeling study [8], a seven-day D2-glucose labeling study [13] and a one-day D2-glucose labeling study [6] and compared the labeling of total CD4+ T lymphocytes (Methods). We analyzed the data from the seven-day D2-glucose and the 9 week D2O labeling study using a multi-exponential model [22]. The multi-exponential model (Methods) consists of N homogeneous subpopulations; proliferation and death within each subpopulation is random and occurs at a constant rate and each subpopulation is assumed to be independently at equilibrium (i.e. proliferation = death). The multi-exponential model allows for saturation of label in rapidly turning over subpopulations and effectively adjusts for the length of the labeling period. There were fewer data points available for the one-day D2-glucose study and so the multi-exponential model could not be used (since it has a relatively large number of free parameters). However, saturation is unlikely to be an issue with such a short labeling period so we used the alternative, kinetic heterogeneity model [6,20]. The kinetic heterogeneity model is a model that deals with heterogeneity implicitly by postulating that labeled cells may not be representative of the whole population and thus the disappearance rate of labeled cells (d*) may not be equal to the proliferation rate of the whole population (p) even for populations at equilibrium (Methods). The advantage of the kinetic heterogeneity model over the multi-exponential model is that it has fewer parameters (2 compared with 2N-1 for the multi-exponential model) and, in the absence of saturation, yields identical estimates of turnover. We confirmed that, even for the seven-day D2-glucose study, the multi-exponential and kinetic heterogeneity models give identical proliferation rate estimates, suggesting that, at least for CD4+ T cells, saturation is only a problem with the long labeling periods associated with D2O. The resulting proliferation rate estimates are shown in Fig 1. The differences between the one-day and seven-day D2-glucose estimates and between the one-day D2-glucose and 9 week D2O labeling estimates were significant (P = 0.01 and P = 0.003 respectively, two-tailed Mann-Whitney). Thus, although correcting for the saturation of label in rapidly turning over subpopulations helps to reduce the difference between the estimates (by increasing the proliferation rate estimates obtained in the 9 week D2O labeling study) significant discrepancies remain. Strikingly, these discrepancies are not only apparent in the comparison of D2-glucose and D2O labeling studies but also between the different D2-glucose labeling studies. That this is not an artefact of the modeling can be seen by studying the normalized experimental data. In the one-day D2-glucose study, labeling for only one-day resulted in deuterium enrichment in CD4+ T cells at day 3–5 similar to the enrichment obtained after 4 days of labeling in the seven-day labeling study (S1 Fig). This basic difference in data underlies the difference in estimated proliferation rates.
We sought to directly compare D2O and D2-glucose by labeling an immortalized human T cell line (Jurkat) in vitro either with D2O or with D2-glucose or with both compounds simultaneously. In all cases cells were labeled for seven-days with a seven-day wash-out phase. Data were normalized using the conventional strategy. Specifically, we adjusted for the level of label availability in both the D2-glucose and the D2O labeling experiments based on measured deuterium levels in culture media, and we estimated the maximum level of label enrichment that could be attained by fitting bw in the case of D2O and using a fixed value of bg = 0.65, based on the maximal end-product enrichment of such cells in vitro, in the case of D2-glucose labeling. Existing mathematical models, which assume lymphocyte steady state, were adjusted to allow for a growing cell population (Methods) and fitted to the normalized data. We assumed cell death in the cultures was negligible. We found that the estimates of proliferation obtained using D2O (p = 0.50±0.06 d-1, bw = 3.46±0.27) were similar to those obtained using D2-glucose (p = 0.52±0.03 d-1), and the difference between the estimates was not significant (P = 0.69, two-tailed Mann-Whitney), Fig 2, S1 Table, and S2 Fig. We conclude that estimates from in vitro D2-glucose and D2O labeling are in good agreement and that biochemical differences between D2O and D2-glucose did not lead to differences in proliferation rate estimates in this experiment.
To compare D2-glucose and D2O labeling in vivo we conducted a seven-day oral labeling study of mice. For one group of mice both the feed and the drinking water were spiked with D2O to an enrichment of 8%; the other group received feed spiked with D2-glucose (comprising ~30% of glucose intake). Identical chemical composition of feed between groups was maintained by spiking the feed for D2O-labeled mice with “unlabeled” (normal) glucose and for the glucose-labeled mice with “unlabeled” water. Food consumption was similar between the groups and all mice continued to gain weight. Deuterium enrichment was measured in blood plasma, PBMC, splenocytes, and thymocytes.
All cell types showed a progressive increase in DNA enrichment. Strikingly, the maximum enrichment in thymocyte DNA in glucose-labeled mice (about 4.5%) exceeded the estimated precursor (D2-glucose) enrichment, which averaged about 3.3% during the labeling period. Initially, the raw data were normalized following the conventional approach. That is, both D2-glucose and D2O data were first adjusted for plasma deuterium enrichment. The data were then scaled to the maximal level of enrichment; for D2O labeling this was determined using the plateau enrichment of a rapidly turning over cell population (thymocytes in this case), and for D2-glucose by using the in vitro derived constant factor bg = 0.65. The kinetic heterogeneity model was then fitted to the normalized data. This analysis (Figs 3A and 4) yielded substantially different proliferation rates for the D2O and D2-glucose labeling experiments (Fig 5). Similar (differences in) estimates were found using a multi-exponential model.
We were concerned by the large variations in D2-glucose plasma enrichment (Fig 3A) and the potential impact of mouse diurnal feeding patterns. By contrast, D2O plasma enrichment showed very little data variability (Fig 4A). In order to improve our estimate of glucose enrichment and reduce sampling error we repeated the D2-glucose arm of the experiment under identical conditions but with more frequent blood sampling (n = 27 plasma samples in 12 mice over seven-days) including both day and night sampling (facilitated by a reverse day-night cycle). As expected, enrichment in DNA from cells (thymocytes, splenocytes, PBMC) reached very similar values to those from the first experiment; specifically thymus labeling reached a maximum of about 5.1% with a modeled plateau of 4.9%. As before, large variations in plasma glucose enrichment were seen (S3 Fig) although these tended to stabilize with time, albeit at lower levels than in the first 24 hours, suggesting that label may be handled differently in the early stages of the experiment compared to later. When light-dark patterns were analyzed, a pattern could be discerned in which labeling reached a minimum in the middle of scotophase in line with previous publications [23] (Fig 6). Importantly, we found similar evidence for a diurnal pattern in humans that were labeled with D2-glucose [6]. In this study, subjects received D2-glucose by continuous infusion for 24h but also ate unlabeled low–carbohydrate meals. The plasma enrichment in these individuals tended to increase during the night (S4 Fig), which is to be expected as the intake of deuterated glucose remained constant whilst the intake of unlabeled glucose ceased as the individual sleeps (subjects were not woken to receive nutrition during the night).
We reasoned that variability in the plasma enrichment may make estimation of the D2-glucose availability problematic, we therefore reanalyzed the D2-glucose experiment normalizing to the estimated plateau enrichment of thymocytes, rather than using the mean plasma enrichment x bg. This reanalysis (Fig 3B) caused a dramatic reduction in proliferation rate estimates from the D2-glucose experiment resulting in good agreement between the proliferation rate estimates from the D2-glucose and D2O labeling experiments. The average proliferation rate of PBMC was 0.050 d-1 with D2-glucose and 0.051 d-1 with D2O; the average proliferation rate of splenocytes was 0.056 d-1 with D2-glucose and 0.064 d-1 with D2O (Fig 5). Again, similar estimates were found using a multi-exponential model. In short, when the D2-glucose and D2O data were analyzed using an identical normalization procedure, the estimates of cell proliferation were in good agreement; however when conventional normalization approaches (i.e. normalizing to the plasma deuterated glucose level) were used, differences emerged. These findings suggest that biochemical differences between these deuterium labeled compounds do not influence lymphocyte proliferation estimates in mice. However, the difficulty of estimating label availability throughout the course of the experiment may cause discrepancies.
Based on these in vitro and in vivo studies we hypothesized that normalization using the observed D2-glucose enrichment in the blood is problematic as this quantity fluctuates markedly and shows systematic diurnal variation, making it difficult to assess the label enrichment over the course of the study. To test this hypothesis we analyzed two published D2-glucose labeling studies (one-day and seven-day D2-glucose labeling of healthy individuals [6,13] where labeling in a rapidly turning over population (monocytes) was also available. If the average plasma enrichment during the labeling period is a good measure of intracellular DNA precursor enrichment, then, if we estimate the plateau enrichment of monocytes (i.e. the maximum label they can attain) this should be 100%. However if, as we hypothesize, the average D2-glucose enrichment in plasma during the labeling period is a poor surrogate for precursor enrichment, then the plateau enrichment of the monocytes will be significantly different from 100%. We constructed a model of monocyte development in which monocyte progenitors in the bone marrow proliferate, mature, exit into the blood and from there migrate into tissue to differentiate into macrophages or dendritic cells (Methods). We fitted this model to the normalized monocyte data from one-day and seven-day D2-glucose labeling studies and estimated the plateau enrichment of monocytes (Table 1 and S5 Fig). This analysis showed that, for the one-day labeling study, subjects tended to have a plateau enrichment significantly above 100% whereas for the seven-day labeling study, mean plateau enrichment was consistent with or slightly lower than 100% (Table 1). Repeating this analysis with a simpler but less physiological model confirmed these results (“delayed observation” model, Methods). For every model, the plateau enrichment was significantly higher in the one-day than in the seven-day labeling study (P = 0.048, P = 0.012, P = 0.012 for the bone marrow, delayed observation and a weighted combination of both models respectively, two-tailed Mann-Whitney). We conclude that there is evidence that the mean plasma enrichment may have underestimated the label available to dividing cells during the one-day D2-glucose study (and potentially overestimated label availability during the seven-day study). This conclusion is supported by an analysis of the deuterium enrichment in blood plasma glucose in the two studies. If we compare the labeling protocol of the one-day and seven-day D2-glucose studies, we see that the subjects in the one-day labeling study were infused with approximately twice as much D2-glucose per day as subjects in the seven-day labeling study. Despite this, the median plasma enrichment measured was only slightly higher in the one-day labeling study (Fig 7). These findings are consistent with the hypothesis either that the measured plasma enrichment in the one-day labeling study may have underestimated label availability (which would in turn have led to an overestimation of the T cell proliferation rate), or that carbohydrate intake, and thus glucose flux, was higher in subjects in the one-day study.
These investigations, which reveal a problem with the normalization in D2-glucose labeling studies, prompted us to also examine the normalization in D2O studies. Again, we focused on the 9 week D2O labeling study [8] where detailed data were available. In the published study the data in all T cell subsets were normalized to the maximum enrichment in granulocytes. Maximum enrichment in granulocytes was estimated based on the measured enrichment in urine and on an estimate of the intracellular scaling factor bw (c in the notation of the reporting paper [8]). The factor bw was estimated using a simple model in which granulocytes in the body were assumed to be a homogeneous population with a single rate of proliferation which was assumed to be equal to a single death rate. This yielded estimates of the intracellular dilution bw in the range bw = 3.78–5.15. Since the label in granulocytes comes close to reaching a maximum, there is little scope for error in this plateau estimation. Nevertheless the model used is unphysiological. We therefore refitted the data using the more physiological “bone marrow” model we developed (Methods), in which granulocyte precursors proliferate in the bone marrow, followed by a lag period before they enter the blood and then exit the circulation. This model fitted the data well (S6 Fig) and in most cases resulted in an AICc (a measure of goodness of fit penalized by number of parameters [24]) considerably lower than the published model despite its increased complexity (the published model consistently overestimated the fraction of labeled cells during the delabeling phase). The estimates of bw from the more physiological bone marrow model were significantly lower than the published estimates (P = 0.01, two-tailed paired Mann-Whitney, S7 Fig), though the effect is numerically small. This presumed overestimate of bw in the published study will have caused an underestimate of the proliferation rate in all subpopulations. Correcting for this underestimate leads to a small proportional increase in T cell proliferation rate estimates of between 1%-10% (P = 0.0003, two-tailed paired Mann-Whitney).
Several independent lines of evidence indicate that the discrepancy between proliferation rates estimated with D2O and D2-glucose, once saturation of highly dynamic pools has been accounted for [18,19], is not explained by a fundamental difference between the compounds but instead can be explained in part by the difficulty of using deuterium enrichment in plasma glucose to quantify label availability. Firstly, we find no evidence, either in vitro or in vivo, for a biochemical difference between D2-glucose and D2O that impacts on cell proliferation rate estimates. Secondly, in mice in vivo, estimates of cell proliferation rates obtained using D2-glucose and D2O labeling agree if labeled precursor availability during the course of the study is estimated using a rapidly turning over population but not if it is estimated using plasma glucose enrichments. Thirdly, in humans there is evidence that deuterium enrichment in plasma glucose fluctuates markedly with a predictable increase in enrichment during the night. If the sampling strategy is unbalanced between day and night (or more specifically between post-prandial and fasting) then the plasma glucose measurements may not reflect label availability throughout the labeling period. In the case of the one-day labeling study, plasma glucose profiles were based on 6–8 readings during the labeling period but sampling during the night was avoided to reduce inconvenience to participants and as a result plasma glucose enrichment during the whole labeling period may be underestimated. Fourthly, if we compare the labeling protocol of the one-day and seven-day D2-glucose studies we see that the subjects in the one-day labeling study were infused with approximately twice as much D2-glucose per day as subjects in the seven-day labeling study. Despite this, the median plasma enrichment measured in the two experiments did not differ two-fold, again indicating a potential problem in assessing precursor availability. Finally, when we estimate the plateau enrichment of monocytes in the one-day and seven-day human D2-glucose labeling studies we find that, although the data have been normalized such that the maximum label incorporation should be 100%, there is evidence in the one-day labeling study that the maximum incorporation exceeds 100%. Again, this would suggest that the plasma glucose enrichment levels used to normalize the data in the one-day D2-glucose study underestimated the label availability.
If, as the data summarized above suggest, the label availability has been underestimated in the one-day human D2-glucose studies then a direct consequence will be that T cell proliferation rates will have been overestimated. Estimation of the magnitude of these errors is nontrivial. The difficulty of estimating the plateau enrichment in monocytes after such a short labeling period means that correction factors for each individual may be unreliable. Instead we focus on correction factors averaged across the cohort (Table 1). This indicates proliferation rates in the one-day study have been overestimated by a factor of 1.2–2.4 (16–57%). In contrast, we find little clear evidence for a problem with the normalization in the seven-day D2-glucose study; if anything there is perhaps a suggestion that label availability may have been overestimated and thus T cell proliferation underestimated (Table 1). In the case of the 9 week D2O study there is evidence again of a (numerically small) problem in the normalization. Reanalysis of existing 9 week D2O data with a new model indicates that both naive and memory CD4+ and CD8+ T cell proliferation rates may have been systematically underestimated by 1–10%. Reducing the T cell proliferation rate estimates from the one-day D2-glucose study and increasing the proliferation rates from the 9 week D2O study will bring the estimates from the two methods significantly closer.
There are at least three potential causes for an underestimate of label availability in the one-day D2-glucose study. Firstly, in the one-day labeling study the sampling of enrichment in plasma glucose was biased towards daytime (5–7 points during the day, only 1 fasted overnight point, S3 Fig) which may have led to a systematic underestimate of label availability as we have shown that enrichment increases during the night. In contrast, in the seven-day labeling study, all of the plasma enrichments were measured after overnight fasting (when plasma enrichment will be at its peak); thus in the seven-day study the average label availability may be overestimated resulting in an underestimate of the proliferation rate (consistent with the average plateau enrichment in monocytes of less than 100%, Table 1). In the case of the seven-day labeling study, this error may have been relatively small as carbohydrate intake was severely restricted thus reducing fluctuations due to dietary glucose. Secondly, the blood glucose pool is a small pool with a rapid flux, which is highly dependent upon dietary intake. Although glucose concentrations are very tightly regulated these regulation mechanisms are unlikely to distinguish between labeled and unlabeled glucose. Thus whilst total glucose levels are stable, the fraction of enrichment may vary markedly, particularly following meals. Accurately capturing the mean plasma enrichment with only a few measurements from a highly variable profile will thus be difficult. Finally, it is routinely assumed that plasma enrichment will rapidly drop to zero after the end of labeling. However, in both the one-day and seven-day labeling studies, subjects were not completely fasted (though diet was restricted) therefore glycogen will have been synthesized from D2-glucose during this period. Following the end of labeling, this store of “heavy glycogen” would be released, either as D2-glucose following glycogenolysis or as D2O following glycolysis in the liver. Consequently, label availability may be higher than expected following the withdrawal of intravenous D2-glucose. Errors caused by recycling of label back into the circulation from glycogen will be particularly problematic for short term labeling as the unaccounted for label will be a higher proportion of total label. Furthermore, because dietary carbohydrate was more tightly restricted in the seven-day labeling study, postprandial fluctuations and recycling would be reduced compared to the one-day labeling study. These factors may explain why a problem with the estimation of label availability is apparent in the one-day but not the seven-day D2-glucose labeling study. One advantage of long-term D2O labeling studies compared with short term D2-glucose labeling is that label can be normalized to a rapidly dividing cell population within the same individual. Consequently the level of label availability is internally controlled. This does however introduce two problems. One is that the normalization is usually performed using a cell type (typically granulocytes or monocytes) that is different from the cell population studied. Any differences in label availability or usage between the reference cell population and the studied population will introduce error. Secondly, the plateau enrichment in the reference population is estimated by modeling and may therefore be dependent on the choice of model. Why bw, which would be expected to be constant, varies so markedly between individuals remains unexplained and may hint at further problems with normalization in the case of D2O. Although the problems associated with normalization appear to be more severe in the case of D2-glucose, we do not advocate replacing D2-glucose labeling with D2O labeling as there are many objectives, including labeling of rapidly turning over populations, evaluating time-courses of appearance and disappearance of labeled cells, and labeling of cohorts of cells, that cannot be achieved with D2O. Instead, we recommend steps (outlined below) to reduce normalization errors.
Based on this work we recommend that, in the case of D2-glucose labeling, plasma glucose enrichment is measured at frequent intervals, during both day and night as well as following the withdrawal of label. It is important that sampling times are not all just before or just after meals (the former would lead to a systematic overestimate of label availability, the latter to a systematic underestimate). Tightly controlled dietary carbohydrate intake will also help to reduce fluctuations in plasma enrichment. Important directions for future work with D2-glucose include the development of approaches to estimate rapidly fluctuating plasma glucose enrichments (since continuous blood sampling is not acceptable) and experiments to more accurately estimate the true in vivo value of bg. In the case of longer term D2O labeling, which permits normalization to a rapidly turning over subpopulation, we recommend checking for model dependence in estimates of bw and if necessary using the range of bw to provide a range of estimated lymphocyte proliferation rates. Additionally, assessing the maximal enrichment from a range of reference cell populations will highlight whether there are between-cell type differences in label usage. Future directions for D2O include firstly, investigating why bw, which would be expected to be constant, varies between individuals and secondly investigating the impact of the choice of the reference cell population on normalization.
In summary, by combining in vitro, murine and human work with modeling we have demonstrated that biochemical differences between D2-glucose and D2O are not responsible for discrepancies in proliferation rates obtained with these methods. Instead we conclude that problems with normalizing the data play an important role. The problems are most acute for short term D2-glucose labeling where the rapid flux, diurnal variation and potential for label recycling make accurate estimation of plasma glucose levels difficult.
To compare CD4+ T cell enrichment curves of different labeling studies, we calculated the enrichment level of total CD4+ T cells from a 9 week D2O labeling study [8], a seven-day D2-glucose labeling study [13] and a one-day D2-glucose labeling study in healthy individuals [6]. The seven-day D2-glucose labeling experiment directly studied total CD4+ T cells but the 9 week D2O experiment and the one-day D2-glucose experiments studied sorted “naïve” and “memory” CD4+ T cell subsets (defined as CD27+ CD45RO- and CD45RO+ respectively in [8] and CD45RO- and CD45RO+ respectively in [6]). For these two studies we used the enrichment data of naïve and memory CD4+ T cell subsets and their relative sizes within the total CD4+ T cell pool to recalculate the enrichment in total CD4+ T cells (total = naïve + memory). This approach could not be applied to total CD8+ T cells because there is a considerable fraction of CD27-CD45RO- CD8+ effector T cells, meaning that combining CD27+CD45RO- and CD45RO+ T cells is not equal to the total CD8+ population.
Six C57Bl/6J males, ~12weeks old, were housed in normal light/dark conditions; namely light for 12h from 08:00h to 20:00h and dark for 12h from 20:00h to 08:00h. A further 6 males were housed in reversed light conditions for 1 week prior to the experiment and then for the duration of the experiment (dark 08:00h-20:00, light 20:00–08:00h); this was to facilitate sampling of mice in scotophase during working hours.
All mice were labeled with deuterated (6,6-2H2-) glucose continuously. Mice ate a 30% deuterated glucose liquid feed ad libitum instead of their normal chow, mice also received normal drinking water.
Ethics statement: all human data were derived from existing studies. All data were anonymized and the original studies were approved by the relevant Ethics Review Boards.
We analyzed label enrichment in the DNA of monocytes from the one-day [6] and seven-day [13] D2-glucose studies. We constructed two new models, “bone marrow” and “delayed observation”:
We constructed a model of monocyte development (Fig 8) in which monocyte progenitors in the bone marrow proliferate, mature for a fixed time, exit into the blood and from there disappear into tissue (to mature into macrophages).
In the absence of label the system is described by the following set of differential equations:
M˙(t)=pM(t)−εM(t)B˙(t)=εM(t−Δ)−dB(t)
(5)
Where M(t) is the number of monocyte progenitors in the bone marrow at time t, p is their proliferation rate and ε their rate of exit into the maturation compartment prior to their exit into the blood. B(t) is the number of monocytes in peripheral blood, Δ is the time it take a monocyte progenitor to mature prior to entry into the blood and d is the rate of disappearance of blood monocytes from the blood into tissue. In the presence of label the system is described by
A˙M*=ψbgpU(t)AM−εAM*t≤τA˙M*=−εAM*t>τA˙B*=εAM*(t−Δ)−dAB*∀t
(6)
Where AM and AM* is total and labeled adenosine in DNA of monocytes in the bone marrow, AB and AB* is total and labeled adenosine in DNA of monocytes in the peripheral blood. bg is the intracellular dilution, ψ is the maximal enrichment (will be equal to 1 if the normalization is working correctly), τ is the length of the labeling phase, p, ε, Δ and d as above. U(t) is the label enrichment in the plasma as defined in Eq(4).
We define FM and FB as the fractional label enrichment in marrow and blood monocytes respectively normalized by intracellular dilution bg (fixed at 0.65) and plasma enrichment:
FM(t)=AM*(t)bgAMU, FB(t)=AB*(t)bgABU
Rearranging Eq (6) and eliminating ε and AM/AB by assuming both blood and marrow monocytes are independently at equilibrium in Eq (5) gives:
F˙M=ψp−pFMt≤τF˙M=−pFMt>τF˙B=dFM(t−Δ)−dFB∀t
This system was solved analytically and the theoretical quantity FB fitted to the normalized monocyte data. Data from all individuals was pooled to avoid overfitting. We fitted p, d, Δ as population parameters (they each take a single value for the whole population) and ψ as an individual parameter (each subject has a different value of ψ). So, for example, in the seven-day D2-glucose study there are 3 subjects and hence 6 free parameters (p, d, Δ, ψ1, ψ2, ψ3).
We also considered a second model, the “delayed observation” model, which is a simple extension of the kinetic heterogeneity model which has been described previously [20]. Briefly, the delayed observation model (like the kinetic heterogeneity model) assumes a heterogeneous pool but deals with the heterogeneity implicitly [20] (to minimize free parameters). In the delayed observation model this pool represents the population of bone marrow monocytes and is therefore observed in the blood with a lag:
F˙M=ψp−d*FMt≤τF˙M=−d*FMt>τFB(t)=FM(t−Δ)e−d*Δ∀t
Where FM is the fractional label enrichment in bone marrow monocytes, p is their proliferation rate, d* is the rate of disappearance of labeled monocytes, ψ is the maximal enrichment and Δ is the lag between an event in the pool and its observation in the blood (FB). The model was solved analytically and the theoretical quantity FB(t) was fitted to the monocyte data normalized with the intracellular dilution and mean plasma enrichment. Data from all individuals was pooled. We fitted p, d*, Δ as population parameters and ψ as an individual parameter.
We reanalyzed label enrichment in the DNA of granulocytes from the 9 week D2O study to estimate the scaling factor bw. We used the “bone marrow” model (described above) adjusted for D2O labeling:
F˙M=bwU(t)p−pFMF˙B=dFM(t−Δ)−dFB
Where FM is the fractional label enrichment in bone marrow granulocytes, p is their proliferation rate, d is the rate of disappearance of labeled granulocytes, U(t) is an empirical function describing the level of enrichment in urine, bw is the scaling factor and Δ is the lag between an event in the bone marrow and its observation in the blood (FB). For each of the five subjects studied (A-E) we used the previously estimated values of β, δ and f [8] to parameterize the function U(t) (Eq (3)). The theoretical quantity FB(t) was fitted to the raw granulocyte data and the parameters p, d, Δ and bw were estimated.
|
10.1371/journal.ppat.1004282 | Ly6Chi Monocyte Recruitment Is Responsible for Th2 Associated Host-Protective Macrophage Accumulation in Liver Inflammation due to Schistosomiasis | Accumulation of M2 macrophages in the liver, within the context of a strong Th2 response, is a hallmark of infection with the parasitic helminth, Schistosoma mansoni, but the origin of these cells is unclear. To explore this, we examined the relatedness of macrophages to monocytes in this setting. Our data show that both monocyte-derived and resident macrophages are engaged in the response to infection. Infection caused CCR2-dependent increases in numbers of Ly6Chi monocytes in blood and liver and of CX3CR1+ macrophages in diseased liver. Ly6Chi monocytes recovered from liver had the potential to differentiate into macrophages when cultured with M-CSF. Using pulse chase BrdU labeling, we found that most hepatic macrophages in infected mice arose from monocytes. Consistent with this, deletion of monocytes led to the loss of a subpopulation of hepatic CD11chi macrophages that was present in infected but not naïve mice. This was accompanied by a reduction in the size of egg-associated granulomas and significantly exacerbated disease. In addition to the involvement of monocytes and monocyte-derived macrophages in hepatic inflammation due to infection, we observed increased incorporation of BrdU and expression of Ki67 and MHC II in resident macrophages, indicating that these cells are participating in the response. Expression of both M2 and M1 marker genes was increased in liver from infected vs. naive mice. The M2 fingerprint in the liver was not accounted for by a single cell type, but rather reflected expression of M2 genes by various cells including macrophages, neutrophils, eosinophils and monocytes. Our data point to monocyte recruitment as the dominant process for increasing macrophage cell numbers in the liver during schistosomiasis.
| Schistosomiasis is an important neglected tropical disease caused by parasitic worms of the genus Schistosoma. During infection with S. mansoni, parasite eggs become trapped in the liver and elicit granulomatous inflammation characterized by accumulations of immune cells intermixed with liver cells around the eggs. This inflammation is responsible for disease symptoms, but also plays an important role in protecting the host against liver damage that can be caused by egg products. Granulomas, by definition, contain a significant number of macrophages (phagocytic cells of the immune system). Recent work has emphasized that macrophage numbers in inflammation can increase due either to recruitment of precursor cells (called monocytes) from the blood, or as a result of proliferation of tissue-resident macrophages. Local proliferation has been noted in other worm infections, during which the immune response is Th2-like and IL-4 produced by Th2 cells promotes macrophages to become “alternatively (or M2) activated”. We examined the origin of the increased numbers of macrophages in liver inflammation due to schistosomiasis, in which there is also a prominent Th2 response. We found that the cells mostly originated from monocytes recruited into the tissue from the blood. This response was critical for host survival during infection.
| Schistosomiasis is a complex multiorgan disease caused by infection with helminth parasites of the genus Schistosoma, and characterized by the development of granulomatous lesions around parasite eggs trapped within organs such as the liver and intestine. Granulomas are, by definition, macrophage-rich, but in schistosomiasis are recognized to contain substantial numbers of additional cells such as eosinophils, reflecting the fact that infection is associated with the development of a marked Th2 response [1].
In general, expanded macrophage populations in inflammatory sites are thought to arise from the recruitment of circulating inflammatory (Ly6Ch) monocytes from the bloodstream, and the in situ differentiation of these cells into macrophages [2]. Mobilization of monocytes from bone marrow is dependent on the interactions of CCL2 and CCL7 with CCR2, and can be promoted by type I interferons and TLR agonists [3]. Tissue-recruitment of monocytes is documented in viral, bacterial, fungal and protozoal infections [4], where monocytes that have entered tissues assume the characteristics of M1 macrophages, or TIP dendritic cells, both of which make TNFα and NO [3]. In contrast, little is known about the contribution of monocytes to inflammation during schistosomiasis. This is an interesting issue because recent reports have stressed that during Th2-dominated responses to nematode helminth parasites, macrophage rich inflammatory infiltrates arise as a result of IL-4-driven local macrophage proliferation, and not monocyte recruitment [5], [6]. A major effect of IL-4 on macrophages is to induce a program of gene expression that defines the M2 phenotype [7]. M2 macrophages are implicated in wound healing, metabolic homeostasis of adipose tissue, and in protective immunity to helminths, where type 2 immunity dominates [8], [9]
Schistosomiasis is a chronic infection. However, in the absence of IL-4, or IL-4Rα, the infection is acutely lethal [10]–[12]. The protective functions of IL-4 are considered to be macrophage-dependent, since infected IL-4Rαfl/flLysMCre mice develop the same lethal acute disease as infected Il4−/− mice [12]. Arginase 1 and Relmα, expression of both of which by macrophages is induced by IL-4/IL-13, are implicated in protective effects of M2 cells in schistosomiasis [13]–[15]. The absence of M2 macrophages during acute infection is associated with the development of cachexia with elevated levels of TNFα and NO, suggesting that commitment to the M2 pathway of activation contributes to the suppression or prevention of proinflammatory mediator production [10]. Other evidence supports the idea that M2 macrophages are playing a regulatory role in schistosomiasis, not only limiting inflammation but also preventing excessive tissue remodeling [14]. Nevertheless, despite the importance of macrophages in schistosomiasis, little is known of the origin of these cells in the liver during infection. Here we set out to establish whether the macrophage population hyperplasia in this setting results from monocyte recruitment or in situ proliferation. Our data suggest that a small percentage of resident macrophages in the liver are in cell cycle during infection, but that monocyte recruitment dominates as the mechanism for increasing macrophage numbers in this setting.
To begin to explore whether monocytes give rise to macrophages in inflamed hepatic tissues, we first examined monocyte numbers in the blood of naïve vs. infected mice. We found that infected mice developed a monocytosis, evident as increased numbers of Ly6Chi CD11b+CD115+ cells in the blood at week 7 of infection (Figs. 1A, B). This is consistent with elevated plasma levels of the chemokines CCL2 and CCL7 (Fig. 1C), the ligands for CCR2 (expressed on Ly6Chi cells, not on Ly6Clo monocytes [16]), that are responsible for mobilizing monocytes from bone marrow and splenic reservoirs into the circulation. We next asked when during infection monocytosis develops. We reasoned that increases in blood monocyte numbers at the time when egg-induced tissue inflammation is initiated would be consistent with the contribution of monocytes to the inflammatory process. We found that the number of monocytes in the blood of infected mice increased between weeks 5 and 7 of infection (Fig. 1D), correlating with the time of onset of parasite egg production [17]. Increased monocyte numbers in the blood were associated with increased monopoiesis, detected as an increased frequency and number of Ly6Chi monocytes in the bone marrow (Fig. 1E, F). Similarly coordinated increases in blood and bone marrow monocyte numbers have been seen in mice infected with L. monocytogenes [4].
If macrophages in diseased liver due to schistosomiasis are derived from monocytes, we would expect to detect extravasated monocytes within this organ. We used flow cytometry to characterize the cellular infiltrate in the liver, and, as anticipated from the histopathological picture [18], the infiltrate was complex. Using the gating strategy depicted in Fig. 2A, we were able to identify monocytes within the liver, and discovered a significant increase in this population as a result of infection (Fig. 2A, B, gate #1). The infiltrate was also rich in neutrophils (Fig. 2A, B, gate #2) and eosinophils (Fig. 2A, B, gate #3). To define macrophages we took advantage of recent findings that co-expression of the tyrosine kinase Mer (MerTK) and CD64 provides a robust marker for macrophages in the liver [19]. F4/80 was not a sufficient marker of macrophages in this setting, since it also marked eosinophils and monocytes (Fig. 1A, and [19]). In our initial experiments we found that the MerTK+CD64+ cells fell within a Siglec-F+ gate (Fig. 2A, gate #4). Siglec-F is an accepted marker for eosinophils, but it is also highly expressed on alveolar macrophages [20]. It clearly additionally marks macrophages within the liver during schistosome infection. We were able to identify a distinct MerTK+CD64+ macrophage population within the Siglec-F+ gate (Fig. 2A, B, gate #4). Enumeration of cells within gates 1–4 revealed significant increases in all populations as a result of infection (Fig. 2C).
To explore the possibility that the expanded population of hepatic macrophages associated with egg-induced inflammation largely results from the recruitment of monocytes rather than in situ proliferation of resident cells, we first asked whether sort-purified monocytes from liver tissues of infected mice were capable of differentiating into macrophages in response to M-CSF in vitro [21]. M-CSF is readily detectable in the serum of infected mice, where it is present at similar levels (7.9 ng/ml) as in naïve mice (6.6 ng/ml). After 7 days of culture with M-CSF, the cells had acquired a macrophage like appearance and expressed high levels of MerTK, CD64 and F4/80 equivalent to macrophages sorted from diseased liver (Fig. 3A). Next, we determined whether macrophages in the liver express CD115 or CX3CR1, markers that define blood monocytes [2], [22]. We found that CD115 was lost on monocytes that we recovered from the liver and therefore could not be used to trace their relationship to macrophages (data not shown). In contrast, CX3CR1, reported by GFP in Cx3cr1gfp/+ mice, was expressed by all Ly6Chi monocytic cells in the liver of infected Cx3cr1gfp/+ mice (data not shown), and in these animals the majority of hepatic MerTK+CD64+ cells were also GFP+ (Fig. 3B), indicating a significant contribution of monocytes to the macrophage pool during infection. In contrast, very few MerTK+CD64+ cells in naïve Cx3cr1gfp/+ mice were GFP+ (Fig. 3B). These data are consistent with a previous report that GFP-positive macrophages could be isolated from the livers of schistosome infected Cx3cr1 reporter mice [23].
In other settings of type 2 immunity due to helminths, macrophages have been reported to robustly incorporate BrdU in situ within 3 hours of BrdU injection as they proliferate in response to IL-4 [5]. We asked whether macrophages were proliferating in the liver during schistosomiasis by injecting mice with BrdU and then using flow cytometry to detect BrdU in MerTK+CD64+ macrophages at times thereafter. We additionally investigated BrdU incorporation into monocytes. The pulse-chase experimental design allowed us to effectively identify cells proliferating at the time of BrdU injection, and then to ascertain whether cells labeled within this period had differentiated into other cell types, not labeled within the first 4 h, but which were BrdU+ at later times (Fig. 3C). In this way the lineage relatedness of monocytes to macrophages could be defined. By 4 h after BrdU injection, 65% of monocytes in the bone marrow were labeled, a percentage that was approximately twice that observed in naive mice (data not shown). Labeled monocytes were not observed in the blood or liver at this time (Fig. 3C). However, approximately 6% of macrophages in the liver were BrdU+ (Fig. 3C), a percentage that was higher than in naïve mice (where <%2 of macrophages, defined by the same criteria, were positive for BrdU at this time point, data not shown). This percentage of macrophages that labeled within 4 h was similar to that observed after a 4 h BrdU pulse label in peritoneal macrophages of mice that had been injected i.p. with IL-4/anti-IL-4 complexes (Fig. 3D, and [5]). At 4 h after labeling, few BrdU+ monocytes were evident in the blood or liver (Fig. 3C). By 24 h approximately 35% of the monocytes in the blood and liver were BrdU+, reflecting the exit of labeled monocytes from the bone marrow into the circulation, and their entry into hepatic tissues (Fig. 3C). At this time, approximately 20% of macrophages in the liver were BrdU+. By 4–7 days after injection, labeled monocytes were rare in the bone marrow, blood and liver, but BrdU+ hepatic macrophages remained detectable, albeit in numbers that were decreasing within this timeframe (Fig. 3C). Nevertheless, when we gated on all BrdU+ bone marrow derived (CD45+) cells within the liver at the various time points of the experiment, we found that by days 4–7 post labeling, 75% of all label-retaining cells could be categorized as macrophages (Fig. 3E). Our conclusion from these studies is that the majority of macrophages in the liver under these conditions are derived from monocytes that have entered from the blood.
Recruitment of Ly6Chi monocytes to inflammatory sites is dependent on CCR2 [24], which is strongly expressed by these cells. We found that, compared to infected WT mice, the frequency and number of Ly6Chi monocytes in the blood (Fig. 4A, B) and the frequency of Ly6Chi monocytes in the liver (Fig. 4C) during infection in Ccr2−/− mice was greatly diminished. Strikingly, macrophage numbers in livers of infected Ccr2−/− mice were significantly reduced as well (Fig. 4F), supporting the hypothesis that macrophages in diseased liver arise from infiltrating monocytes. Nevertheless, the magnitude of the overall hepatic infiltrate in infected Ccr2−/−mice was similar to that in WT mice (Fig. 4D). This was accounted for by the fact that monocytes (Fig. 4E) and macrophages (Fig. 4F) were replaced by a marked increase in the number of Ly6G+ neutrophils in the absence of CCR2 (Fig. 4G). We noted little difference between these two strains in the infection-induced eosinophilia (data not shown). This picture, revealed by flow cytometry, was confirmed by histopathology (data not shown).
Given the reported importance of macrophages for survival in schistosomiasis, we were intrigued that infected Ccr2−/− mice survived despite a relative failure to increase liver macrophage numbers. However, in repeated experiments we noted that Ccr2−/− mice appeared sick, as indicated by immobility, huddling and piloerection between weeks 6 and 7 post infection (not shown). This was not accompanied by weight loss and while infected mice were sacrificed between weeks 7 and 9 post infection for the experiments shown, we found that infected Ccr2−/− mice would survive into chronic infection if permitted (data not shown).
To further explore the importance of monocyte-derived macrophages in schistosomiasis, we examined the outcome of infection in CCR2-Diphtheria Toxin Receptor (DTR) mice, in which cells that express CCR2 also express DTR and can be acutely deleted by the injection of DT [25]. We injected groups of infected CCR2-DTR and wild type (WT) mice with DT at week 6 of infection, which is just prior to the increase in blood monocyte numbers (Fig. 1D). This treatment resulted in rapid weight loss in CCR2-DTR mice but not WT mice, such that by day 4 post injection CCR2-DTR mice had lost approximately 20% of their body weight (Fig. 5A); the mice succumbed from infection by day 8 post injection. Naïve CCR2-DTR mice treated with DT remained healthy and did not lose any weight over the time period of the study (data not shown). To analyze the effects of CCR2+ cell depletion, we sacrificed mice at day 5 post initiation of DT treatment, and used flow cytometry to assess blood monocyte levels. As expected, we found that CD11b+Ly6Chi cells, which were also CD115+ (not shown), were lost in infected CCR2-DTR mice following DT treatment (Fig. 5B). DT treatment also depleted LyChi monocytes from the livers of infected CCR2-DTR mice (Fig. 5C). Serum levels of the liver enzyme alanine aminotransferase (ALT), a marker of hepatocellular injury, were no higher in DT treated infected CCR2-DTR mice than in DT treated infected WT mice, and DT treatment per se did not lead to liver damage in infected WT controls, since serum ALT levels in these mice were no higher than in untreated infected mice (Fig. 5D).
We next used flow cytometry to assess the effect of acutely deleting monocytes on hepatic macrophage populations in infected mice. We found no overall change in the frequency of MerTK+CD64+ cells in DT treated infected CCR2-DTR mice compared to DT treated WT mice (Fig. 6A), although there were substantially fewer hepatic leukocytes following monocyte depletion (data not shown). However, when we further examined MerTK+CD64+ macrophages for expression of F4/80 and CD11c, which can be used to distinguish subsets of macrophages within tissues [19], we found that a population of F4/80+CD11chi cells was markedly increased as a result of infection and depleted by DT treatment in infected CCR2-DT mice (Fig. 6A). The CD11chi macrophages additionally expressed high levels of MHC II (Fig. 6B). We used Ki67 staining to broadly identify cycling cells in the resident CD11clo macrophages and monocyte-derived CD11chi macrophages. We found that Ki67 expression was increased in the CD11clo population in infected vs. naïve mice (Fig. 5B), suggesting that resident macrophages are responding to infection, a conclusion further supported by the fact that MHC II expression is increased on this population in infected mice (Fig. 6B). In our analyses, CD11chi macrophages were also positive for Ki67 (Fig. 6B), suggesting that these cells were proliferating, however this was true whether the cells were from naïve or infected mice, suggesting perhaps that this is an intrinsic property of these monocyte-derived cells (Fig. 6B).
We used histology to assess whether deletion of monocytes had an effect on hepatic granuloma formation. We found that DT treatment in infected CCR2-DTR mice resulted in significantly reduced granulomatous inflammation (Fig. 7). Many eggs in these mice were associated with little or no inflammation (Fig. 7A), a situation observed only rarely in infected WT mice. Moreover, when granulomas were apparent in DT-treated CCR2-DTR mice (Fig. 7B–F), they were significantly smaller than those in infected DT-treated WT mice (Fig. 7G,H,I,J). We also noted marked accumulations of cells close to blood vessels in infected DT treated CCR2-DTR mice that were not apparent in infected DT-treated WT mice (Fig. 7D); the majority of cells in these accumulations were lymphocytic in appearance.
Disease progression marked by weight loss in DT treated infected CCR2-DTR mice was similar to that reported for infected IL-4−/− and IL-4Rα−/− mice [10]–[12], raising the possibility that depletion of monocytes had secondary effects on the Th2 cell response, with a resultant reduction in IL-4/IL-13 production. To examine this possibility, we prepared splenocytes and measured cytokine production following restimulation with soluble schistosome egg antigen (SEA). We found that the production of IFNγ, IL-5, IL-10 and IL-13 were largely similar between DT treated WT and CCR2-DTR mice (Table 1), although there was a trend towards greater production of IL-5 and IL-13, and lower production of IL-10 in the absence of CCR2+ cells. Strikingly, the only marked difference observed was in IL-4 production, where levels of this cytokine were higher in culture supernatants of splenocytes from DT treated CCR-DTR vs. WT mice. IL-17 levels were below the level of detection in our assay.
The fact that monocyte recruitment rather than in situ macrophage proliferation was primarily responsible for the increased hepatic macrophage population in infected mice led us to question the extent to which liver macrophages are M2 activated as a result of infection. We know that the M2-polarizing cytokine IL-4 is present within the hepatic milieu during infection [26], and previous reports indicate that hepatic macrophages are M2 activated in this setting [27], [28]. However, previous reports preceded the identification of stringent markers of macrophages [19]. Therefore, we sorted macrophages using MerTK and C64 expression and used real time RT-PCR to assess gene expression in these cells as well as in whole liver tissue. Using real time RT-PCR, we found that expression of the M2 markers Arginase-1 (Arg1), Relmα (Retnla), YM1 (Chi3l3), as well as the M1 markers TNFα (Tnf), and iNOS (nos2), was increased in liver from infected vs. naive mice (Fig. 8A), pointing to a complex array of signals that could favor monocyte recruitment to the liver even as IL-4 might also trigger macrophage proliferation. To determine which cell type(s) was responsible for expressing the genes under study, we sort-purified infiltrating cells based on the approach shown in Fig. 2A. We found that the overall M2 bias in the liver is accounted for by the expression of Relm-α in macrophages and eosinophils, Arginase-1 in monocytes, and YM-1 in monocytes and neutrophils (Fig. 8B). Our results for Relmα are consistent with previous reports, where macrophages and/or eosinophils have been shown to be major sources of this protein during schistosome egg associated inflammation [13], [15]. Thus, the overall M2 bias observed in the liver during acute infection is the result of a mosaic of gene expression in different cell types. Our gene expression analysis additionally revealed that Ly6Chi monocytes in the liver of infected mice were expressing nos2 and tnf (Fig. 8C).
Granulomatous inflammation rich in macrophages is common in helminth infections, when large parasite life stages become trapped in and or die within tissues during targeted or aberrant migration. While many features of helminth infection-associated granulomas are well documented, the origin of macrophages within them has remained relatively unclear. This question became particularly intriguing when it was reported that macrophages activated by IL-4, the quintessential product of Th2 cells which tend to dominate the immune response during helminth infections, undergo a proliferative response that can result in local population expansion. This process differs fundamentally from the recruitment of blood monocytes into tissues, and their subsequent differentiation into macrophages, which is considered to be the dominant source of macrophages associated with Th1 cell-associated inflammation in, for example, infections with bacteria or protists. During infection with S. mansoni, a strong Th2 response develops within the week following the initiation of egg production by mature female parasites and parasite eggs become trapped in the liver sinusoids and form focal points for intense CD4+ T cell dependent granulomatous inflammation. Because of the well-documented bias towards Th2 immunity in this system, and indeed the importance of IL-4 and of IL-4Rα expression by macrophages for the survival of mice infected with schistosomes, we hypothesized that macrophage activation within diseased liver tissue would be clearly M2 in nature (a notion supported by previous reports) and that local proliferation would therefore account for a significant component of the macrophage infiltrate. Our data reveal a different picture, in which the macrophage population exhibits signs of classical as well as alternative activation, and the recruitment of monocytes is the major process supporting the expansion of the macrophage population within the liver.
Our data clearly show a role for CCL2 in the recruitment of monocytes into liver tissue during infection with schistosomes. Levels of this chemokine were significantly elevated in the blood as a result of infection, and mice in which cells expressing the CCL2 receptor CCR2 were depleted had significantly diminished blood and liver monocyte numbers compared to infected WT mice. Our data are consistent with findings from previous reports that utilized a model system in which mice are injected intravenously with schistosome eggs, which embolize in the pulmonary vasculature and induce granulomatous inflammation. CCL2 is produced in the lungs in this model [29], and in mice that lack CCR2 the macrophage content of the lung granulomas that develop as a result of egg injection is initially reduced compared to that in egg injected WT mice [30]. Moreover, ccl2-/- mice were found to develop smaller granulomas in this model of egg induced granuloma formation [31]. Taking all of the findings together, it is clear that schistosome eggs induce CCL2 production and that CCR2/CCL2 play a role in monocyte recruitment during egg induced inflammation, and that monocytes recruited in this way are contributing to the expansion of the macrophage population at these sites. This is consistent with a role for CCL2 in increased macrophage numbers at sites of infections with other helminths [32], and with microbial pathogens [33], [34]. Monocytes can differentiate into macrophages in response to M-CSF [21] and into dendritic cells in response to GM-CSF [2], [35]. While we were able to detect M-CSF in the serum of infected mice, GM-CSF was undetectable (data not shown). However, it remains possible that GM-CSF is made within the hepatic environment of infected mice, and it will be interesting to examine dendritic cell differentiation from monocytes in this setting in the future.
Earlier work suggested that in addition to Ly6Chi monocytes, Ly6Clo monocytes are also recruited to the site of L. monocytogenes infection, via CX3CR1, and give rise to M2-like macrophages [36]. Our work argues that, in schistosomiasis, it is rather the Ly6Chi monocytes that contribute to the pool of M2-like macrophages. In this regard, our findings are broadly consistent with recent findings that M2 macrophages in allergic skin are derived from inflammatory monocytes [37]. Our findings using CCR2-DTR mice revealed a critical role for monocytes for survival during acute schistosomiasis. While the system does not let us delineate between the importance of monocytes themselves and cells that differentiate from them, it was notable that in these animals, injection of DT resulted in loss of a subpopulation of CD11chi hepatic macrophages as well as of circulating and hepatic monocytes. Macrophages with this phenotype were present at significantly higher frequency following infection, and their loss in infected mice was evident histologically as significantly reduced hepatic granulomatous pathology, which nevertheless was associated with exacerbated disease. The latter is not altogether unexpected since granulomas are recognized to perform a host-protective function by sequestering egg hepatotoxins (discussed in [17]). Nevertheless, if this was the primary protective mechanisms disrupted by loss of CD11chi macrophages, we might have expected to see evidence of hepatocyte damage within the vicinity of parasite eggs, and this was not readily apparent. The lack of any difference in serum levels of the liver enzyme ALT between DT-treated WT and DT-treated CCD2-DTR mice supports the view that deletion of monocytes did not exacerbate liver damage in infected mice. It is possible that the major effect of monocyte-deletion during infection is in the intestine, although there was no gross evidence of severe intestinal damage or hemorrhage. Moreover, we failed to detect endotoxin in serum samples from DT treated infected WT or CCR2-DTR mice (<0.005 EU/mL, data not shown), which supports the view that intestinal integrity was maintained even when monocytes were acutely depleted (although a caveat of this conclusion is that we did not specifically test portal blood in this regard). We are also unclear why infected CCR2−/− mice and DT-treated CCR2-DTR mice did not phenocopy each other. This lack of concordance is consistent with the development of a compensatory mechanism in the germline knockouts that provides the mice sufficient advantage to make it through acute infection. This could be related to the intense neutrophil infiltrate in the infected Ccr2−/− mice, which was not apparent in the DT-treated infected CCR2-DTR mice (data not shown). Understanding the difference in outcome of infection in these two strains of mice, along with the identification of the cause of death of infected mice acutely depleted of monocytes, awaits further investigation.
Recent work indicated that, in the steady state, liver macrophages are embryonically derived and genetically distinct from monocyte-derived cells and persist at the population level independently of replenishment from the bone marrow [38]. These macrophages are likely those that are CD11clo and which resist deletion by DT treatment in CCR2-DTR mice (although an alternative possibility is that these cells are derived from monocytes prior to the initiation of DT treatment). This population responds to infection by increased expression of MHCII and proliferation, but its exact function in the context of schistosomiasis remains unknown at this time. We suspect that it may play an important role, because broad depletion of macrophages using clodronate liposome injection at 7 weeks of infection caused a significantly more rapid onset of lethal disease than did DT administration to CCR2-DTR mice, such that all treated animals (n = 15) died within 24 h (data not shown; clodronate treatment of naïve mice caused no noticeable ill effects). A tentative conclusion from these results is that the CD11clo macrophages cooperate with the CD11chi macrophages in a host-protective role during schistosome infection.
One protective effect of M2 macrophages in schistosomiasis is to regulate the intensity of Th2 immune responses. This is critical in schistosomiasis, where unregulated Th2 responses can lead to excessive and sometimes lethal immunopathology. Arginase 1 and Relmα, both of which are considered to be canonical markers of M2 activation, have been shown to play roles in these processes, although their importance appears to be greater at times later in infection than those tested here [14], [15]. Nevertheless, it is possible that the marked effects of deletion of CCR2 cells that we observed here reflect the loss of monocytes or macrophages capable of making Arginase1 and Relmα. Of interest in this regard, we observed a trend towards increased production of IL-4, IL-5 and IL-13, and decreased production of IL-10 by splenocytes from infected DT treated CCR2-DTR mice, which is consistent with the loss of mechanisms that regulate Th2 cells. Our failure to detect IL-4 in the supernatants of SEA-stimulated splenocytes from infected WT mice is likely due to the consumption within the cultures of this cytokine, since in previous reports we have shown that the addition of blocking anti-IL-4Rα antibodies allows subsequent accumulation of IL-4 to measurable levels. A corollary of this is that CCR2+ cells are precisely those responsible for consuming IL-4. This possibility will be addressed in future studies.
Our data support the view that there is a conflict of pro- and anti-inflammatory events occurring within the liver during infection, and suggest that the overall pattern of immune gene expression in this organ reflects the involvement of different cellular populations within the infiltrate. Interestingly, both NO and TNFα, which we found to be made by infiltrating monocytes during infection, are implicated in severe morbidity and increased mortality due to schistosomiasis in the absence of IL-4 [17], raising the possibility that it is unregulated inflammatory monocyte invasion and activation that underpins this condition. Whether individual monocytes express Arginase1, iNOS and TNFα, or whether the RT-PCR data reflect the existence within the monocyte gate of cells committed to the production of one, or more than one of these mediators, is unknown at present. These questions are being addressed in ongoing research.
The difficulty of correlating findings from in vitro macrophage polarization experiments and in vivo infections has been addressed previously [39], and our findings support this view. Despite the well-documented dominance of the Th2 response during schistosomiasis [17], it is clear that in this infection macrophages do not conform fully to the M2 phenotype observed in IL-4 stimulated bone marrow derived macrophages [7], [40]. This likely reflects the impact of other cytokines, such as IFNγ, which may modulate the M2 phenotype during infection [15]. Indeed, the presence of iNOS and TNFα expressing cells in the liver may reflect the superimposition of an inflammatory signal arising from the leakage of TLR agonists from the intestinal flora into the portal bloodstream as schistosome eggs follow their natural route of egress from the vasculature into the gut lumen, perforating the epithelium en route [17]. In this way, schistosomiasis differs from the “sterile” nematode infection or implantation models in which resident macrophage proliferation in the absence of monocyte recruitment has been documented [5], [41].
All work was conducted in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee of Washington University in St. Louis School of Medicine (Animal Welfare Assurance Number: A-3381-01).
C57BL/6J, B6.129P-Cx3cr1tm1 Litt/J (Cx3cr1gfp/gfp), B6.129S4- Ccr2tm1Ifc/J (Ccr2−/−), male mice (all from Jackson Laboratories) were used in this study. B6 CCR2-DTR mice were a gift from Dr. Eric Pamer, and were bred at Washington University. For experiments using these mice, CCR2-DTR heterozygotes were compared with WT littermates. Cx3cr1gfp/+ mice were obtained by breeding Cx3cr gfp/gfp mice with C57BL/6J mice. Cx3cr1 gfp/+ mice have one Cx3cr1 allele replaced with cDNA encoding Egfp, and can be used to track monocytes [22]. Mice were kept under specific pathogen–free conditions and were infected at 8–12 weeks of age. Mice were each percutaneously exposed to 150 Schistosoma mansoni cercariae. Snails were provided by the Schistosome Research Reagent Resource Center for distribution by BEI Resources, NIAID, NIH: Schistosoma mansoni, Strain NMRI Exposed Biomphalaria glabrata, Strain NMRI, NR-21962. Uninfected mice were used as controls. Mice were euthanized at 7–9 weeks post infection.
Tissues were fixed in neutral buffered formalin, embedded in wax, sectioned and stained with hematoxylin and eosin. Slides were initially read and interpreted in a blinded fashion. The diameters of granulomas (around single eggs) were measured in a horizontal plane bisecting central eggs using the LAS imaging program (Nikon).
CCL2, CCL7, M-CSF and GM-CSF levels in plasma pools from 5 uninfected or 5 infected mice were measured using the Rodent MAP v2.0 by RBM-Myriad. Cytokines in splenocyte supernatants were measured by cytometric bead array (CBA, BD). Bacterial Gram negative endotoxin in serum samples was measured using the Chromo Limulus Amebocyte Lysate assay (Associates of Cape Cod Inc.). ALT in serum samples was detected using a coupled enzyme assay to colorimetrically measure enzyme activity (Sigma-Aldrich).
Peripheral blood was drawn via cardiac puncture with heparin. Erythrocytes were removed by incubating with red blood cell lysis buffer. Bone marrow was flushed out of femurs using RPMI. To analyze hepatic cell populations livers were removed from PBS-perfused animals, crushed with a syringe plunger, and incubated in RPMI (Hyclone) containing 250 µg/ml of Collagenase D (Roche) at 37°C for 60 min. The resulting suspension was disrupted through a 100 µm metal cell strainer and centrifuged at room temperature for 20 min at 2500 RPM through 40% isotonic Percoll/RPMI. The resulting pellet was washed, and used for analysis. Splenocytes were prepared and cultured alone or with soluble egg antigen (SEA, 50 µg/ml), as previously described [42].
Diphtheria toxin (DT) was from List Biological Laboratories (Cat. No. 150), reconstituted at 1 mg/ml in PBS, and frozen at −80°C. Mice received 10 ng/g DT via the i.p. route in 0.2–0.3 ml PBS [25].
Mice were each treated by i.v. injection of 250 µl of liposomes containing clodronate (clodronateliposomes.com).
Mice were injected (i.p.) with 1.0 mg BrdU (BD Biosciences) dissolved in PBS at 4 hours, 24 hours, 4 days, and 7 days before experimental end-points. Cells were prepared as described above and the BrdU in situ detection Kit (BD Bioscience) was used according to the manufacturer's instructions to stain for BrdU incorporation.
Samples were blocked with 5 µg/mL of anti-CD16/CD32 (eBioscience) and incubated with combinations of the following antibodies, which were either directly conjugated to fluors, or biotinylated: anti-Ly6C (HK 1.4, BD Biosciences); anti-Ly-6G (1A8, BD Biosciences); anti-CD11b (M1/70, BD Biosciences); anti-F4/80 (BM8; eBioscience); anti-CD11c (HL3, BD Biosciences); anti-I-A/I-E (M5/114.15.2, Biolegend); anti-CD115 (AFS98, Biolegend); anti- CD45 (30F11, Biolegend); anti-CD64 (X54-5/7.1,BD Bioscience); anti-mMer (MerTK, R&D Systems); anti-Siglec F (E50-2440, eBioscience); Ki67 (B56, BD Biosciences, used in conjunction with elements of the BrdU detection kit for cell permeabilization). Strep-PECy7 (BD Biosciences) was used to detect biotinylated antibodies. Cells were stained with LIVE/DEAD (Invitrogen) before analysis. Data were acquired on a Canto II (BD Biosciences) and analyzed with FlowJo v.8.8.6 (Tree Star, Inc.). Cells were sorted on a BD FACSAria (BD Biosciences). For morphologic characterizations, sorted cells were prepared on slides by cytocentrifugation at 1000 RPM for 5 min, and stained with HEMA-3 (Fischer Scientific).
FACS-sorted Ly6Chi monocytes from blood and livers were treated with M-CSF (0.02 µg/mL) in RPMI (Hyclone) supplemented with 10% FCS (Hyclone), 50 µM 2-Mercaptoethanol (Cellgro, Mediatech, Inc., Va) and 100 U/mL Penicilin-Streptomycin (Cellgro, Mediatech, Inc., Va). Cells (105) were plated in triplicate in 96-well round bottom plate (Costar, Corning Inc., NY) and cultured in a humidified incubator at 37%, 5% CO2 [21]. Cells were harvested at day 7, and the expression of the surface marker F4/80, CD64 and MertK were determined.
Classical monocytes, neutrophils, eosinophils and macrophages were sorted from the leukocyte populations isolated from livers pooled from 5–10 naïve or infected mice. Pieces of liver were also used for RNA isolation. Cells or liver samples were resuspended in 1.0 mL Trizol (Invitrogen) and RNA was extracted using the manufacturer's instructions. Qiagen's RNeasy Micro “RNA clean-up” protocol with an on-column DNase treatment was used to purify the RNA further. RNA yield and integrity were assessed with a NanoDrop spectrophotometer (Thermo-Scientific). PCR cDNA was generated from 500 ng of RNA per sample by reverse transcription (RT). First strand cDNA was synthesized using isolated RNA, Superscript II reverse transcriptase (Invitrogen), and oligo dT as a primer. Relative quantification of the genes was performed by Applied Biosystems 7500 real-time PCR system with Taqman Gene Expression Master mix (Invitrogen) and the 2-ΔΔCt analysis. Total reaction volume was 20 µL with 300 nM of each primer/probe, 10 µl of master mix, and 1 µl of cDNA as template (or water as a negative control). Expression of Arg1 (Mm00475988_m1), Retnla (Relmα; Mm00445109_m1), Chi3l3 (YM1; Mm00657889_s1), tnf (Mm00443260_g1) and Nos2 (Mm00440502_m1), were normalized to β-actin, Actb (Mm00607939_m1).
Results are expressed as mean ± SEM, or ± SD. Statistical tests included unpaired, 2-tailed Student's t test. P values of 0.05 or less were considered to denote significance.
|
10.1371/journal.pgen.1003721 | Environmental Stresses Disrupt Telomere Length Homeostasis | Telomeres protect the chromosome ends from degradation and play crucial roles in cellular aging and disease. Recent studies have additionally found a correlation between psychological stress, telomere length, and health outcome in humans. However, studies have not yet explored the causal relationship between stress and telomere length, or the molecular mechanisms underlying that relationship. Using yeast as a model organism, we show that stresses may have very different outcomes: alcohol and acetic acid elongate telomeres, whereas caffeine and high temperatures shorten telomeres. Additional treatments, such as oxidative stress, show no effect. By combining genome-wide expression measurements with a systematic genetic screen, we identify the Rap1/Rif1 pathway as the central mediator of the telomeric response to environmental signals. These results demonstrate that telomere length can be manipulated, and that a carefully regulated homeostasis may become markedly deregulated in opposing directions in response to different environmental cues.
| Over 70 years ago, Barbara McClintock described telomeres and hypothesized about their role in protecting the integrity of chromosomes. Since then, scientists have shown that telomere length is highly regulated and associated with cell senescence and longevity, as well as with age-related disorders and cancer. Here, we show that despite their importance, the tight, highly complex regulation of telomeres may be disrupted by environmental cues, leading to changes in telomere length. We have introduced yeast cells to 13 different environmental stresses to show that some stresses directly alter telomere length. Our results indicate that alcohol and acetic acid elongate telomeres, while caffeine and high temperatures shorten telomeres. Using expression data, bioinformatics tools, and a large genetic screen, we explored the mechanisms responsible for the alterations of telomere length under several stress conditions. We identify Rap1 and Rif1, central players in telomere length maintenance, as the central proteins directly affected by external cues that respond by altering telomere length. Because many human diseases are related to alterations in telomere length that fuel the disease's pathology, controlling telomere length by manipulating simple stressing agents may point the way to effective treatment, and will supply scientists with an additional tool to study the machinery responsible for telomere length homeostasis.
| Telomeres are nucleoprotein structures located at the ends of chromosomes. Telomeres are essential for chromosome replication and stability [1], and protect chromosome ends from degradation and deleterious chromosomal rearrangements [1], [2]. In human embryonic cells, telomeres are elongated by the enzyme telomerase [3]. In somatic cells, however, telomerase expression is low, and telomeres shorten with each cell division due to the incomplete replication of the linear chromosome ends by conventional DNA polymerases. This progressive telomere shortening constitutes a “molecular clock” that underlies cellular aging [4]. Accordingly, telomere length is associated with cell senescence and longevity [5], as well as with age-related disorders and cancer [6]. While short telomeres have been reported to predict early mortality [7], recent work has shown that telomerase reactivation may reverse tissue degeneration in aged telomerase-deficient mice [8].
Three systematic genome-wide surveys in the yeast Saccharomyces cerevisiae [9]–[11] have revealed that mutations in at least 6% of the genes lead to alterations of telomere length. These TLM (Telomere Length Maintenance) genes span a broad range of functional categories and different cellular compartments. Integration of data from these large-scale mutant screens with information about protein–protein interactions has further permitted charting of the cellular sub-network underlying telomere length regulation in yeast [12], [13], revealing a complex set of interactions responsible for a very tight length homeostasis.
Environmental stresses affect the regulation and the activity of many genes and accordingly may perturb telomere length homeostasis by altering the expression or activity of genes in the TLM network described above. Previous studies have suggested that emotional stress in humans is associated with telomere shortening, presumably through its effect on oxidative stress [14], [15]. These studies, however, establish a correlation, but not causality. Here, we use controlled experimental approaches to explore a possible effect of the environment on yeast telomere length, and to identify the molecular mechanisms by which external signals exert their effect.
We exposed yeast cells (S. cerevisiae) to thirteen different environmental signals for 100–400 generations (Figure 1 and Table S1). Our results show that some stresses, such as high temperature, the addition of caffeine, and low levels of hydroxyurea resulted in telomere shortening, while others, such as added acetic acid and alcohols including ethanol, methanol, and isopropanol, caused a significant increase in telomere length (Figure 1). Strikingly, under alcohol stress telomeres were not only longer, but also exhibited length heterogeneity, indicating that the mechanism responsible for telomere length homeostasis, which preferentially elongates short, but not long telomeres [16], was disrupted (Figures 1, 2). The effect of alcohols on telomere length was independent of the ability of these cells to metabolize the alcohol: Upon ethanol treatment, isogenic petite yeast strains (lacking mitochondrial function, and thus unable to utilize ethanol) exhibited elongated telomeres (Figure S1).
Importantly, however, many other environmental stresses, including oxidative stress, did not significantly alter telomere length (Figure 1 and Table S1), indicating that telomere length homeostasis is robust under many other environmental conditions. The effect of each stress on telomere length was concentration-dependent. In all cases, removal of the stressing agent resulted in a gradual restoration of wild type telomere length (Figure 2A–C), demonstrating that the changes in telomere length were physiological rather than genetic, and thus may have been mediated by altered gene expression and protein activity.
Under unperturbed conditions, telomere length can be modified either by disrupting the regulation of telomerase/telomere-associated nucleases or by recombination. To distinguish between these two mechanisms, we analyzed the response to stresses of cells unable to carry out homologous recombination due to a deletion of the RAD52 gene. rad52 cells responded to the stresses much as would a wild type strain, indicating that telomere length alteration in response to these stresses is not recombination-dependent (Figure S2) and that the external signals affect telomerase or telomere-associated nucleases.
To understand how external signals affect telomere length and to identify the mechanism behind this telomeric response to stress, we measured genome-wide transcript levels in yeast cells grown for 20 generations in the presence of stresses that showed an effect on telomere length (ethanol, caffeine or high temperature), as well as in the presence of H2O2, a stress that does not alter telomere length. The results were compared to genome-wide transcript levels of the same strain grown under standard conditions (YEPD medium, 30°C). Using Significance Analysis of Microarrays (SAM) [17] with a false discovery rate (FDR) below 0.01, we obtained a set of 1,744, 1,404, 1,670 and 1,019 differentially expressed genes for caffeine, 37°C, ethanol and H2O2, respectively. General environmental stress responding (ESR) genes were not induced under these conditions, as expression level was measured after a long-term exposure to the stresses while ESR genes are induced for a short time period [18]. To identify the mechanisms responsible for telomere elongation and shortening, we sought genes that were differentially expressed only under shortening or only under elongating conditions (Figure S3). We integrated transcript abundance data with the known TLM network [13] that uses protein-protein interactions data, connecting TLM genes to the telomere maintenance machinery. The (unweighted) pairwise distances between stress-specific differentially expressed TLM genes were compared with pairwise distances of other TLM genes. This revealed that stress-specific, differentially expressed TLM genes lie significantly closer to each other for ethanol, caffeine and 37°C (p<2E-33,p<3E-27 and p<3E-50, respectively), but not for hydrogen peroxide stress, which does not affect telomere length (Materials and Methods). This phenomenon was unique to TLM genes under stresses that affect telomere length, suggesting that the differentially expressed TLM genes may be involved in transducing the external signals and disrupting telomere length homeostasis.
Based on the analysis above, we generated a list of candidate genes for further analysis. Using strains from the yeast deletion library [19] and the DAmP library of hypomorphic mutants [20] we screened mutants in this list to identify genes important for telomere length maintenance under stress conditions. Strikingly, we found a strong correlation between the rate of change in telomere length and the initial length of the mutant: in ethanol, long tlm mutants elongate more rapidly than the wild type, while short tlm mutants elongate more slowly (Pearson correlation, r = 0.61, p<E-12, Figure 3A). Similarly, in caffeine and at 37°C long tlm mutants shorten more rapidly, while short tlm mutants shorten more slowly than does the wild type (Pearson correlation, r = −0.78, p<2E-22 and r = −0.96, p<9E-34, respectively; Figure 3B–C). This correlation between abnormal telomere length and response magnitude to the stresses suggests that telomere elongation/shortening in the presence of external cues is carried out by the same basic mechanisms that maintain telomere length under unperturbed conditions.
To identify the genes that mediate the telomeric response to stress and to understand how external signals are transduced to altering telomere length, we focused on mutants that disrupt this transduction and, therefore, show an atypical response to each stress (Figure 3). A remarkable such tlm mutant is rif1Δ, which exhibited a reduced response to ethanol and caffeine but normal response to 37°C (Figures 3A and 4), indicating that elongation by ethanol and shortening by caffeine are Rif1-dependent, while telomere shortening by high temperature relies on a different mechanism.
The Rif1 and Rif2 proteins are negative regulators of telomerase that interact with the C-terminus of Rap1, an essential protein that binds to the telomeric repeats [21]. Under normal growth conditions, short telomeres are preferentially elongated by a mechanism that depends on Rap1. Mutations in the carboxy-terminus of RAP1 or down-regulation of the RAP1 gene lead to extreme telomere elongation and to an increase in telomere length variability, similar to what we observed in the presence of ethanol ([22], [23]; Figure 2). Our transcript measurements detected a reduction in the level of Rap1 expression in cells grown in the presence of ethanol [(Table S2); and [24]]. These results suggest a model in which telomere elongation under ethanol stress is primarily due to reduced levels of Rap1, which reduce Rif1 recruitment to telomeres. To test this hypothesis, we used a strain in which RAP1 was expressed from a Tetracycline-inducible promoter [25]. In this strain the level of Rap1 remained unchanged in the presence of ethanol (Figure 4A) and only a slight telomere elongation was observed (Figure 4B). Also consistent with the model, a rap1-17 strain (deleted for the C terminus of Rap1), a rif1Δ single mutant and a rif1Δ rif2Δ double mutant exhibited attenuated responses to ethanol (Figure 4B). Thus, the telomere elongation response to ethanol was abolished when a steady level of Rap1 protein was maintained or when Rif1 activity was eliminated, indicating that the Rap1- Rif1 pathway is central to telomere elongation in response to ethanol. Consistent with this hypothesis, chromatin immunoprecipitation (ChIP) experiments showed that upon exposure to ethanol there is a two-fold reduction in the level of Rif1 at telomeres, as well as a slighter reduction in the level of Rif2 (Figure 4C). Since it is necessary for both elongation and shortening responses, Rif1 may play a general sensing/structural/regulatory role, rather than a catalytic one, in the telomeric response to environmental signals. This is consistent with recent studies that found a role for Rif1 in the regulation of chromatin structure and of DNA replication origin firing [26], [27].
Remarkably, rif2Δ cells exhibited a strong response to ethanol (Figure 3A), underscoring the different roles of Rif1 and Rif2 in telomere length maintenance [28]–[32]. We suggest that exposure to ethanol reduces the recruitment of the Rif proteins at the telomere ends, resulting in conditions permissive for indiscriminate telomerase recruitment, elongating both short and long telomeres, and yielding a broad distribution of telomere lengths (Figure 2A). The insensitivity of rif1Δ mutants to ethanol could be due to the importance of Rif1p for the telomere elongation response, and/or the increased binding of Rif2 to telomeres in the absence of Rif1. In agreement with this model, deletion of RIF2 caused over-extension of telomeres in ethanol (Figure 3A); a reduction of Rif1 telomere recruitment by ethanol in the strain deleted for RIF2 mimics a rif1Δ rif2Δ double mutant, which exhibits increased levels of telomere elongation. In contrast to these results, the RIF2 deletion had no effect on the reduction in telomere length upon exposure to caffeine or 37°C (Figure 3B,C).
Mutations in the TEL1 gene, which encodes the yeast ortholog of the mammalian ATM protein kinase, result in very short telomeres. Tel1 regulates the preferential elongation of short telomeres [33] by a pathway that also includes the MRX complex (Mre11, Rad50, Xrs2; [34]). A separate regulatory branch includes the yeast Ku proteins [35]. Figure 3D shows that the tlm mutants with very short telomeres could be clearly separated into two groups: telomeres of mutants of the Tel1 pathway (tel1Δ, mre11Δ, rad50Δ, xrs2Δ) were hyper-responsive, while mutants of the NMD (nonsense mediated decay, nmd2Δ, nam7Δ and upf3Δ) and Ku pathways had only a mild response to ethanol. The fact that telomeres can be elongated by ethanol in the absence of Tel1 or of components of the MRX complex is surprising; notably, the wide size distribution observed upon exposure to ethanol (Figures 1, 2), is consistent with a mechanism independent of the one that preferentially elongates the shortest telomeres, which depends on the Tel1 pathway [16].
The NMD pathway degrades mRNAs carrying nonsense mutations. In addition, it affects the steady state level of hundreds of mRNAs, including those known to act at telomeres (e.g., Est1, Est2, and two components of the CST telomeric capping complex, Stn1 and Ten1 [36]). Mutations in the NMD machinery lead to higher mRNA levels of these proteins and to short telomeres [37]. The NMD pathway has been recently shown to affect the fitness of cdc13-1 and yku70 mutants by controlling the expression of Stn1, an essential telomere capping protein, which interacts with Cdc13 and participates in the recruitment of telomerase [38]. In nmd mutants, the response of telomeres to ethanol stress is reduced relative to wild-type strains, indicating that the NMD pathway is involved in telomere elongation during ethanol stress. We asked if upregulation of Ten1 and Stn1 is involved in this effect by overexpressing these genes in naïve cells and measuring the effect of ethanol on telomere length in these cells (Figure S4). Overexpression of Stn1 reduced the ethanol response and overexpression of both Stn1 and Ten1 completely abolished the telomere length response to ethanol. These results suggest that the level of CST activity, controlled by the NMD pathway, plays an important role in the telomere elongation response to ethanol. This is consistent with the proposed role of the CST complex in telomerase activation. Interestingly, mutations in the CST proteins are lethal when combined with a deletion of RIF1 [28]–[32], indicating the existence of an essential overlapping function between the two telomere regulatory components. The roles of the CST and Rif1 in transducing the ethanol signal to the telomeres will be the subject of future research.
Among the additional mutants with a reduced response to ethanol were doa4Δ, snf7Δ and did4Δ (Figure 3A). DOA4 encodes an enzyme that removes ubiquitin from membrane proteins destined for vacuolar degradation. The Doa4 protein resides in the late endosome, where it interacts with the ESCRT-III machinery, which includes Did4 and Snf7 [39]. A role was previously observed for vacuolar traffic proteins in telomere length maintenance [40]; however, the precise mechanism remains enigmatic. Another mutant that shows apathy towards ethanol is hpr1Δ, defective for a component of the THO complex. Consistent with these results, mutations in HPR1 were recently shown to affect the expression levels of RIF1 [41].
In contrast to these genes, a deletion of HSP104 was hyper-responsive to ethanol. Hsp104 is a stress chaperone that plays an important role in maintaining prion particles in the cell [42]. It is unclear whether its role in telomere length regulation is related to its role in prion maintenance.
Deletion of Rif1 and mutations in Rap1 also significantly decrease the telomeric response to caffeine, indicating that Rif1-Rap1 is not only involved in telomere elongation under ethanol stress, but also in telomere shortening under caffeine. Caffeine is a known inhibitor of phosphatydyl inositol-3 kinase related kinases (PI3K-like kinases) such as human ATR and ATM [43] and their yeast counterparts, Tel1 and Mec1 [44]. Therefore, we tested whether mutations in these target genes would abolish the telomere shortening caused by caffeine. Indeed, deletion of either TEL1 or MEC1 individually does not prevent the response to caffeine, but a double mutant tel1Δ mec1Δ is completely insensitive to the telomeric effect of caffeine (Figure S5), consistent with the known redundant function that these two kinases play in telomere biology [45]. Thus, caffeine causes telomere shortening by inhibiting the ATM/ATR-like regulatory kinases.
Mutations in Rap1 and the deletion of Rif1 affect only the shortening rate in the presence of caffeine but do not affect the response to high temperature. High temperature has a broad, pleiotropic effect, and may alter telomere length via several mechanisms. Several TLM genes that, when mutated, result in short telomeres, are down regulated by high temperature (Table S3). However, no single deletion mutant failed to respond to high temperature by shortening its telomere length, suggesting that there are redundant functions among these responding genes. This result is consistent with a recent study [46] proposing that one or more telomerase components are intrinsically thermolabile.
Accurate telomere length homeostasis is dependent on a large genetic network that includes ∼400 (largely evolutionarily conserved) genes [9]–[11]. Our results show that this network can be disrupted by several environmental signals, and by different regulation mechanisms that lead to altered telomere length. These responses are distinct from the stereotypic responses to stress [18], and seem to be specific only to particular conditions.
Telomere length and telomerase activity are important factors in the pathobiology of human disease. Age-related diseases and premature aging syndromes, for example, are characterized by the shortening of telomeres [47]. Tumor cells, on the other hand, prevent telomere shortening and telomere loss by up-regulating telomerase, thereby perpetuating cells with short telomeres and high chromosomal instability [48]. Thus, although the mechanisms at work differ, changes in telomere length fuel disease pathology in cancer and other premature aging syndromes. While previous studies have identified correlations between telomere length and environmental conditions such as mental stress [49], socioeconomic status [50], and health-related behavior in adults [51], we extend those findings here by demonstrating direct causality between environmental cues and changes in telomere length. This identification of mechanisms by which external signals modify telomere length significantly advances our understanding of the complex interplay of genes and environment. More critically, however, these findings also point a future path to strategic manipulations of telomere length that may well have important therapeutic implications in the treatment of human disease.
All the yeast strains used in this study are derivatives of BY4741 (MATa ura3Δ met15Δ leu2Δ his3Δ), unless otherwise specified. Mutants were obtained from the yeast deletion library [19] or from the DAmP library of hypomorphic alleles [20]. Strains carrying genes with tetracycline-inducible promoters were taken from the library described in [25]. Petite BY4741 derivatives were obtained by plating cells on YEPD plates containing ethidium bromide. Strains deleted for MEC1, TEL1 and SML1 were in the MS71 background [52] (kindly provided by T. Petes).
Telomeric Southern blots were carried out as in [53]. PCR fragments containing telomeric sequences and a genomic region that hybridizes to two size marker bands (2044 and 779 bp) were used as probes. The telomere length was measured with an in-house software (TelQuant) using the size marker bands as reference. Telomere length was ∼1250 bp in wt cells [composed of the sub-telomeric region (∼900 bp) and the telomere repeats (∼350 bp)].
Stress levels were calibrated to reduce growth by 40%–60%. Cells were subjected to the various stresses by serial transfer growth: a single colony of BY4741 was grown in rich medium (YEPD), and 5 µl were used to inoculate 5 ml cultures under the various stress conditions (in triplicates). The cultures were grown ∼10 generations before being diluted (1∶1000) into fresh medium.
We analyzed stress-induced RNA response for caffeine, temperature of 37°C, ethanol and hydrogen peroxide (H2O2), using Affymetrix GeneChip Yeast Genome 2.0 arrays. Transcript levels were measured for three independent cultures grown in the presence of the stress agent, and were compared to a control set comprised of four wild-type measurements. To obtain differentially expressed genes between the stress-induced response and the control measurements, we (i) employed the Robust Multi-array Average (RMA) method for normalization and summarization of the Affymetrix arrays [54]; (ii) filtered probes which had more than half of their detection calls marked as absent; and (iii) employed the Significance Analysis of Microarrays (SAM) [17] with false discovery rate (FDR) below 0.01. Following these procedures, we obtained a set of 1,744, 1,404, 1,670 and 1,019 differentially expressed genes for caffeine, 37°C, ethanol and H2O2, respectively.
We used the un-weighted TLM-based network described in [13], representing the most likely network connecting TLM genes to the telomere maintenance machinery. We next compared the pairwise shortest (unweighted) distances in the network between stress-specific differentially expressed TLM genes and other TLM genes, revealing that stress-specific differentially expressed TLMs for ethanol, caffeine and 37°C lie significantly closer to each other than other TLM genes (Wilcoxon ranked sum test, p<2e−33,p<3e−27 and p<3e−50 for ethanol, caffeine and 37°C stresses, respectively). Reassuringly, the hydrogen peroxide stress showed no significant difference between the two types of TLM genes. Last, using an assembled yeast protein-protein interaction network [13], we verified that stress-specific differentially expressed TLM genes are significantly closer in this network than other stress-specific differentially expressed genes (p<6e−9 for all stresses), verifying that closeness on the network is not a general property of differentially expressed genes.
In an attempt to identify stress-response related genes, we examined the elongation or shortening of the telomere for each knockout gene in the absence or presence of the stress. The elongation/shortening of the telomere in the presence of the stress displayed a linear relation with the initial length of the telomere (Pearson correlation coefficient between the two variables is ρ = −0.77 (p<9e−25), −0.95 (p<2e−38) and 0.36 (p<7e−6) for caffeine, 37°C and ethanol, respectively). In order to detect outliers, we performed a robust linear regression analysis. Following [55], we assumed that the residuals follow a normal distribution and identified the outlier genes as the most extreme 5% (2.5% from each side). The computations were performed using Matlab.
Chromatin immuno-precipitation (ChIP) was carried out by standard methods [56]. The association of Rif1-HA, and Rif2-HA with Y′-element telomeres was detected using Santa Cruz Mouse anti HA monoclonal IgG antibodies (SC-7392). Real-time PCR (RT-PCR) reactions were carried out using the following primers: Y′-element : 5′-GGCTTGATTTGGCAAACGTT-3′, and 5′-GTGAACCGCTACCATCAGCAT-3′. ARO1: 5′-GTCGTTACAAGGTGATGCC-3′, and 5′- CGAAATAGCGGCAACAAC-3′. The relative fold enrichment\depletion of the telomere-associated proteins Rif1 and Rif2 was calculated as follows: [telIP/ARO1IP]/[tel input/ARO1input] [57].
|
10.1371/journal.pntd.0004177 | Population Dynamics of Owned, Free-Roaming Dogs: Implications for Rabies Control | Rabies is a serious yet neglected public health threat in resource-limited communities in Africa, where the virus is maintained in populations of owned, free-roaming domestic dogs. Rabies elimination can be achieved through the mass vaccination of dogs, but maintaining the critical threshold of vaccination coverage for herd immunity in these populations is hampered by their rapid turnover. Knowledge of the population dynamics of free-roaming dog populations can inform effective planning and implementation of mass dog vaccination campaigns to control rabies.
We implemented a health and demographic surveillance system in dogs that monitored the entire owned dog population within a defined geographic area in a community in Mpumalanga Province, South Africa. We quantified demographic rates over a 24-month period, from 1st January 2012 through 1st January 2014, and assessed their implications for rabies control by simulating the decline in vaccination coverage over time. During this period, the population declined by 10%. Annual population growth rates were +18.6% in 2012 and -24.5% in 2013. Crude annual birth rates (per 1,000 dog-years of observation) were 451 in 2012 and 313 in 2013. Crude annual death rates were 406 in 2012 and 568 in 2013. Females suffered a significantly higher mortality rate in 2013 than males (mortality rate ratio [MRR] = 1.54, 95% CI = 1.28–1.85). In the age class 0–3 months, the mortality rate of dogs vaccinated against rabies was significantly lower than that of unvaccinated dogs (2012: MRR = 0.11, 95% CI = 0.05–0.21; 2013: MRR = 0.31, 95% CI = 0.11–0.69). The results of the simulation showed that achieving a 70% vaccination coverage during annual campaigns would maintain coverage above the critical threshold for at least 12 months.
Our findings provide an evidence base for the World Health Organization’s empirically-derived target of 70% vaccination coverage during annual campaigns. Achieving this will be effective even in highly dynamic populations with extremely high growth rates and rapid turnover. This increases confidence in the feasibility of dog rabies elimination in Africa through mass vaccination.
| Rabies is a deadly disease caused by a virus that in Africa is maintained in populations of owned, free-roaming domestic dogs. Rabies can be controlled by mass vaccination, by ensuring that a certain proportion of the dog population is immune to the disease. Maintaining this proportion of immune animals creates herd immunity, reducing the spread of disease even among non-immune individuals, eventually leading to its elimination from the population. Maintaining herd immunity to rabies in free-roaming dog populations can be challenging, particularly in communities that lack regular access to veterinary services. In these communities, mass vaccination is usually implemented in annual campaigns, of relatively short duration. Between campaigns, the proportion of immune individuals in the population declines, often dropping below the critical threshold as vaccinated dogs die and susceptible dogs enter the population through birth or migration. We measured these rates of birth, death and migration in a typical population of free-roaming dogs in South Africa, and showed that vaccinating 70% of the population during annual campaigns would be sufficient to maintain herd immunity to rabies in the period between campaigns. This is achievable even in populations that have high turnover and are growing rapidly—the most challenging circumstances to maintaining herd immunity. These findings increase confidence in the feasibility of eliminating dog rabies from Africa through mass vaccination.
| Rabies is a serious yet neglected public health threat in resource-limited communities in sub-Saharan Africa [1]. In these settings, the virus that causes this deadly disease is largely maintained in populations of free-roaming domestic dogs, and is transmitted to people through bites or other contact with the saliva of infectious rabid dogs. Rabies in dog populations (and consequently in humans) can be controlled and in certain circumstances eliminated through the mass vaccination of dogs against the virus [2]. The control of an infectious disease through vaccination relies on vaccinating a sufficient proportion of the host population to effect herd immunity: the phenomenon whereby the risk of infection among susceptible individuals in a population is reduced by the presence of immune individuals [3]. Thus, if a threshold proportion of individuals in a population are immune, the incidence of infection in that population will decline, eventually to zero [3]. This critical vaccination threshold is a function of the basic reproductive number R0; that is, the number of secondary cases of infection generated by a typical infectious individual in an otherwise fully susceptible population [4]. Hampson et al. [5] estimated R0 for outbreaks of rabies in domestic dog populations around the world. From their estimates (R0 < 2), they calculated the critical vaccination threshold for rabies to be lower than 40% in the populations reviewed. Thus, theory and empirical evidence predicts that outbreaks of rabies in dogs can be controlled if at least 40% of the population is immune at any time. However, achieving this goal in free-roaming dog populations in resource-limited communities is hampered by the rapid turnover of these populations, compounded by the lack of affordable and accessible veterinary services. In these areas, mass dog vaccination against rabies is usually implemented by the state or other agencies in annual or less frequent campaigns, of relatively short duration. Between campaigns, the proportion of immune individuals in the population declines as vaccinated dogs die and susceptible dogs enter the population through birth or migration. To maintain population immunity above the critical threshold in the period between campaigns requires that a larger proportion of the dog population be vaccinated during campaigns [5]. The actual target vaccination coverage to be achieved during campaigns is thus dependent on the demographic rates of the dog population, as well as the interval between campaigns and the duration of vaccine-induced immunity.
The World Health Organization (WHO) recommends that, to achieve control and eventual elimination of dog rabies, programmes must ensure that mass dog vaccination campaigns achieve a vaccination coverage of at least 70% of the population in a given area, and that such campaigns recur, usually annually [6]. The figure of 70% is an empirically-derived consensus, stemming from work on the control of dog rabies in New York State during the 1940s [7]. It is assumed that this coverage, achieved during a campaign of relatively short duration, is sufficient to maintain the population immunity above the critical threshold for at least 12 months, despite dog population turnover due to births, deaths and migrations during this period [6]. To date, little work has been done to test this assumption using real data on demographic rates from free-roaming dog populations in rabies-affected communities. Using demographic characteristics for a cohort of owned dogs over a 12-month period in northwest Tanzania, retrospectively collected through household questionnaires, Hampson et al. [5] estimate that a target vaccination coverage of 60% is sufficient to avoid coverage falling below the critical threshold of 40% between annual campaigns. Using prospective cohort studies in four populations of owned, free-roaming dogs in South African and Bali, Indonesia, Morters et al. [8] estimate that 60–70% coverage is sufficient for the same purpose.
Knowledge of the population dynamics of free-roaming dog populations, particularly the core demographic rates of birth, death and migration, may therefore help to inform effective planning and implementation of mass dog vaccination campaigns to control rabies in resource-limited communities, and to design strategies for the eventual elimination of dog rabies and associated human deaths. Knowledge of these rates, and their interplay with population vaccination coverage levels, may also improve understanding of the possible contribution of humane dog population management to rabies control efforts [9, 10]. Despite the ubiquity of free-roaming dogs in sub-Saharan Africa, little is known about the demographic rates of these populations, or the factors that affect them. Studies that have provided estimates of demographic parameters of these populations have largely relied on cross-sectional household questionnaire surveys of dog owners [5, 11–13]. These analyses make certain assumptions which may not hold true, such as stable age- and sex-distributions or consistency of demographic rates over time. In addition, human-mediated migration of domestic dogs may play an important role in population dynamics and rabies epidemiology [11, 12, 14–18], yet few studies have examined the contribution of migration to population turnover.
A number of studies of free-roaming dog populations in communities in Africa have revealed that, despite appearances, there is little evidence for the presence of large numbers of unowned dogs in these populations [11, 12, 16–19]. Adequate demographic surveillance of dog populations is therefore possible through on-going monitoring of owned dogs within households [8]. Here, we report data from a demographic surveillance system covering all owned dogs in a rabies-affected, resource-limited community in South Africa. Data span a 24-month period, from 1st January 2012 through 1st January 2014. The aim of the study was to quantify demographic parameters in this population of dogs, particularly the core demographic rates of births, deaths and migrations, and to assess the implications of dog population dynamics for rabies control through mass vaccination.
The data collection method used in our study follows the model of health and demographic systems in human public health. A health and demographic system (HDSS) monitors all individuals, households and residential units in a defined geographic area, known as a demographic surveillance area (DSA) (www.indepth-network.org). HDSSs are used in the field of public health, to meet the need for reliable population-based data on health in many low- and middle-income countries where there is limited registration of vital events, including births, deaths (by age and sex), and medical causes of death [20]. The overall objective of these HDSS sites is to establish a reliable information base to help policy-makers set health priorities and allocate resources more efficiently. Following an initial census of the defined population in which all residential units and occupants are enumerated, longitudinal measurement of demographic and health variables is undertaken through repeated visits at regular intervals to all residential units within the DSA. We applied this model to create a health and demographic surveillance system in dogs (HDSS-Dogs) in a population of owned, largely free-roaming dogs in a resource-limited community in South Africa. The aim of the HDSS-Dogs is to provide long-term, reliable, population-based data for evidence-based approaches to the control and elimination of dog rabies, and for humane dog population management in resource-limited communities.
The DSA of the HDSS-Dogs (Fig 1) was arbitrarily defined prior to the start of the study, making use of natural and man-made features recognisable on the ground by field teams. The DSA encompasses Hluvukani settlement (S 24°39’; E 31°20’), which includes parts of two administrative areas (Eglington and Clare A villages) in Bushbuckridge Local Municipality, Mpumalanga Province, South Africa. The total human population of Bushbuckridge Municipality in 2011 was 541,249, with a growth rate of 0.79% from 2001–2011 [21]. Over two thirds of the population aged 20 years and older had not completed secondary school. The unemployment rate in Bushbuckridge in 2011 was 52.6%, substantially higher than the national rate of 29.8%. The majority of the population (96%) lives in formal housing, with a mean household size of 4.0 persons. One fifth of households do not have access to piped water, and fewer than 10% to refuse removal services. The mean annual household income in 2011 was ZAR 36,569 (about US$ 5,000), less than half of the mean annual income for the province as a whole [21].
In Hluvukani, families live in houses on separate stands (a stand is a plot or parcel of land). Stands are permanently identified by municipal stand numbers that are unique within each administrative area. Stands without municipal numbers were assigned a unique number by the study team. All stands in the DSA are georeferenced as part of the study. There are no private veterinary services in the study area. The provincial state veterinary services have strengthened regular dog rabies vaccination campaigns since the disease re-emerged in Bushbuckridge in 2008, after a long period of apparent absence [22]. In addition to the veterinary services provided by the state, the animal health needs of the community in Hluvukani and the surrounding areas are also met at a subsidised rate by the Hluvukani Animal Clinic, located in the centre of Hluvukani and run by the Faculty of Veterinary Science of the University of Pretoria (UP). The majority of dogs in the community have the morphological appearance of the Africanis landrace [23], although there is some phenotypic evidence of interbreeding with western breeds.
An initial census of the dog population was conducted from July through October, 2011. This census was combined with a house-to-house rabies vaccination campaign in conjunction with the provincial veterinary services and the Faculty of Veterinary Science, UP. To uniquely and permanently identify individual dogs, microchips (BackHome BioTec, Virbac RSA) were subcutaneously implanted into dogs present at the start of the study, and into those dogs that entered the population during the study period. Dogs that could not be handled to implant a microchip were assigned a unique identification code. Dogs were also identified by name and appearance. All dogs enrolled in the study were photographed. Following the census (Round 1), five follow-up rounds (Rounds 2–6) were conducted from December 2011 through May 2014, resulting in all households being visited approximately every six months during this period. All households in the DSA were visited during each round. All owned dogs were recorded at each visit, as were the demographic events that occurred in the period between visits, including births, deaths, and migrations into and out of households. Data including sex, age and rabies vaccination status were collected during the census for all dogs and at each round for new dogs. For vaccination history, we used owners’ reports. Although vaccination certificates are issued by the veterinary services, not all owners consistently keep such certificates. Vaccination data of the veterinary services are aggregated at a local administrative level, and are not always readily available for individual dogs. Dogs vaccinated at any point in the preceding 36 months were considered vaccinated [24, 25]. Variables with time units (e.g. age, or time since entry or exit of a dog) were estimated by the owners. Owners were asked to give a lower and upper estimate, in the time unit of their choice (days, weeks, months, years), reflecting the precision of their estimate. The date of the event was assigned as the midpoint of the estimated range, and the range of the estimate converted to days. All ‘residence episodes’ of individual dogs in households were tracked and aggregated to produce the denominator of dog-time in the population. Residence episodes within households begin with birth or in-migration (e.g. purchase or receipt of a new dog), and terminate with death or out-migration (e.g. sale or gifting of dog to another household). Data from 1st January 2012 through 1st January 2014 are presented here. Data are presented in 3-month periods (annual quarters, abbreviated Q). Point data are provided for the first day of each quarter.
Mortality rates were determined as the total number of deaths in a defined population during a specified period, divided by the total number of dog-years lived in the same sub-population over the same period. The subpopulations were characterized by three variables: sex (male or female), age class (0–3 months, 4–11 months, 12–23 months, 24–35 months, ≥36 months), and rabies vaccination status (unvaccinated or vaccinated). Time periods were 2012 and 2013. If dogs changed subpopulations over time, they were split in different records; each record is independent as it describes the dog in one or other specific subpopulation. To model mortality rates in the subpopulations across the two periods, we fitted a Poisson regression model in R [26]. For each record, the number of dog-days was calculated and the log (dog-days) included in the regression as an offset. The four variables (sex, age class, vaccination status and year) as well as all possible interactions (two-, three- and four-way) were included in a maximal model. Model simplification was done by stepwise removal of non-significant variables from the maximal model, starting with the highest-order interactions, and examining the resulting change in deviance of the model. Variables whose removal did not result in a significant change in deviance (p>0.05) were dropped from the model. Mortality rate ratios (MRR) and confidence intervals were calculated from the final minimal adequate model.
To assess whether a 70% vaccination coverage achieved during annual campaigns is sufficient to maintain population immunity above the critical threshold of 40% for a 12-month period, we simulated two hypothetical vaccination campaigns, one on 1st January 2012 and another on 1st January 2013. We randomly assigned a positive vaccination status to 70% of all dogs present in the population on those dates. The number of these vaccinated dogs still present in the population 12 months later was divided by the total number of dogs in the population on that date, to give the vaccination coverage one year later. The process of random assignment was repeated 1,000 times to produce Monte Carlo estimates of vaccination coverage. Similarly, we assessed the minimum proportion of dogs to be vaccinated during campaigns on the 1st January 2012 and on the 1st January 2013 required to ensure a 40% coverage in the population 12 months after the respective dates.
The study was approved by the University of Pretoria Animal Ethics Committee (protocol no. V033-11). Written informed consent was obtained from dog owners to participate in the study. The protocol adhered to the specifications in the South African National Standard (SANS 10386–2008): “The Care and Use of Animals for Scientific Purposes”.
The area of the DSA is 10.4 km2. There are a total of 2,373 stands in the DSA, including 68 empty or non-residential stands (Fig 1). Stands occupy an area of 4.6 km2. The total number of households in the DSA recorded during Round 6 was 2,116. A number of households occupy more than one stand. As residence episodes of households within the DSA were not tracked for this study, we present household-level data for Round 6 only. The number of occupants in these households was 9,652, with a mean household size of 4.6 (range: 1–27). The mean number of dogs per household in Round 6 was 0.36 (range 0–9), and the number of dogs per 100 people in the DSA was 7.9. The percentage of dog-owning households (DOHHs) was 17%. The proportion of surgically-sterilized dogs in the population was very low (1–1.5%). We recorded the number of new households (taken occupancy within the previous 12 months) in the DSA during Round 6. Of the 2,116 households, only eight were new, with missing data from a further eleven households.
Table 1 shows the demographic characteristics of the owned dog population present in the DSA on the first day of each quarter, from 1st January 2012 to 1st January 2014. The population of owned dogs declined by 10% over the period of the study, but this overall decline masks a substantial fluctuation (Fig 2). Annual population growth rate (as a percentage of the population at the start of the period) was +18.6% in 2012 and -24.5% in 2013. The total number of dog-years in the population was 915 in 2012 and 821 in 2013. Crude annual birth rates were 451 puppies born per 1,000 dog-years of observation (dyo) in 2012 and 313 per 1,000 dyo in 2013. Crude annual death rates were 406 per 1,000 dyo in 2012 and 568 per 1,000 dyo in 2013. Crude birth and death rates by quarter are shown in Fig 2. The rate of natural increase of the population (birth rate minus death rate) was +4.5% in 2012 and -25.5% in 2013. The net migration rate, measured as the total number of in-migrations to households minus the total number of out-migrations from households, was 12.3% in 2012 and -2.1% in 2013.
Data were recorded for 1,093 household entry events and 1,173 exit events (Fig 3). Most dogs entered households through birth (61%) or as gifts (31%), and exited through death (71%) or being given away (21%). Owner-reported causes of death by quarter are shown in Fig 4. Of the exits and entries, a small proportion were dogs that were bought and sold. Over the course of the study, 63 dogs were purchased while only 5 were sold, suggesting that the majority of dogs purchased were from outside the study area. The median purchase price for dogs was ZAR 30, about US$ 5 (n = 60, range: ZAR 5 to ZAR 4,500). The sex ratio of dogs migrating in to the population was skewed towards males (1.79 males per female; data on characteristics of external migrants was collected in Rounds 5 & 6 only). During this period, the sex ratio of dogs entering households from outside the study area (external in-migrants, n = 25) did not differ significantly from those of dogs entering households from within the study area (internal in-migrants, n = 184) (1.79 vs. 1.78; Fisher’s exact test p-value = 1).
Mortality rates by sex, age group and vaccination status are shown in Table 2. There was no evidence of overdispersion of the regression model (residual deviance 4453 on 4590 degrees of freedom; goodness-of-fit test p-value = 0.9). Reducing the number of levels for the age class variable from five to three (0–3 months, 4–11 months and ≥12 months) did not cause a significant increase in deviance of the model (p = 0.09), and simplified interpretation of the model outputs; age class was therefore reduced to three levels. The initial model revealed a significant three-way interaction between sex, age class and year. To aid interpretation of this interaction, we split the data into two sets by year (2012 and 2013) and modelled these separately. The adjusted mortality rate ratios (MRRs) are presented in Table 3. In 2012, there was no significant difference in mortality rates between the sexes, but in 2013, females suffered a significantly higher mortality rate (MRR = 1.54, 95% CI = 1.28–1.85). Sex-specific mortality rates by quarter are shown in S1 Fig. In both 2012 and 2013, there was a significant two-way interaction between age class and vaccination status (Tables 3 and 4). Among unvaccinated dogs, mortality rates were significantly lower in the 4–11 months and ≥12 months age classes when compared to the 0–3 months age class across both years, but this effect of age on mortality rates was not seen among vaccinated dogs in either year (Table 4). In the age class 0–3 months, the mortality rate of vaccinated dogs was significantly lower than that of unvaccinated dogs (2012: MRR = 0.11, 95% CI = 0.05–0.21; 2013: MRR = 0.31, 95% CI = 0.11–0.69).
Vaccination coverage (based on owner-reported vaccination history of 520 dogs for which this information was available) was 33% before the start of the house-to-house vaccination campaign in Round 1. After the campaign, this increased to 78%. Based on owner reports of individual dog vaccination history (and assuming a duration of protection of three years), vaccination coverage remained well above the threshold level of 40% until a second vaccination campaign in 2013 (Fig 5a).
The results of the simulation of vaccination campaigns reaching 70% of the dog population on the 1st January 2012 and the 1st January 2013 are shown in Fig 5b and 5c. Despite the high turnover and substantial growth of the population in 2012, the simulated coverage remained above the threshold of 40% for that year. The decline in the population in 2013 slowed the rate of decrease of vaccination coverage from the 2012 campaign, such that coverage only dropped below the threshold level around 18 months after the campaign (Fig 5b). A second simulated campaign in January 2013 kept coverage well above the threshold for that year (Fig 5c). To ensure a 40% coverage in the population 12 months after vaccination, the minimum vaccination coverage needed for the campaigns was 61% in 2012 and 52% in 2013.
The above simulation assumes that vaccination coverage is randomly distributed in the dog population. We tested for heterogeneity in actual vaccination coverage across sexes and age groups, using owner-reported vaccination data as on 1st January 2012 (coverage = 66%). We found no association between sex and vaccination status (Χ2 = 1.02, d.f. = 1, p-value = 0.31), but a strong association between age and vaccination status (Χ2 = 71.88, d.f. = 4, p-value < 0.0001), with significantly fewer dogs vaccinated in the 0–3 and 4–11 month age groups than in the oldest age group (≥ 12 months).
We studied the dynamics of an owned, free-roaming dog population over a period of 24 months. We show that this is a highly dynamic population, with rapid turnover and significant heterogeneity in demographic rates over time and across segments of the population. Despite this, routinely achieving 70% vaccination coverage during mass dog vaccination campaigns conducted every 12 months will be sufficient to maintain coverage above the critical threshold of 40%, even during periods of rapid growth and high turnover.
The population declined by 10% over the course of the study. Previous estimates of growth rates of owned dog populations in sub-Saharan Africa predict steady high growth rates of between 7–10% [5, 11–13], but these estimates, derived from retrospective data based on owner recall and collected during household surveys, may not capture the highly dynamic nature of these populations. Moreover, methods used to derive demographic rates from retrospective data collected during cross-sectional surveys may rely on assumptions whose validity is questioned by our findings, such as stable age- and sex-distributions or consistency of demographic rates over time. By contrast, Morters et al. [8], who undertook a longitudinal study of two populations of owned dogs in resource-limited communities in Johannesburg, South Africa using methodologies similar to ours, observed no population growth in one site and an overall decline in the other site, over a three-year period. Together, these longitudinal studies demonstrate that, while owned dog populations are certainly capable of rapid growth, sustained growth in a given area is not a general phenomenon.
The mortality rate in this population of dogs was very high over the study period, particularly in puppies aged 0–3 months. Hampson et al. [5] recorded a similarly high mortality rate (450 deaths per 1,000 population) in dogs older than 3 months in owned, free-roaming dogs in northwest Tanzania. By comparison, the crude death rate of dogs in U.S. households in 1996 was estimated as 79 per 1,000 population [27], and 39 per 1,000 dog-years in insured Swedish dogs from 1995 to 2000 ([28]; this figure excludes puppies and dogs older than 10 years). High mortality rates in the first year of life in free-roaming dogs in resource-limited communities have been reported elsewhere ([11–13, 29]). Over half of all deaths in our study were reported by owners to be due to disease or parasites, while 21% were due to unknown causes. Only a relatively low proportion of deaths (7%) were reportedly caused deliberately by owners or others; however, as Morters et al. (2014) argue, killing of unwanted dogs by owners may be underreported. If it is accepted that human demand rather than environmental resources determines the ‘carrying capacity’ of owned dogs in a given area (as demonstrated by [8]), it can be hypothesized that, in the absence of affordable options for humane population management, killing of dogs will increase as the population grows and demand becomes satiated. Continued observation of the study population through a growth phase may shed further light on this and other mechanisms that regulate population growth.
Mortality rates by age class were highest in the 0–3 month group, and significantly lower in subsequent age classes. Notably, rabies vaccination removed the effect of age on mortality rates, due to its association with significantly reduced mortality rates in the 0–3 month age group. Plausible explanations for this association include i) specific protective immunological effects of the vaccine against rabies in this age group, ii) nonspecific protective immunological effects of the vaccine against other infections (heterologous immunity), and iii) confounding effect of other interventions by owners who have their puppies vaccinated against rabies, compared with those who don’t. It is highly unlikely that the reduction in mortality is due to the specific protective effect of rabies vaccination, given the low incidence of the disease in the population during the periods in question (2012/2013), particularly among the age group concerned [30]. Another explanation might be that dogs that are vaccinated against rabies are simultaneously vaccinated against other infections such as canine distemper or parvovirus, or are more likely to receive therapeutic interventions such as endo- or ectoparasite treatment. The reduction in mortality in the rabies-vaccinated puppies may therefore be as a result of the specific protective or therapeutic effects of associated interventions. Although the overall level of vaccination against diseases other than rabies is assumed to be extremely low, it remains plausible that some owners of young puppies may seek veterinary care at Hluvukani Animal Clinic. Such care might include vaccinations and other healthcare interventions that could significantly increase the survival rates of these puppies. However, cursory examination of clinic records and discussions with the clinician in charge suggest that veterinary health-seeking behaviour among the population within the DSA is not sufficiently advanced to account for this as an explanation. During vaccination campaigns, the local state veterinary services only administer rabies vaccine to dogs (including those in the 0–3 month age class), and no other routine vaccination or healthcare intervention is given. This leaves the intriguing possibility that the reduction in mortality in the 0–3 month age group associated with rabies vaccination is due to a non-specific protective effect of the vaccine. There is strong evidence from human HDSS sites that vaccines have substantial nonspecific effects in children in high-mortality regions [31]; for example, in randomized trials tuberculosis and measles vaccines are associated with a substantial reduction in child mortality, which cannot be explained by prevention of the target disease [32]. Further studies in dogs (observational studies in which the details and timings of vaccinations and other healthcare interventions are carefully recorded, or randomised controlled trials), are needed to determine if rabies vaccine does in fact induce a protective nonspecific immune response sufficient to reduce puppy mortality. Such a finding would have implications for the use of rabies vaccine in this age group (particularly in light of a recent field study showing that rabies vaccine is effective in this age group [33]).
The population sex ratio was strongly skewed towards males (around 1.4 males per female dog in 2012), and became increasingly so during 2013 (increasing from 1.5 to 1.8 males per female). Male-skewed sex ratios are a consistent feature reported in other studies of the demographics of owned dog populations in sub-Saharan Africa [5, 8, 11–13, 19, 34, 35]. This is attributed to a preference of owners for male dogs for guarding of households and livestock, and the reduced nuisance factor of males compared to adult females (e.g. oestrus behaviour, unwanted puppies); however, few studies have directly examined the demographic mechanisms that give rise to this male skew. Preference for male dogs in an owned population implies either decreased retention of female dogs (through higher mortality rates and/or out-migration rates) or increased recruitment of male dogs (through higher in-migration rates or male-skewed birth rates). While Kitala et al. [13] found uniformly lower survival rates for females, other studies in male-skewed populations found no significant differences in survival rates between the sexes [5, 12]. Our study shows that sex-specific mortality rates vary over time, with a significantly higher mortality rate in females compared to males in 2013. This may explain the increase in the sex ratio during that year. (Although the sex ratio of external in-migrants was also male-skewed, the number of external in-migrants was too small to affect the overall population sex ratio). These results highlight the need for longitudinal studies of dog demographics, as contributory factors to population structure will change over time. A conspicuous feature of the mortality rates in this population is the spike in mortality in the second quarter of 2013 (Q6; Fig 2). Although not definitively determined, observations by the authors suggest that a distemper epidemic occurred in the dog population at this time (a retrospective serological study is underway to investigate this hypothesis). Females appear to have suffered a disproportionately greater increase in mortality rates during this period (S1 Fig).
The birth rate in this population was very high (313–451 per 1,000 dog-years), although not sufficient to compensate for the high mortality rates during the latter part of the study period, resulting in the overall natural decline in the population. Hampson et al. [5] report a similar high annual birth rate (530 dogs born per 1,000) for a population of owned, free-roaming dogs in northwest Tanzania. By contrast, New et al. [27] estimated a crude birth rate four times lower (114 dogs born per 1,000) for owned dogs in the United States. Other studies provide proxy measures for birth rates in free-roaming dog populations. Reece et al. [36] report that 48% of roaming adult females became pregnant in any given year from 1995 through 2006 in Jaipur, India. Kitala et al [13] recorded 249 puppies born to a cohort of 305 dogs (128 females and 192 males) over a one-year period in Machakos District, Kenya in 1992–1993, equating to an annual birth rate of 816 dogs born per 1,000 population.
The dynamics of this dog population are strongly seasonal, driven by seasonality in birth rates (peaking April–June) and subsequent mortalities and migration of puppies. Seasonality of breeding in dog populations is not consistently reported. Reece et al. [36] and Totton et al. [37] report seasonal oestrus and pregnancy in two populations of free-roaming dogs in India, while Butler and Bingham [11] deduced a peak in births in June–August in Zimbabwe, neighbouring South Africa to the north. Conversely, Morters et al [8] reported no significant difference in the proportion of dogs pregnant by month in their study populations in Johannesburg, some 350 km south-west of our study area. Seasonality of reproduction, combined with differences in mortality rates across segments of the population, may have implications for the cost-effectiveness of vaccination campaigns conducted at different time periods.
Our simulations show that, despite the highly dynamic nature of this population, achieving 61% vaccination coverage during an annual campaign of relatively short duration will be sufficient to maintain coverage above the critical threshold of 40%, even in the face of rapid growth and high turnover, as was the case in 2012. This is consistent with the predictions of Hampson et al. [5] and Morters et al. [8] from populations of owned, free-roaming dogs in resource-limited communities elsewhere. The predictions from our simulation are conservative, in that they assume that all in-migrating dogs are unvaccinated, and that no supplementary vaccination occurs in the period between campaigns. This may explain the higher-than-predicted estimates seen in the owner-reported vaccination coverage (Fig 5a). Furthermore, the simulation assumed random vaccination of 70% of the dog population present. Because proportionately more puppies were considered vaccinated in our simulation than in reality, and because of the disproportionately higher mortality rate in this age group, our estimates should again be seen as conservative compared to the real-world situation in which a higher proportion of the vaccinated dogs fall within older, more stable age groups. Furthermore, the simulation does not take account of the unanticipated finding in our study, that vaccinated puppies have a significantly lower mortality rate than unvaccinated puppies; this also makes our predictions more conservative. Future refinements of the simulation should take account of heterogeneity in vaccination coverage and demographic rates (particularly mortality rates) across segments of the dog population. Although allowance should be made for the fact that not all dogs who receive the vaccine will develop a protective immune response, the proportion of non-responders is likely small: recent field studies have shown that the vast majority of dogs (>90%) seroconvert to the vaccine, regardless of health status [38].
Overall, the findings of our study are consistent with WHO recommendations that, to achieve control and eventual elimination of dog rabies, programmes must ensure that mass dog vaccination campaigns achieve a vaccination coverage of at least 70% of the population in a given area, and that such campaigns recur, usually annually [6]. Meeting this target ideally requires an accurate estimate of the total dog population in a given area. Obtaining such an estimate is complicated by the highly dynamic nature of dog populations, as evidenced by this study. Estimates based on the number of dogs per household, or per 100 people, should therefore only be used to obtain a rough estimate of dog numbers for planning purposes. Novel, low-cost, robust methods are needed to provide accurate estimates of dog numbers. One such approach may be to engage community members to complete a census of owned dogs immediately ahead of a vaccination campaign. Such an exercise could also be used to raise awareness among dog owners of the upcoming campaign. This approach would be in line with that of the community-directed interventions that have proven successful in the control of other neglected tropical diseases, in which health interventions are undertaken at the community level under the direction of the community itself [39]. This approach could be extended to include planning of the vaccination campaign itself by communities, in partnership with local veterinary services, to ensure maximum vaccination coverage. The incorporation of a simple dog census/household survey into the dog rabies vaccination campaign, as practised in the study area, also offers a direct assessment of the achieved coverage.
Our results show that effective rabies control is possible without adjunct dog population control measures, such as fertility control though sterilization or contraception [10, 40]. One potential benefit of adjunct sterilisation programmes to rabies control could be to reduce population turnover rates and so help sustain vaccination coverage between campaigns, possibly extending the period between campaigns. Extending the period between campaigns (for example, from 12 to 24 months) could result in significant savings in operational costs and reduced ‘vaccinator fatigue’ (a major factor in reduced campaign effectiveness over time; [41]), thereby improving the long-term cost-effectiveness of rabies eradication programmes; however, further cost-benefit studies are needed to weigh this up against the increased cost and time needed for sterilization programmes. Furthermore, the assumption that reducing birth rates through sterilization programmes will result in lower turnover in owned dog populations must be carefully examined. If mortality rates remain high in the face of lowered birth rates, demand for dogs may soon exceed supply, and new dogs may be sourced from outside the population. If not vaccinated, these dogs will contribute to the turnover of the population and offset the benefits of reduced recruitment of unvaccinated puppies. In addition, increased human-mediated in-migration of dogs may foreseeably increase the rate of incursion of rabies [15, 42, 43], complicating eradication efforts. The focus of adjunct population management measures should be to help create stable, healthy, vaccinated populations of dogs; this may require the identification and inclusion of cost-effective interventions to reduce mortality rates as well as birth rates. Such efforts should not detract from the primary goal for rabies control, which is to achieve at least 70% vaccination coverage of the dog population during campaigns.
In conclusion, we emphasise that mass dog vaccination campaigns which reach 70% of the population will be effective in bringing rabies under control and can contribute to rabies elimination, even in populations undergoing extremely high growth rates and rapid turnover. The results of this study demonstrate that demographic surveillance of an entire owned, free-roaming dog population in a resource-limited community in a rabies-affected area is feasible and provides reliable, accurate data that are needed for decision-making. In much the same way that the INDEPTH network has provided reliable population-based data on human health in low-resourced areas [20], we feel there is value in establishing a network of health and demographic surveillance sites where similar methodologies are applied in owned dog populations in resource-limited communities, to provide a platform for evidence-based policies for rabies control and humane dog population management.
|
10.1371/journal.pgen.1004328 | Epistatically Interacting Substitutions Are Enriched during Adaptive Protein Evolution | Most experimental studies of epistasis in evolution have focused on adaptive changes—but adaptation accounts for only a portion of total evolutionary change. Are the patterns of epistasis during adaptation representative of evolution more broadly? We address this question by examining a pair of protein homologs, of which only one is subject to a well-defined pressure for adaptive change. Specifically, we compare the nucleoproteins from human and swine influenza. Human influenza is under continual selection to evade recognition by acquired immune memory, while swine influenza experiences less such selection due to the fact that pigs are less likely to be infected with influenza repeatedly in a lifetime. Mutations in some types of immune epitopes are therefore much more strongly adaptive to human than swine influenza—here we focus on epitopes targeted by human cytotoxic T lymphocytes. The nucleoproteins of human and swine influenza possess nearly identical numbers of such epitopes. However, mutations in these epitopes are fixed significantly more frequently in human than in swine influenza, presumably because these epitope mutations are adaptive only to human influenza. Experimentally, we find that epistatically constrained mutations are fixed only in the adaptively evolving human influenza lineage, where they occur at sites that are enriched in epitopes. Overall, our results demonstrate that epistatically interacting substitutions are enriched during adaptation, suggesting that the prevalence of epistasis is dependent on the underlying evolutionary forces at play.
| Mutations can fix during evolution for two reasons: they can be beneficial and fix for adaptive reasons, or they can be neutral or deleterious and fix solely by chance. Most studies focus on adaptation, where the evolving population is increasing in fitness due to a new selection pressure. Such studies have found an important evolutionary role for epistasis, the phenomenon where the effect of one mutation depends on another mutation. But adaptation only accounts for a fraction of overall evolutionary change. Here we investigate whether epistasis is as common during non-adaptive as adaptive evolution. We do this by comparing the same protein from human and swine influenza. Human influenza is constantly adapting to escape from the immunity that people acquire from previous influenza infections. But swine influenza is under less pressure to escape from acquired immunity since pigs have shorter lifetimes and are less likely to be infected with influenza multiple times. We find that epistasis is less common during the evolution of the swine influenza protein than its human influenza counterpart. Overall, our results suggest that mutations that interact via epistasis are more likely to fix during adaptive evolution.
| Epistasis occurs when the effect of a change at one site in a genome depends on the presence or absence of a change at another site. Understanding epistasis is of profound importance in evolutionary biology, as epistasis can constrain evolutionary pathways and shape patterns of sequence change. As a result, epistasis has been extensively studied at an experimental level. Nearly all of these studies have focused on adaptive evolution, where the population is undergoing changes that improve its fitness in response to some new selection pressure. Examples include bacterial adaptation to new environmental conditions [1]–[3], the acquisition of drug resistance [4]–[7], and changes in enzyme activity or specificity [8]–[10]. These studies have almost universally emphasized a crucial role for epistasis in adaptive evolution.
But adaptive evolution accounts for only a portion of total evolutionary change, which can also be driven by stochastic forces such as genetic hitchhiking and drift [11]–[15]. In many cases, these stochastic forces probably drive a greater fraction of overall sequence change than does adaptive evolution [13]–[17]. Do insights about epistasis from studies of adaptive evolution also apply to evolutionary change by non-adaptive forces?
There are reasons to suspect that epistatically interacting substitutions may be more prevalent in adaptive than non-adaptive evolution. Two main mechanisms have been identified for the fixation of epistatically interacting mutations during adaptive evolution: compensatory mutations and permissive mutations. In the compensatory-mutation mechanism, selection favors an initial mutation that confers an overall adaptive benefit but also creates secondary defects, which are then remedied by a subsequent compensatory mutation. An example is the evolution of broad-spectrum antibiotic resistance, where an initial mutation that confers resistance to a new antibiotic but impairs protein stability is followed by a compensatory mutation that restores stability [5], [18], [19]. In this compensatory-mutation mechanism, both epistatic mutations are immediately beneficial.
In the permissive-mutation mechanism, an initially neutral or mildly deleterious [20] mutation that rises in frequency due to stochastic forces is essential for permitting the subsequent than adaptive mutation. An example is the evolution of steroid-receptor specificity, where initial neutral mutations modulate protein conformational stability in a way that permits subsequent adaptive mutations to alter specificity [8]. In this permissive-mutation mechanism, only the subsequent adaptive mutations are directly favored by selection – but selection for the adaptive mutations indirectly favors linked permissive mutations, leading to expansion of lineages carrying the combination of mutations and increasing their rate of fixation [21].
Crucially, in both the compensatory-mutation and the permissive-mutation mechanisms described above, adaptive evolution is ultimately responsible for driving fixation of the epistatic mutations. It is possible to imagine scenarios for the fixation of epistatic mutations by stochastic forces in the absence of adaptation – but it is not immediately obvious whether epistatic mutations would fix as commonly in the absence of a driving selective force. This idea that the frequency of epistatically interacting substitutions might differ between adaptive and non-adaptive evolution would be consistent with theoretical work suggesting that patterns of epistasis depends on the selective forces at play [22], [23].
Here we examine whether epistasis is more common during adaptive evolution by comparing a pair of protein homologs of which only one is subject to a known selection pressure for adaptation. Specifically, we compare nucleoprotein (NP) homologs from human and swine influenza. In both of these influenza lineages, NP has a highly conserved and essential function in the packaging and transcription of viral RNA, and this function is under strong stabilizing selection [24], [25].
Because human influenza circulates in a population of long-lived hosts that are infected with influenza repeatedly during their lifetimes, human influenza is also under constant diversifying selection for adaptive mutations that escape immune memory that accumulates in the host population [26]–[29]. A major way in which human immune memory targets NP is via cytotoxic T lymphocytes (CTLs), and mutations in CTL epitopes are therefore of adaptive value to human influenza [30]–[33]. We have previously shown that the evolution of NP from human influenza involves the fixation of mutations involved in strong epistatic interactions, and that these epistatic mutations occur in epitopes targeted by CTLs [34]. This prior work hints at an association between epistasis and adaptation.
To systematically test the hypothesis that epistasis is enriched during adaptation, here we compare human influenza NP with its swine influenza homolog. Swine influenza is not targeted by human CTLs (CTL epitopes are highly species specific [35], [36]) – so mutations in human CTL epitopes are not of any special significance to swine influenza. Furthermore, swine influenza is unlikely to be under strong diversifying selection even from swine CTLs. In contrast to human influenza, swine influenza circulates in a population of short-lived hosts that have much less opportunity to acquire anti-influenza immune memory before they are slaughtered [37]. As a result, swine influenza is under less pressure to escape from host immune memory. For example, the HA of classical swine influenza underwent minimal antigenic change from 1918 through the late 1990s [37]–[42] – a timeframe during which human influenza HA underwent extremely extensive antigenic change [43], [44]. Although reassortment events and swine vaccination may have recently somewhat increased antigenic change [38]–[40], overall antigenic change in swine influenza is clearly far less than in human influenza [40], [43], [44].
For this reason, the NPs from swine and human influenza represent an ideal pair of homologs for comparative studies of how adaptation affects patterns of epistasis during evolution. While both NPs are under strong stabilizing selection to maintain their essential and conserved biochemical functions [24], [25], only NP from human influenza is under substantial diversifying selection to change sequence epitopes recognized by CTLs. Comparison of the evolution of NPs from these two influenza lineages therefore provides a naturally occurring case study of how ongoing adaptation affects evolutionary patterns.
In the work described below, we first infer evolutionary trajectories for human and swine NP homologs. We then comprehensively mine existing experimental data to define sites in both NP homologs that are targeted by human CTLs. We show that the human NP homolog exhibits an increased frequency of substitutions in these sites relative to the swine NP homolog, a finding consistent with the expectation that mutations to these sites are adaptive only to human influenza. We then experimentally show that the swine NP homolog lacks the type of epistatic mutations that are fixed in the adaptively evolving human NP homolog. Finally, we use our comprehensive analysis of human CTL epitopes to systematically verify that epistatic interactions within the human NP homolog occur at sites that are targeted by CTLs, where mutations are of adaptive value. Overall, these results demonstrate that during NP evolution, epistatically interacting substitutions are enriched during adaptation.
We set out to compare the evolution of NP homologs from human and swine influenza. Figure 1 shows a phylogenetic tree of NP from human and swine influenza lineages that derive this gene from a common ancestor closely related to the viruses that caused concurrent human and swine pandemics in 1918 [45], [46]. The NP genes of the human influenza lineages in Figure 1 have circulated exclusively in humans since 1918 [45], [46], while the NP genes of the swine influenza lineages in Figure 1 have circulated exclusively in swine since 1918 [38], [47].
Upon transfer into a new host, influenza undergoes a process of adaptation to the ecology, physiology, cell biology and innate immunology of the new host [48]. Because the details of this host adaptation are incompletely understood, we confined our studies to NP homologs that had already been circulating in their respective hosts for several decades. Our expectation is that during these decades of host-specific evolution, the NP homologs will have become highly adapted to the genetically encoded characteristics of their hosts – and that any further adaptation will be driven largely by non-genetic changes in the hosts, such as the acquisition of immune memory due to prior infections.
We therefore focused on the two evolutionary trajectories indicated in Figure 1. For human influenza, we examined the trajectory separating the H3N2 strains A/Aichi/2/1968 and A/Texas/JMM 49/2012. For swine influenza, we examined the trajectory separating the H1N1 strains A/swine/Wisconsin/1/1957 and A/swine/Indiana/A00968365/2012. In both cases, the starting strains for these trajectories meet the criterion specified in the previous paragraph – they are viruses with NPs that have had several decades to adapt to their respective hosts.
In order to map the mutations along these evolutionary trajectories, we utilized a previously described approach [34] for estimating the posterior distribution of mutational paths through protein sequence space by probabilistically placing mutations [49], [50] on trees sampled from a posterior distribution using BEAST [51]. The inferred mutational paths are shown in Figure 2. The human influenza NP accumulated 40 amino-acid mutations along the roughly 44-year trajectory, corresponding to 34 unique mutations relative to the initial Aichi/1968 NP (six mutations are reversions). The swine influenza NP accumulated 18 amino-acid mutations along the roughly 55-year trajectory, corresponding to 18 unique mutations relative to the initial swine/Wisconsin/1957 NP (there are no reversions).
We posit that two factors contribute to the slower rate of amino-acid substitution along the swine NP evolutionary trajectory relative to that of the human NP. First, as discussed in the previous section, the swine NP homolog is under less direct selection from immune memory than its human counterpart. Second, the strongest selection on influenza is from antibodies against the viral surface proteins, and so much of NP's sequence evolution is driven by stochastic genetic hitchhiking with adaptive antibody-escape mutations in these surface proteins [27], [52]. The reduced immune selection on these surface proteins in the swine lineage [37]–[43] probably curtails opportunities for similar genetic hitchhiking by mutations to the swine NP homolog. However, it is important to note that NP function is absolutely essential for viral replication in all strains of influenza [24], [25], and that decreases in NP function dramatically impair viral fitness [34]. Therefore, both adaptive and hitchhiking mutations in NP must first satisfy the stringent stabilizing selection for retention of protein function before they have an opportunity to fix.
In order to examine the association between NP evolution and selection from CTLs, we comprehensively mapped human CTL epitopes in the human and swine influenza NP homologs. Numerous experimental studies have identified epitopes in NP that are targeted by human CTLs (see for example [30], [53]–[57] plus many others). The Immune Epitope Database [58] contains a comprehensive listing of such experimentally characterized epitopes. We created a software package (https://github.com/jbloom/epitopefinder) to systematically parse this database for MHC class I epitopes with an experimentally verified human T-cell response that are between 8 and 12 residues in length and align with no more than one mismatch to NP. We considered epitopes to be present in human influenza NP if they matched to either the Aichi/1968 or Texas/2012 NP, and to be present in swine influenza NP if they matched to either the swine/Wisconsin/1957 or swine/Indiana/2012 NP. We removed redundant epitopes from the same MHC class I gene allele group (see http://hla.alleles.org/nomenclature/naming.html) or from the same supertype [59] if the allele group was not specified.
Figure 3A shows the number of characterized epitopes that contain each site in NP. As can be seen from this figure, the distribution of CTL epitopes is non-uniform along NP's sequence, with some sites falling in many known epitopes and others falling in none. The distributions of epitopes along the NP sequence are highly similar for the human and swine NP homologs. Figure 3B shows the distribution of number of epitopes per site for the human and swine NP homologs. These distributions are nearly indistinguishable (see the Figure 3 legend for statistical testing). Overall, Figure 3 indicates that the human and swine NP homologs contain nearly identical numbers of known human CTL epitopes.
If the NP from human influenza is under selection from human CTLs, we might expect this to lead to an increased rate of fixation of mutations in CTL epitopes. No such selection is expected to occur for the NP from swine influenza, as swine influenza is definitely not under pressure from human CTLs, and is probably not under strong selection even from swine CTLs for the reasons discussed in the Introduction.
To compare the relative rate of substitution in known CTL epitopes for the two NP homologs, we determined the number of epitopes at the sites of the mutations that fixed along the evolutionary trajectories from Figure 2. As shown in Figure 4, for the human NP homolog, the typical fixed mutation falls in more epitopes than an average site – whereas for the swine NP homolog, the typical fixed mutation falls in fewer epitopes than an average site. We interpret these results as follows: the known epitopes in NP tend to involve sites that are less inherently mutationally tolerant than the average site, either due to a tendency of CTLs to target conserved regions or a bias towards the experimental discovery of epitopes in conserved regions of NP (the tendency of characterized CTL epitopes to fall in conserved regions of viral proteins has also been noted by others [60], [61]). This tendency for the epitopes to fall in less mutationally tolerant regions of NP means that in the absence of CTL selection, the site of the typical fixed mutation contributes to fewer epitopes than an average site – this is the case for the swine NP homolog. But for the human NP homolog, selection for adaptive mutations in sites targeted by CTLs is sufficient to cause the fixed mutations to fall in more epitopes than an average site – and in significantly more epitopes than mutations fixed in the swine NP homolog (P = 0.008, see the Figure 4 legend for statistical testing).
The results in the previous section support the idea that there is pressure for adaptive change in human CTL epitopes for human influenza NP, but not for swine influenza NP. The facts discussed in the Introduction also strongly suggest that swine influenza NP is also under much less selection from swine CTLs than human influenza NP is from human CTLs. How do these differences in adaptive pressures influence the prevalence of epistasis during evolution?
We have previously performed a systematic test for a specific form of epistasis in the Aichi/1968 human influenza NP [34]. Specifically, we introduced all single mutations from the human NP evolutionary trajectory (Figure 2A) into the initial Aichi/1968 NP parent sequence, and quantified the effect of the mutations on total transcriptional activity by the influenza polymerase in transfected 293T cells. The previously described results from these experiments are shown in Figure 5A. Three of the 34 single mutations are highly deleterious as individual changes to the Aichi/1968 NP, despite the fact that they eventually fixed during the virus's evolution. We have previously shown that these three individually deleterious mutations were able to fix during NP's natural evolution due to epistatic interactions with other mutations that alleviated their deleterious effects [34].
Do similar epistatic interactions occur during the evolution of the swine influenza NP? To experimentally address this question, we introduced all of the single mutations from the swine NP evolutionary trajectory (Figure 2B) into the initial swine/Wisconsin/1957 NP parent sequence, and quantified the effect on transcriptional activity. These results are shown in Figure 5B. None of the mutations have a substantial deleterious effect as individual changes, indicating that none of them were dependent on epistatic interactions with other mutations. Therefore, while the 44-year evolutionary trajectory of the adaptively evolving human influenza NP involved the fixation of three mutations involved in strong epistatic interactions, we see no evidence of similar epistatically interacting substitutions along a 55-year evolutionary trajectory of the swine influenza NP. We acknowledge that the difference in the numbers of substitutions involved in epistatic interactions (3 out of 34 for human influenza NP, 0 out of 18 for swine influenza NP) is not statistically significant, and therefore merely provides anecdotal support for the idea that epistatically interacting substitutions are more common in the adaptively evolving human NP homolog. However, this anecdotal support becomes much more convincing when combined with the observations in the next section.
Is the presence of epistasis in the human but not the swine influenza NP due to the fact that only the former is adaptively evolving to escape from CTL selection? One way to test this idea is to examine whether the epistatic mutations in the human NP are at sites that contribute disproportionately to CTL escape. We have previously noted that the three epistatically constrained mutations in human NP are in known CTL epitopes [34]. Here we use our new comprehensive mapping of CTL epitopes described above to more thoroughly test the hypothesis that epistasis in the human NP is associated with CTL escape. Figure 6 shows that the epistatic mutations occur at sites that contain significantly more CTL epitopes than either average sites in NP or the set of sites that actually substituted along the evolutionary trajectory. Therefore, not only are epistatically interacting substitutions enriched during the evolution of the adaptively evolving human influenza NP relative to its swine influenza homolog – furthermore, the epistasis involves mutations that play an especially important role in the protein's adaptive evolution.
We have used a combination of computational and experimental analyses to examine whether epistasis is more common during adaptive protein evolution. We did this by comparing the evolution of an adaptively evolving NP from human influenza with a closely related homolog from swine influenza that is not under similar pressure for adaptive change. Experimentally, we find that strong epistatic interactions are fixed only during the evolution of the adaptively evolving human influenza NP homolog. Our computational analyses strongly suggest that the different patterns of epistasis are due to the fact that only the human influenza NP homolog is undergoing continuing adaptive evolution. Specifically, mutations that fix in the human influenza NP are significantly more likely to be in sites targeted by human immune memory than are mutations in the swine influenza homolog – and the epistatic interactions all involve sites that are heavily targeted by such immune selection. Overall, these results suggest that epistatically interacting substitutions are significantly enriched in adaptive versus non-adaptive evolution.
Why are epistatically interacting substitutions more prevalent during adaptive evolution? Our experiments probe for epistatic interactions involving a mutation that is individually deleterious but becomes neutral or adaptive when paired with secondary mutations. As discussed in the Introduction, there are two mechanisms by which such epistatic interactions have been shown to fix during adaptive evolution: compensatory mutations and permissive mutations. Our prior work suggests that the epistatic mutations in human influenza NP fix primarily via the latter mechanism, although compensatory mutations may also play a lesser role [34]. Crucially, the driving force for both mechanisms is adaptation. For the compensatory-mutation mechanism, this driving force is obvious: an initial deleterious mutation is more likely to persist long enough to be paired with a compensatory mutation if the initial mutation also confers some adaptive benefit (although mildly deleterious mutations can also fix without compensation, albeit at a lower rate). Somewhat less obviously, a similar force drives the permissive-mutation mechanism: although the initial permissive change is stochastic, the fixation of its subsequent pairing with the mutation that it permits is more likely if the latter change is adaptive [21]. Although epistatically interacting mutations can fix during non-adaptive evolution by similar temporal mechanisms, there is no underlying force to favor these relatively rare epistatic combinations over more abundant and easily accessible non-epistatic mutations.
This explanation can be stated more succinctly in terms specific to the NP homologs studied here. In the absence of adaptation, evolution tends to fix easily accessible non-epistatic mutations that have no adverse effect – in other words, the evolution of the swine influenza NP is dominated by stabilizing selection for retention of function. The human influenza NP is also under strong stabilizing selection for retention of function, but in addition experiences diversifying selection for change in immune epitopes. Some of these adaptive immune-escape mutations have adverse effects on NP function, and so selection biases evolution towards epistatic combinations that enable the adaptive mutations to fix while retaining NP function.
Most experimental studies of epistasis have focused on its role in constraining adaptation [1]–[10]. Our results suggest that caution may be warranted in extrapolating findings about the frequency of epistatically interacting substitutions during adaptation to more general evolutionary scenarios, since such substitutions appear to be more common during adaptive than non-adaptive evolution.
The input sequences for construction of the phylogenetic tree (Figure 1) and mutational paths (Figure 2) were downloaded from the Influenza Virus Resource [62]. For human influenza, up to 5 sequences per year were retained from the following lineages: H1N1 (isolation dates from 1918 to 1957, and then from 1977 to 2008), H2N2 (isolation dates from 1957 to 1968), and H3N2 (isolation dates from 1968 to 2012). For swine influenza, up to 5 sequences per year and subtype were retained from North American swine influenza. For the human H1N1 isolated in 1977 or later, 24 years were subtracted from the isolation dates because these sequences are from an influenza lineage revived after being frozen for roughly 24 years [45]. We excluded sequences that were classified as mis-annotated by [63] or that are strong outliers from the molecular clock based on an analysis with RAxML [64] and Path-O-Gen (http://tree.bio.ed.ac.uk/software/pathogen/).
The sequences were translated, date-stamped, and used as input to BEAST [65] with a strict molecular clock, a JTT [66] model of substitution, and a relatively loose coalescent-based prior on the tree. Figure 1 shows a maximum clade credibility tree rendered with FigTree (http://tree.bio.ed.ac.uk/software/figtree/).
The mutational paths in Figure 2 were constructed using the approach described in [34], and were rendered using GraphViz (http://www.graphviz.org/).
The source code, input data, and detailed documentation for the construction of the phylogenetic tree and the mutational paths can be accessed on GitHub via http://jbloom.github.io/mutpath/example_influenza_NP_1918_Descended.html
The CTL epitopes were identified by downloading from the Immune Epitope Database [58] all epitopes with a positive T-cell response with source organism Influenza A virus and host Homo sapiens. We created a new software package, epitopefinder (https://github.com/jbloom/epitopefinder), to map specific epitopes to NP.
This mapping was done by parsing all MHC class I peptide epitopes of 8 to 12 residues, and removing as redundant any epitopes that overlapped by 8 or more residues and were from the same MHC class I allele group (see http://hla.alleles.org/nomenclature/naming.html) or from the same MHC class I supertype [59] if no allele group was specified. For redundant epitopes, the shortest epitope sequence was retained. The non-redundant epitopes were aligned to NP: if they aligned to Aichi/1968 or Texas/2012 with no more than one mismatch then they were considered to be present in the human NP homolog, and if they aligned with no more than one mismatch to swine/Wisconsin/1957 or swine/Indiana/2012 with no more than one mismatch then they were considered to be present in the swine NP homolog. The number of epitopes in which each site participates is listed in Tables S1 and S2.
The source code, input data, and detailed documentation for mapping the epitopes and for the computing the P-values can be accessed on GitHub via http://jbloom.github.io/epitopefinder/example_NP_CTL_epitopes_H3N2_and_swine.html
We measured the function of the NP mutants by using flow cytometry to quantify the mean fluorescent intensity of 293T cells 20 hours after they had been transfected with plasmids encoding the NP variant in question, the three influenza polymerase proteins (PB2, PB1, PA), and the fluorescent reporter pHH-PB1flank-eGFP [67]. The data for the human NP homolog in Figure 5A were originally described in [34], and are reprinted here.
The data for the swine NP homolog in Figure 5B were generated by following the protocol described in [34] with the following modifications: the polymerase proteins were derived from the A/California/4/2009 swine-origin H1N1 strain, and the measured signal was normalized to that obtained using the wild-type swine/1957 NP. The polymerase plasmids (pHWCA09tc-PB2, pHWCA09tc-PB1, and pHWCA09tc-PA) have been described previously [68], while the insert for the swine/1957 NP plasmid (pHWswine57-NP) was synthesized commercially and cloned into pHW2000 [69]; the viral-RNA sequences for all four plasmids are in Dataset S1. The A/California/4/2009 swine-origin H1N1 polymerase proteins were chosen because the NP of this strain is closely related to NPs from the latter part of the swine influenza trajectory in Figure 1. We verified that the NP plasmid concentration used in [34] gave signal that was near the midpoint of the assay's dynamic range when using this combination of NP and polymerase genes (Figure S1). The data in Figure 5B represent the mean and standard error of at least three independent replicates; numerical values are in Table S3.
|
10.1371/journal.ppat.1003585 | Viral Membrane Fusion and Nucleocapsid Delivery into the Cytoplasm are Distinct Events in Some Flaviviruses | Flaviviruses deliver their genome into the cell by fusing the viral lipid membrane to an endosomal membrane. The sequence and kinetics of the steps required for nucleocapsid delivery into the cytoplasm remain unclear. Here we dissect the cell entry pathway of virions and virus-like particles from two flaviviruses using single-particle tracking in live cells, a biochemical membrane fusion assay and virus infectivity assays. We show that the virus particles fuse with a small endosomal compartment in which the nucleocapsid remains trapped for several minutes. Endosomal maturation inhibitors inhibit infectivity but not membrane fusion. We propose a flavivirus cell entry mechanism in which the virus particles fuse preferentially with small endosomal carrier vesicles and depend on back-fusion of the vesicles with the late endosomal membrane to deliver the nucleocapsid into the cytoplasm. Virus entry modulates intracellular calcium release and phosphatidylinositol-3-phosphate kinase signaling. Moreover, the broadly cross-reactive therapeutic antibody scFv11 binds to virus-like particles and inhibits fusion.
| Many viruses package their genetic material into a lipid envelope. In order to deliver their genome into the host-cell cytoplasm, where it can be replicated, viruses must fuse their envelope with a cellular lipid membrane. This fusion event is therefore a critical step in the entry of an enveloped virus into the cell. In this study, we used various cell biological and biochemical approaches to map precisely the cell entry pathway of two major human pathogens from the flavivirus family, yellow fever virus and Japanese encephalitis virus. We discovered that these viruses co-opt cellular phospholipid signaling to promote the fusion of their envelope with the lipid envelope of small compartments inside the host-cell endosomes. The viral genome remains trapped in these compartments for several minutes until the compartments fuse with the surrounding endosomal membrane. It is this second membrane fusion event that delivers the viral genome into the cytoplasm. We also showed that the antibody fragment scFv11 inhibits the fusion of the viral envelope with small lipid compartments, explaining the therapeutic activity of the scFv11 antibody. Our work identifies new vulnerabilities in the entry pathway of flaviviruses, including the formation of small endosomal compartments and two distinct membrane fusion events involving these compartments.
| Many enveloped RNA viruses utilize the endocytic pathway to enter host cells [1], [2]. Endocytosis begins at the cell membrane, where these viruses bind to their cellular receptors and ends at the lysosome, the “stomach” of the cell. Along the endocytic pathway, changes in the lipid composition and environmental pH provide a series of distinct milieus for specific cellular or viral functions to occur [3]. Enveloped viruses and bacterial toxins enter the endocytic pathway by binding receptors on the cell surface that are coupled to the endocytic machinery, in particular clathrin adaptors. These microbial cargoes undergo sorting at two different checkpoints [4], [5], [6]. The first is in early endosomes (EEs) where the vesicular contents are either directed back to the cell membrane via tubular structures, or targeted to the trans-Golgi network (TGN). Alternatively, the cargo contents are sorted into intraluminal vesicles and transported to late endosomes via endosomal carrier vesicles (ECVs). ECVs require functional microtubules to be transported to the second sorting station, the late endosomes. In late endosomes, cargo contents can be forwarded to the TGN, the cytoplasm, or for lysosomal degradation. ECVs originate from EEs. Both the EEs and ECVs are rich in cholesterol, phosphatidylserine (PS) and phosphatidylinositols (PI) [7], [8], [9]. The level of cholesterol decreases along the endocytic pathway and is replaced with ceramide in late endosomes and lysosomes, where it maintains membrane fluidity [10]. Unlike cholesterol and PS, the anionic lipid BMP (bis(monoacylglycero)phosphate), also known as LBPA (lysobisphosphatidic acid), is abundant in internal membranes of lysosomes and late endosomes, and depleted in the EEs [7]. BMP regulates membrane sorting and dynamics in the late endosome. Autoantibodies against this lipid result in human disorders such as Niemann-Pick type C (NPC) syndrome, characterized by dysfunctional sorting and trafficking in late endosomes [11].
The genus flavivirus includes important human pathogens such as dengue, Japanese encephalitis (JE), West Nile (WN) and yellow fever (YF) viruses. Flaviviruses contain a lipid envelope and a positive-stranded RNA genome encoding for a polyprotein that is processed by the host- and viral proteases to yield the viral proteins. Three structural proteins (C, M and E) form the virions; the nonstructural proteins (NS1-5) are required for virus replication, transcription and modulation of the host innate immune system [12]. Flaviviruses assemble in specialized structures within the endoplasmic reticulum and mature in the Golgi network [13]. Glycoprotein E forms the outer protein shell of the virion, mediates cellular attachment, and catalyzes the fusion of the viral and cellular membranes necessary to deliver the genome into the cytoplasm. The E ectodomain contains three domains (I–III) connected by hinges [14], [15]. Conserved histidine residues at the domain I-domain III interface become protonated at the reduced pH of early endosomal compartments (pH 6–6.5), thereby triggering a conformational rearrangement in E that drives membrane fusion [16], [17].
Although flaviviruses generally follow the clathrin-mediated endocytic pathway, other mechanisms of entry have also been proposed. Dengue virus has been reported to fuse primarily from within Rab7-positive late endosomes. However, certain dengue strains have been reported to infect cell independently of Rab7 [18] and alternative entry pathways have been proposed for specific viruses and cell types [19], [20]. Moreover, in certain flaviviruses low pH does not appear to be sufficient to trigger fusion suggesting that additional factors may be required [21]. Indeed, different compartment-specific lipids are required for fusion of dengue and Japanese encephalitis viruses [22], [23], [24], [25].
Flaviviruses produce small noninfectious virus-like particles (VLPs) during infection [26]. Little is known about the role of these VLPs in infection or in host immunity however, recombinant flavivirus VLPs are the subject of intense study as vaccine candidates and gene delivery vehicles [27], [28], [29], [30]. It is not known whether VLPs have the same requirements for fusion and the same cell entry pathways as full-sized virions.
In this study, we use JE-VLPs and YFV virions as model systems to study the cell flavivirus entry pathway. Using a combination of approaches—including single-particle tracking in live cells, a liposome-based membrane fusion assay, a quantitative RT-PCR RNA delivery assay, and viral infectivity assays—we show that our model viruses modulate cellular signal transduction to promote membrane fusion to ECVs, which occurs several minutes before nucleocapsid delivery into the cytoplasm, suggesting that these are two distinct events in virus entry. Our observations are consistent with an entry pathway in which certain flaviviruses fuse with ECVs and require host proteins to deliver the nucleocapsid to the cytoplasm. This pathway has a precedent in vesicular stomatitis virus (VSV) [5], although VSV does not impact PI-3-phosphate kinase activity [31], [32] nor trigger intracellular calcium release during entry [33]. Moreover, we demonstrated the ability of a broadly cross-reactive therapeutic antibody, scFv11, to block membrane fusion with the host cells and EE/ECV-like liposomes.
Recombinant Japanese encephalitis virus-like particles (JE-VLPs) were obtained by overexpressing the prM and E genes (see Materials and Methods) in human HEK 293T cells (Figure S1A), and in insect Tni cells using a baculovirus-based expression system (Figure S1B). JE-VLPs and YFV were purified by precipitation of secreted cellular products with polyethylene glycol, followed by sedimentation in sucrose density gradient (Figure S2). Since flaviviruses E proteins bind to heparan sulfate [34], the virus particles could alternatively be purified by affinity chromatography on a heparan sulfate column (Figure S2A–B). Both methods showed highly purified secreted VLPs. Coomassie stained SDS-PAGE confirmed that prM was cleaved to pr and M in the purified particles (Figure S2D), indicating a high degree of maturation in the JE-VLPs. The concentration of the purified VLPs was estimated by enzyme linked immunosorbent assay (ELISA- see Materials and Methods). Negatively stained electron microscopy (EM) show that the JE-VLPs have rough surfaces and diameters ranging from 30 to 40 nm (Figure S2F). Purified yellow fever virus (YFV) particles had a similar appearance but were 50 nm in diameter (Figure S2G). Dynamic light scattering (DLS) analysis indicated an average diameter of approximately 40 nm for the JE-VLPs (Figure S2E), consistent with the EM data.
Upon attachment to the plasma membrane, flaviviruses localize to clathrin-coated pits and undergo endocytosis [35], [36]. Although, VLPs, like full virions, are expected to enter the endocytic pathway, this has not yet been demonstrated experimentally. We treated Vero cells, a commonly used cell line to study flaviviruses, with JE-VLPs and stained fixed cell at different time points with anti-E protein and the endosomal markers Rab5 and Rab7. VLPs attached to the Vero cells immediately. After 5 minutes, E protein colocalized with Rab5. At 15 minutes, E protein colocalized with both Rab5 and Rab7 (Figure 1A) but more so with Rab7, as indicated by the Pearson's coefficients of 0.19 and 0.34 for colocalization with Rab5 and Rab7, respectively. By 25 minutes, most of the particles colocalized with Rab7 indicating their arrival to late endolysosomal compartments. This confirms that the secreted particles follow the same general cell entry pathway as full mature virions.
To dissect the mechanism and kinetics of JE-VLP cell entry, we tracked the membrane fusion step in real time in live cells using confocal microscopy. The lipid envelope of JE-VLPs was labeled with self-quenching concentrations of the hydrophobic dye rhodamine C18 (R18). Fusion of the VLP membrane with an endosomal membrane was detected as a sudden dequenching of R18 fluorescence as the concentrated dye was diluted with lipids from the host membrane. Fluorescent puncta that showed no dequenching were excluded from our analyses. Fusion events were first detected approximately 250 s after treatment of Vero cells, consistent with fusion occurring in early to intermediate endosomal compartments (Figure 1B–C). Notably, the R18 fluorescence signal for individual fusion events remained at its maximal level for 251±97 seconds (n = 14) before starting to decay and the R18 dye remained concentrated in puncta during this fluorescence plateau (Figure 1B–D). Fluorescence was expected to decay immediately after dequenching due to continued dilution with host lipids. The consistent persistence of fluorescent puncta for several minutes after fusion suggests that the R18 dye becomes trapped in an endosomal subcompartment after the initial membrane fusion event. The rate of fluorescence decay of the puncta after the fluorescence plateau, with a decay half-time of 94±64 (n = 14), is too great to be attributed to photobleaching alone, suggesting that the decay is due to a second and distinct event leading to dilution of the R18 dye to below detection levels (Figure 1E). Taken together the particle tracking data suggest that JE-VLPs are fusogenic, that fusion occurs in early to intermediate endosomes, and that the viral lipids remain trapped in a small endosomal subcompartment for several minutes, until a distinct event releases the lipids into a much larger compartment. R18-labeled YFV strain 17D particles produced in BHK cells using an NS1 trans-complementation strategy described previously [37] had similar R18 dequenching kinetics, as judged from analysis of a smaller number of YFV particles.
Chloroquine is a widely used lysosomotropic drug that acts by inhibiting the acidification of the endocytic pathway. Clinical studies have demonstrated the safety, tolerability, and efficacy of chloroquine as a treatment against enveloped RNA viruses [38]. Treatment with 194 µM (0.1 g/l) chloroquine was not toxic to Vero cells (Figure 2). In the presence of chloroquine, fluorescence dequenching of R18 in R18-labeled JE-VLPs and YFV (Figure 2B–C) was completely inhibited, indicating that membrane fusion of JE-VLPs is dependent on the acidic pH of endosomal compartments. This is consistent with the inhibitory effect of chloroquine and other endosomal pH-neutralizing agents reported in other flaviviruses [35], [39].
To confirm that mature flavivirus virions are also dependent on acidic endosomal pH, and that chloroquine inhibits not only fusion but also nucleocapsid delivery into the cytoplasm, we measured the effect of chloroquine on viral RNA release and infectivity of YFV. To measure delivery of YFV genomic RNA into the cytoplasm, infected Vero cells were fractionated into cytoplasmic and endosomal fractions (Figure S3), and viral RNA in the cytoplasmic fraction was detected by relative quantitative RT-PCR (Figure 2D). Chloroquine blocked 95–97% of YFV RNA release in Vero cells. Moreover, in a plaque assay for YFV infectivity in BHK cells, chloroquine completely inhibited infectivity (Figure 2E). These results confirm that the acidity of endosomal compartments is required for flavivirus infection, and that infection can be blocked with lysosomotropic drugs such as chloroquine that raise the endosomal pH.
In mammalian cells, endosomal carrier vesicles (ECVs) are transported from early endosomes (EEs) to late endosomes on microtubules. ECVs then dock onto and fuse with late endosomal membrane [40]. Inhibition of microtubule-dependent transport with the microtubule depolymerizing agent nocodazole inhibits West Nile virus infection [35]. Treatment with 20 µM nocodazole was not toxic to the cells, although changes in cell morphology were observed in the treated cells (Figure 3A). In cell pretreated with nocodazole, fluorescence dequenching of R18 in R18-labeled JE-VLPs and YFV occurred with similar kinetics as in untreated cells (Figure 3), indicating that nocodazole does not affect membrane fusion. However, the R18 fluorescence intensity gradually increased after fusion (Figure 3B–C), rather than decaying after a 3–4 min plateau as in untreated cells (Figure 1C–D). The gradual increase in fluorescence in the presence of nocodazole may be attributed to sequential homotypic fusion events with other non-fluorescent ECVs in early endosomal compartments. The resulting vesicles would be still relatively small (hence the lack of dilution-dependent decay) but would allow additional R18 dequenching (hence the gradual increase in fluorescence). Additionally, inefficient lipid mixing in the presence of nocodazole may contribute to the gradual increase in fluorescence. Consistent with this interpretation, YFV RNA release into the cytoplasm, measured by RT-PCR as described above, was inhibited by 80% in the presence of nocodazole (Figure 3D). Moreover, virus infectivity was completely inhibited by nocodazole (Figure 3E). Together, these results indicate that certain flaviviruses fuse with ECVs, and that membrane fusion and RNA delivery into the cytoplasm are two distinct events.
Having established that membrane fusion occurs early in the endocytic pathway whereas the YFV nucleocapsid is delivered into a late endosomal compartment, we sought next to determine the importance of factors specific to late endosomal compartments for fusion and nucleocapsid delivery. BMP (also known as LBPA) is an anionic lipid that is present in internal membranes, but not the limiting membrane, of late endosomes. An antibody against BMP accumulates on these internal membranes [41] and interferes with the protein sorting and membrane transport functions of the late endosome. Treatment with anti-BMP antibody causes a phenotype characteristic of Niemann-Pick disease type C (NPC) [11], [42]. To assess the role of late endosomal trafficking in flavivirus cell entry, we incubated Vero cells with an anti-BMP antibody and assay membrane fusion activity, RNA delivery and infectivity. The staining patterns of endocytosed anti-BMP antibody and of a total mouse IgG control in BHK cells are shown in Figure 4. Pretreatment with anti-BMP antibody followed by infection with R18-labeled JE-VLPs or YFV produced similar R18 fluorescence profiles as in cells pretreated with nocodazole, with normal R18 dequenching kinetics but no fluorescence decay (Figure 4C–D). We conclude that membrane fusion is not inhibited by blocking the protein and lipid sorting functions of late endosomes. In contrast, the endocytosed anti-BMP antibody reduced both YFV RNA delivery to the cytoplasm of Vero cells and YFV infectivity in BHK cells by 35% (Figure 4E–F). These data suggest that BMP in internal late endosomal membranes is required for virus infectivity, either to ensure correct ECV trafficking, or possibly to promote “back-fusion” of ECVs to the limiting membrane of the late endosome.
In our emerging model of flavivirus cell-entry, virions fuse with ECVs and the nucleocapsid is delivered into the cytoplasm when the ECVs fuse back to the limiting late endosomal membrane. To test this model, we set out to evaluate the importance of factors required for ECV formation for fusion and nucleocapsid delivery and for trafficking of ECV to the late endosome. The lipid phosphatidylinositol-3-phosphate (PI(3)P) is generated by PI(3)P kinase and is required for endocytic trafficking [43]. Vero cells infected with either fusogenic or chemically inactivated YFV or JE-VLPs induced robust activation of PI(3)P kinase as indicated by phosphorylation of AKT, also known as protein kinase B (Figure S4). The PI(3)P kinase inhibitor wortmannin inhibits ECV formation in mammalian cells [44]. PI(3)P is abundant in ECVs and early endosomes, but not in the late endosome [45]. To determine whether PI(3)P kinase activity is required for membrane fusion, we pretreated Vero cells with 60 nM wortmannin and tracked fusion of R18-labeled JE-VLPs and YFV as described above. This concentration of wortmannin was not toxic but did cause vacuoles to form inside the cells as expected (Figure 5A). The resulting R18 fluorescence profiles were similar to those with nocodazole or anti-BMP antibody pretreatment, with the same R18 dequenching kinetics but no fluorescence decay (Figure 5B–C and Movie S3). Since wortmannin inhibits ECV formation, we attribute the gradual increase in fluorescence in the presence of wortmannin to inefficient lipid mixing. RT-PCR analysis and plaque assay showed that wortmannin pretreatment blocked RNA release into the cytoplasm of Vero cells and virus infectivity in BHK cells, respectively (Figure 5D–E). Collectively, these results suggest that in the absence of ECVs, JE-VLPs and YFV fuse with other as yet unidentified membranous structures or compartments, where the nucleocapsid remains trapped. Alternatively, the lipid composition or curvature of the limiting early endosomal membrane may preclude full membrane fusion of JE-VLPs. The observed R18 dequenching may then be attributed to transient hemifusion of the viral and endosomal membranes, which would allow lipid mixing of the proximal viral and endosomal lipid monolayers, and therefore dilution of R18, without nucleocapsid delivery into the cytoplasm.
In certain flaviviruses low pH does not appear to be sufficient to trigger fusion [21]. We note that YFV fusion is only partially inactivated by a pretreatment under conditions (pH 6.2) but that infection is nevertheless completely inhibited in acidic media (Figure S5). To determine the minimal physicochemical requirements for membrane fusion of JE-VLPs and YFV, we measured fusion of R18-labeled virus particles in vitro in a bulk fusion assay with synthetic liposomes. The liposomes were 0.1 µm in diameter (see Materials and Methods) and their lipid composition was chosen to correspond to those in EEs/ECVs: cholesterol, phosphatidylcholine (PC), phosphatidylethanolamine (PE), PI(3)P, and phosphatidylserine (PS) at a molar ratio of 3∶4∶1∶1∶1 [7], [9]. Fusion was measured by R18 dequenching. We found that both JE-VLPs and YFV fused with the liposomes at pH 5.5, but not at neutral pH (Figure 6A). In flaviviruses, a conserved cluster of histidine side chains acts as a “pH sensor”, which triggers the fusogenic conformational change in response to the reduced pH of the endosome [16], [17]. The histidine modifying agent diethylpyrocarbonate (DEPC) inactivates VSV [46] and dengue virus [22] by inhibiting the fusogenic conformational change. We found that DEPC also blocked fusion of R18-labeled YFV with liposomes at pH 5.5 (Figure 6B). In conclusion, synthetic liposomes with a lipid composition similar to ECVs are sufficient to induce flavivirus fusion in vitro at low pH.
PS and PI(3)P are abundant in mammalian EEs/ECVs [7]. We found that JE-VLPs and YFV bound to PS-coated beads (Figure 6C). Similarly, heparan sulfate beads also bind the virus particles, consistent with reports that flaviviruses bind heparan sulfate through the viral envelope protein [34], [47]. However, the virus particles did not bind to PI(3)P beads (Figure 6D), suggesting that the binding to PS is not due to nonspecific electrostatic interactions and that PS may act as an intracellular receptor or fusion cofactor for flaviviruses. PS and PI(3)P beads bound with equal affinity to polyarginine peptides, indicating that the surface charges of the two types of beads are comparable (Figure S6).
Calcium released into the cytoplasm during viral infection can result in the translocation of PS from cytoplasmic to extracellular lipid leaflets [48], [49], [50]. Imaging of cells with the calcium-dependent dye Fluo-4 showed that calcium is released into the cytoplasm within one minute of infection with YFV or JE-VLPs (Figure S7). To determine whether intracellular calcium release promotes flavivirus infection we measured the effect of the cell-permeable calcium chelator BAPTA on YFV infectivity. BAPTA reduced YFV infectivity to less than 5% of the untreated infected control (Figure S7C). We propose intracellular calcium release during flavivirus infection may cause a redistribution of PS towards extracellular or luminal leaflets, which may be important for flavivirus infectivity.
Antibodies that inhibit fusion by targeting the fusion loop of the E protein are important determinants in the humoral response to flavivirus infection [51], [52], [53]. The therapeutic scFv11 antibody fragment, which recognizes the fusion loop, was selected by phage display for binding to West Nile virus E protein [54], [55]. scFv11 protects mice from a lethal challenge of West Nile virus and also protects against dengue virus types 2 and 4 [54]. To probe the reactivity of scFv11 against other flaviviruses, we used an ELISA assay to measure the binding of scFv11 antibody to either JE-VLPs or YFVs. scFv11 bound tightly to JE-VLPs but did not bind to YFV (Figure 7A). These results were confirmed by co-elution of scFv11-VLP complexes in size-exclusion chromatography (Figure 7B), and by a plaque assay with YFV showing that scFv11 had no effect on YFV infectivity.
The location of the scFv11 epitope in the fusion loop of E suggests that scFv11 inhibits viral membrane fusion [55]. To test whether scFv11 inhibits fusion of JE-VLPs, we used the in vivo and in vitro fusion assays described above. Preincubation of R18-labeled JE-VLPs with scFv11 inhibited membrane fusion in Vero cells, as judged by the lack of R18 dequenching. Subsequent addition of untreated JE-VLPs produced R18 dequenching as expected (Figure 7C). In the bulk fusion assay, scFv11 reduced the acid-induced fusion of JE-VLPs with EE/ECV-like synthetic liposomes by 50% (Figure 7D). These data suggest that the fusion loop of JE-VLPs is accessible to scFv11, which inhibits fusion of JE-VLPs in EEs/ECVs.
The dissociation equilibrium constant of scFv11 from JE-VLPs, measured by isothermal titration calorimetry, was at 150 nM (Figure 7E). scFv11 was previously shown to bind soluble form of West Nile E with a 5 nM dissociation constant [54]. The higher affinity for West Nile E could be due to the fusion loop epitope being partially occluded in JE-VLPs, or to differences in the amino acid sequence or structure of the non-cognate JE E and the cognate West Nile E. The stoichiometry of binding was 0.648±0.013 scFv11 molecules per E protein. Thus, if the JE-VLPs each contain 60 E proteins, as is the case in an electron microscopy structure of tick-borne encephalitis VLPs [56], each JE-VLP would be capable of binding approximately 40 scFv11 molecules. The observed substoichiometric binding of scFv11 suggests that one third of the E-protein epitopes in VLPs do not bind scFv11 either because they are not fully exposed or because they are clustered too closely together to allow full occupancy by scFv11 without steric clashes. The latter is more likely given the presumed T = 1 icosahedral symmetry of the VLPs, in which each E protein displays identical surface epitopes [56]. In contrast, in mature virions in the E proteins are distributed in three distinct chemical environments with slightly different surface epitopes. Two different neutralizing antibodies against dengue and West Nile viruses bind to only two thirds of the E proteins in their cognate virions [57], [58], providing precedents for the stoichiometry reported here for scFv11 binding to JE-VLPs.
Many enveloped viruses enter the endocytic pathway and rely on specific features of the endosomal environment, in particular the reduced pH and the lipid composition, to trigger membrane fusion and productive delivery of the viral genome into the cytoplasm. Although it has been established that flaviviruses generally undergo clathrin-mediated endocytosis [35], [36], [59], the sequence and kinetics of the steps required for cell entry remain unclear. It has been reported that approximately 20% of dengue virions fuse early in the endocytic pathway while the rest fuse in late endosomes [60], but it is unclear whether all of these fusion events lead to productive infection. We note that diphtheria toxin, a bacterial bipartite toxin complex composed of carrier and toxin subunits, inserts its carrier subunit into early and late endosomal membranes but the active toxin subunit is only delivered to the cytoplasm from early endosomes, and most of diphtheria toxin complexes are degraded in the lysosome [4].
Little is known about the physical and biological properties of flavivirus VLPs- how and when they assemble, and what their possible roles are in infection and in virus evolution. In this study, we have dissected the cell entry steps of VLPs and virions from two different flavivirus species. Tracking of single virus particles in live cells and virus infectivity measurements in the presence of various cell biological inhibitors are consistent with an entry mechanism in which virus particles fuse preferentially with small endosomal carrier vesicles (ECVs), with nucleocapsid delivery into the cytoplasm occurring several minutes later, when the ECVs fuse with the limiting membrane of the late endosome. Alternatively, instead of fusing completely with ECVs, the virus particles may form metastable hemifusion intermediates with the ECVs, with full fusion only occurring in late endosomes, consistent with a previous report that dengue forms ‘restricted hemifusion’ intermediates [22]. Either way, we conclude that flavivirus membrane fusion and nucleocapsid delivery into the cytoplasm are distinct events in space and time (Figure 8).
While novel for flaviviruses, a sequential cell entry mechanism involving delivery into ECVs followed by back-fusion of the ECVs to the limiting late endosomal membrane is not unprecedented. VSV utilizes this mechanism of nucleocapsid delivery into the cytoplasm to reach the cytoplasm [5] although VSV does not impact PI(3)P signaling or trigger intracellular calcium release [31], [32], [33]. Similarly, anthrax lethal toxin (LT) from Bacillus anthracis, a bipartite toxin complex composed of carrier and toxin subunits, inserts carrier protein (protective antigen) into ECVs in response to endosomal acidification, and delivers its toxin subunit (lethal factor) into the cytoplasm upon back-fusion of the ECV with the limiting late endosomal membrane to deliver lethal factor to the cytoplasm [6].
The back-fusion of ECVs with the endosome's limiting membrane depends on the anionic lipid BMP [5], which is found in internal membranes and vesicles within late endosomes but not in the limiting endosomal membrane [8]. Treatment of cells with anti-BMP antibody did not inhibit membrane fusion of JE-VLPs or YFV but strongly inhibited both YFV infectivity and RNA delivery into the cytoplasm. This suggests that the cell entry mechanism of these viruses is dependent on back-fusion of ECVs to the limiting late endosomal membrane. Since anionic lipids are required for efficient fusion of dengue virus [22], the presence of the anionic lipid phosphatidylserine (PS) in ECVs may be responsible for the preference of JE-VLPs and YFV to fuse with ECV membrane over limiting endosomal membranes. Additionally, the presence of cholesterol in the target membrane promotes fusion [61], [62] and cholesterol chelation reduces flavivirus infectivity, although addition of exogenous cholesterol at the cellular attachment step has been reported to block JE and dengue virus cell entry [63]. Cholesterol is abundant in early endosomes and ECVs [8].
While anionic lipids are important in a general sense for cell entry of enveloped RNA viruses, PS may perform a more specific function. PS is abundant in early endosomes and ECVs, where PS represents about 9% of all phospholipids [7], [9]. YFV and JE-VLPs fused with liposomes of this lipid composition. Moreover, PS-coated beads bounds JE-VLPs and YFV whereas beads coated with PI(3)P, which is also anionic and present in ECVs, did not bind the virus particles. Interestingly, PS is expressed on the plasma membranes of insect cells [64] and malignant and non-apoptotic cells [65], [66].
The therapeutic antibody fragment scFv11 binds JE-VLPs with reasonably high affinity. Noninfectious flavivirus VLPs produced during infection [26] may thus serve as antibody decoys to promote immune evasion of the infectious virions. VLPs have recently been in focus as vaccine candidates [27]. Our study establishes that flavivirus VLPs can be used as a model for virus entry and for screening of therapeutic antibodies.
In summary, our work suggests a novel mechanism for flavivirus cell entry in which the virus fuses to ECVs and depends on a second cell-mediated membrane fusion event to deliver the viral genome from the vesicle lumen to the cytoplasm. We propose that flavivirus infection modulates PI(3)P-dependent signaling in the host and modifies host phospholipid distribution to promote fusion with endocytic compartments. Our findings provide a framework for future studies to determine the physicochemical basis of the preference for membrane fusion with ECVs, the nature of the contribution of specific lipids (BMP, PS, cholesterol) to fusion activity, and the precise sequence and kinetics of the molecular steps required for membrane fusion and nucleocapsid delivery.
Horse anti-WNV E antibody was a generous gift from L2 Diagnostics (New Haven). The scFv11 construct, a kind gift from Erol Fikrig, was expressed and purified as described [54]. The purified protein showed the expected purity and molecular weight in SDS-PAGE and size-exclusion chromatography (Figure S8). The following reagents were purchased from commercial sources: rabbit anti-human Rab5, Rab7, Phospho-AKT and Total-AKT antibodies (Cell Signaling), anti-LPBA (BMP) antibody (Echelon biosciences), Fluorescein (FITC)-labeled anti-horse antibody (Bethyl Labs.), horseradish peroxidase (HRP)-conjugated anti-horse (Sigma), Texas Red anti-rabbit antibody (Invitrogen), total mouse IgG (Sigma), Texas Red anti-mouse antibody (Invitrogen).
All inhibitors were freshly prepared before use according to the manufacturers' recommendations. 25 µM BAPTA (Sigma) was added to the virus, maintained during cellular attachment, and removed before addition of agarose plugs in plaque assays. Wortmannin (Sigma) was used at a final concentration of 0.1 µM while nocodazole (Sigma) was used at 20 µM final concentration. Chloroquine (Sigma) was used at a final concentration of 194 µM (0.1 g/l). All inhibitors were added to the cells before starting the experiments. In the plaque assay, all of the inhibitors were diluted to their half concentrations and kept in the medium after adding the agarose plugs as described in plaque assay method. Diethylpyrocarbonate (DEPC) (Sigma) was added at final concentration of 2 mM to purified YFVs for 30 min and then removed by buffer exchange using an ultrafiltration unit.
The prM-E sequence of JEV strain CH2195LA was cloned into the pAcgp67 vector (BD Biosciences). Sf9 cells were co-transfected with the linearized baculovirus genome and the pAcgp67-prM-E construct. The secreted recombinant baculovirus encoding the prM-E sequences was amplified in Sf9 cells. For protein expression, Tni cells (Expression Systems) were infected with the baculovirus in ESF921 medium (Expression Systems) at 27°C. Alternatively, JE-VLPs were expressed in HEK293T cells transiently transfected with the pcDNA-prM-E mammalian expression vector. Cell media was clarified by centrifugation at 4°C. We then added 2.5% (w/v) NaCl slowly with continuous stirring, followed by 10% (w/v) polyethylene glycol 6000 (PEG6000) to precipitate the VLPs. VLPs were pelleted at 20,000 rpm for 20 min. The pellet was resuspended in 20 ml of buffer (50 mM Tris pH 8.4, 0.1 mM EDTA, 0.15 M NaCl) and centrifuged at 20,000 rpm for 2 min to remove excess PEG6000. The VLPs were then purified on a 10–40% sucrose gradient by centrifugation for 9 h at 30,000 rpm in a SW32 rotor at 4°C. The gradient was then separated into 1-ml fractions and VLPs were detected by immunoblotting. Horse anti-WNV E antibody and anti-horse secondary antibody conjugated to HRP were both used at a dilution of 1∶104. Fractions containing VLPs were concentrated in ultrafiltration units with a 30-kDa molecular weight cutoff (Millipore). Buffer was exchanged for all positive layers to 50 mM Tris pH 7.4, 0.14 M NaCl and the VLPs were checked for the purity by SDS-PAGE on 4–20% acrylamide gels with Coomassie staining. The hydrodynamic radius of the JE-VLPs was determined by dynamic light scattering analysis (Malvern).
Generation of the BHK cell line expressing the NS1-GFP fusion protein for trans-complementation of YFV 17D genome lacking NS1 has been described previously [37]. Non-infectious ΔNS1 viruses were collected from the cell culture medium. The YFV particles were purified by PEG6000 precipitation and sucrose gradient as described above for JE-VLPs. In the absence of an effective antibody to detect YFV structural proteins, YFV was quantified in fractions from the sucrose gradient as plaque-forming units in a plaque assay with the BHK-NS1-GFP cells (see below).
Serial dilutions of YFV were added to BHK cells (5×105 cells/ml) in DMEM medium. Viruses were allowed to attach to the cells for 1 h at 37°C. A 1∶1 mixture of 2× DMEM medium (Gibco) and autoclaved 1.6% (w/v) agarose (37°C) was then layered onto the cells in 6-well plates. After 2–3 days, the wells were fixed with 7% formalin (Sigma) and the agarose plugs were removed. Cells were stained with 0.5% crystal violet (Sigma) to visualize the plaques. Excess stain was removed with water. For acid pretreatment of YFV, the buffer was exchanged to 50 mM HEPES pH 6.2 with an ultrafiltration device (Millipore), viruses were allowed to attach to cells for 15 min in DMEM pH 6.5 or 7.4, and cells were then washed with DMEM pH 7.4 before proceeding to the next step of plaque assay.
5 µl of JE-VLP suspension was applied for 2 min on the carbon surface of a glow-discharged carbon-coated grid (Microscopic Science). Excess sample was removed using absorbent paper and the grid was air-dried before examination. Data were collected using a Zeiss EM900 electron microscope.
JE-VLPs were labeled with Rhodamine C18 (R18; Invitrogen). The dye was added to the PEG-precipitated fraction of either YFV or JE-VLPs at a final concentration of 20 ng/µl and incubated for 15 min before the sucrose gradient centrifugation step.
Vero cells were grown on a coverslip petri dish (MatTek) overnight at a density of 5×105 cells/ml at 37°C, 5% carbon dioxide. Before microscopic examination, the medium was changed to serum-free OptiMEM (Gibco) and cells were stained with Hoechst stain (Invitrogen). JE-VLPs were added from a stock at 17 pM (50 ng/ml E protein) and the particles were kept in the medium during data collection. Time-lapse confocal microscopy was performed using a Zeiss microscope connected to 37°C incubated and buffered with 5% CO2. Time-lapse images were collected using a slice of 4 µm to avoid changes in confocal planes during data collection. Images were collected every 10 s. Data were analyzed with ImageJ [67].
Immunostaining was performed as described [68]. Briefly, Vero cells were grown on a coverslip overnight at 5×105 cells/ml and treated at different time points with 34 pM JE-VLPs (50 ng/ml E protein). 1 µl of Hoechst stain (10 g/l) was added to the cells for 10 min at 37°C. Cells were fixed with 4% paraformaldehyde and permeabilized with 1% Triton ×100. Cells were then blocked for 1 h with 10% fetal bovine serum (FBS) and stained with the primary anti-Rab5/7 antibody according to the manufacturer's recommendation (Cell Signaling). The cells were washed 10 times with PBS. For BMP staining, cells were fed 50 µg/ml anti-BMP antibody overnight and the cells were washed and fixed as above. The cells were then permeabilized with 0.3% Tween-20 in PBS and blocked for 1 h in 10% FBS. Cells were blocked for 1 h with 10% FBS and then stained with secondary antibody conjugated to Texas Red (Invitrogen). Cells were washed 10 times with PBS and mounted with fluoromount G (Microscopic Science) before examination. Semi-quantitative colocalization analysis (Pearson coefficient calculation) was performed with ImageJ.
Cholesterol, phosphatidylethanolamine (PE), PI(3)P, phosphatidylserine (PS), and phosphatidylcholine (PC) were mixed in chloroform at a molar ratio of 3∶1∶1∶1∶4, respectively, and then dried with argon gas and under vacuum for 2 h. The lipids were resuspended in 3 ml TEA buffer (10 mM triethanolamine pH 8.3, 0. 14 M NaCl) and subjected to 10 freeze-thaw cycles using liquid nitrogen and a 37°C water bath. The lipid suspension was then extruded through 0.1 µm membranes 21 times with a lipid extruder (Avanti Polar Lipids). The liposome suspension was added to R18-labeled virus particles. After a 5 min incubation, the pH was modified with either sodium acetate pH 5.5 buffer or with Tris pH 8.4 buffer (70 mM final concentration). R18 fluorescence was monitored after 1 min. with a QuantaMaster cuvette-based spectrofluorometer (Photon Technology International) or a time-domain plate-based fluorimeter (HoriBa).
ELISA plates were coated with 0.1 M carbonate buffer pH 9.6 and either JE-VLPs or YFV overnight at 4°C. 6×1010 JE-VLPs (300 ng E protein) or 6×107 plaque forming units of YFV were added to each well. The coated wells were blocked with 10% FBS in PBS for 1 h at room temperature. Primary antibody was added and plates were incubated for 1 h at room temperature followed by 10 washes with PBS. Secondary antibody was added and the plate was incubated for 30 min at room temperature. Plates were then washed 10 times with PBS. HRP substrate TMB (Sigma) was added and stop solution (Sigma) was used to stop the color development. Absorbance was measured at 450 nm. Standard curves of JE-VLPs were generated by serial dilution of purified VLPs. The concentration of E protein was estimated by comparing Coomassie staining on SDS-PAGE with staining from known concentrations of bovine serum albumin (Sigma). The E protein concentration was used to determine the concentration of JE-VLPs in this study, assuming 60 copies of E per VLP.
1 ml of media from BHK cells infected with a YFV multiplicity of infection (MOI) of 0.1, or insect cells expressing prM-E were mixed with a 50 µl suspension of beads coated with heparan sulfate (HS; Sigma), phosphatidylserine (PS; Echelon biosciences), or phosophoinisitol-3-phosphate (PI(3)P). The beads were pre-equilibrated with 50 mM Tris pH 7.4 and 0.1 M NaCl. Uncoated beads (Echelon biosciences) were used as a negative control. The beads were collected by centrifugation, washed with equilibration buffer and eluted with 50 mM Tris pH 7.4 and 0.5 M NaCl. Samples containing JE-VLPs were concentrated and the buffer was exchanged to 50 mM Tris pH 7.4, 0.14 M NaCl using an ultrafiltration unit (MilliPore). Samples were then analyzed by SDS-PAGE on 4–20% acrylamide gels and analyzed by Western blotting using the anti-WNV-E antibody. For samples containing YFV, RNA was extracted from the bead eluates with Trizol (Invitrogen) in the presence of 50 µg/ml yeast transfer RNA (RNase and DNase free; Sigma) as a viral RNA carrier. RNA was quantified by RT-PCR as described below. To control for the surface charges of the PI and PI(3)P beads, we tested binding of 2.5 µg/µl polyarginine (5–15 kDa, Sigma) to each types of bead, using 100 µl beads. The binding, wash, and elution steps for polyarginine were performed as described above for JE-VLPs and YFV. Polyarginine in the eluate was quantified using Bradford reagent.
We employed an established protocol to isolate endosomes from the cytosolic fraction [5], [11], [41], [69]. Briefly, Vero cells (6×106 cells/ml) in DMEM were infected with YFV (MOI = 1) and incubated for 1 h. Cells were grown in 2 g/l HRP (Sigma) for 40 min post-infection. Cells were washed with PBS and harvested by centrifugation at 1500 rpm for 5 min at 4°C. Cells were suspended in homogenization buffer (3 mM imidazole and 8.5% sucrose pH 7.4 plus protease inhibitors (Roche)) and passed through a steel syringe needle 20 times. The nuclear fraction was isolated by centrifugation for 10 min at 1 kg and 4°C. Sucrose was added to the post-nuclear supernatant (PNS) to a final concentration of 40% (w/v). This mixture was overlaid with 35%, 27%, and 8.5% sucrose cushions in 10 mM HEPES pH 7.4. Samples were centrifuged for 1 h at 100 kg in a SW60 Ti swinging-bucket rotor. Late endosomes were collected at the 27/8.5% sucrose interface while early endosomes were collected at the 35/40% interface. Total RNA was extracted with Trizol (as described above) from the load (cytosolic) fraction. The purity and integrity of the purified RNA was determined by OD260/280 and by 1% formaldehyde agarose gel electrophoresis. To confirm that the isolated endosomes were intact, infected cells were grown with HRP for the last 20 min of the infection and the HRP activity of the endosomal or cytosolic fractions was measured after lysis with 1% Tween-20 (Figure S3A). Effective separation of the cytosolic and endosomal fractions was also confirmed by Western blotting with antibodies against Rab5 and Rab7 (Figure S3C).
The 3′ untranslated region downstream primer (3UTR, bases 10109–10128: 5′-AACCCACACATGCAGGACAA-3′) and glyceraldhyde-3-phosphatedehydrogenase downstream primer (GAPDH, bases 1157–1175: 5′-TCCACCACCCTGTTGCTGT3′) were used to reverse-transcribe the viral RNA and cellular housekeeping gene GAPDH, respectively. Quantitative real-time PCR (qRT-PCR) was performed using the same downstream primers and the 3UTR upstream primer (10337–10318 bases, 5′-GTTGCAGGTCAGCATCCACA-3′) and GAPDH upstream primer (bases 724–742: 5′-ACCACAGTCCATGCCATAC-3′). PCR reactions were carried out in triplicate using an RT-PCR kit (Roche) and an ABI 9700HT RT-PCR instrument (Applied Biosystems). The amplified products (228 bp for 3UTR and 450 bp for GAPDH) were identified on 2% agarose gels. RT-PCR products were relatively quantitated with the software SDS. Endogenous GAPDH was used as a control for the quality of the total extracted RNAs. Neither 3UTR nor GAPDH formed primer dimers as judged by the dissociation curve.
Binding of scFv11 to JE-VLPs was analyzed in 50 mM Tris pH 7.4, 0.14 M NaCl, 2 mM β-mercaptoethanol at 25°C, with an iTC200 calorimeter (MicroCal). The sample cell contained 3.1 µM JE-VLPs or buffer only, and the titrant syringe contained 40 µM scFv11. An initial injection of 1.5 µl of scFv11 was followed by 20 serial injections of 2.0 µl scFv11, each at 10 min intervals. The stirring speed was 1,000 rpm and the reference power was maintained at 11 µcal/s. The net heat absorption or release associated with each injection was calculated by subtracting the heat associated with the injection of scFv11 to buffer. Thermodynamic parameters were extracted from a curve fit to the data to a single-site model with Origin 7.0 (MicroCal). Experiments were performed in triplicate.
JE-VLPs, scFv11 and VLP-scFv11 complexes were separated on a Superdex 200 10/300 GL column (GE Healthcare) in 50 mM Tris pH 7.4, 0.14 M NaCl.
Vero cells were loaded with 5 µM of Fluo-4 (Invitrogen) for 15 min in DMEM medium. Cells were either infected with YFV (MOI = 5) or treated with 200 µl JE-VLPs (at 17 pM or 50 ng/ml E protein). As a positive control we used 10 mM ionomycin (Invitrogen). As a negative control we pretreated Vero cells with 25 mM BAPTA for 30 min before loading the cells with Fluo-4. Images were collected at one frame every 2 s. Fluo-4 fluorescence was analyzed with ImageJ.
Supporting Information includes seven figures and three movies and can be found with this article at the Journal's website.
|
10.1371/journal.pntd.0003319 | Progress and Impact of 13 Years of the Global Programme to Eliminate Lymphatic Filariasis on Reducing the Burden of Filarial Disease | A Global Programme to Eliminate Lymphatic Filariasis was launched in 2000, with mass drug administration (MDA) as the core strategy of the programme. After completing 13 years of operations through 2012 and with MDA in place in 55 of 73 endemic countries, the impact of the MDA programme on microfilaraemia, hydrocele and lymphedema is in need of being assessed.
During 2000–2012, the MDA programme made remarkable achievements – a total of 6.37 billion treatments were offered and an estimated 4.45 billion treatments were consumed by the population living in endemic areas. Using a model based on empirical observations of the effects of treatment on clinical manifestations, it is estimated that 96.71 million LF cases, including 79.20 million microfilaria carriers, 18.73 million hydrocele cases and a minimum of 5.49 million lymphedema cases have been prevented or cured during this period. Consequently, the global prevalence of LF is calculated to have fallen by 59%, from 3.55% to 1.47%. The fall was highest for microfilaraemia prevalence (68%), followed by 49% in hydrocele prevalence and 25% in lymphedema prevalence. It is estimated that, currently, i.e. after 13 years of the MDA programme, there are still an estimated 67.88 million LF cases that include 36.45 million microfilaria carriers, 19.43 million hydrocele cases and 16.68 million lymphedema cases.
The MDA programme has resulted in significant reduction of the LF burden. Extension of MDA to all at-risk countries and to all regions within those countries where MDA has not yet reached 100% geographic coverage is imperative to further reduce the number of microfilaraemia and chronic disease cases and to reach the global target of interrupting transmission of LF by 2020.
| The mass drug administration (MDA) programme to eliminate lymphatic filariasis (LF) was initiated in 2000. By the end of 2012, the programme was in place in 55 endemic countries. During these first 13 years (2000–2012) of programme implementation, 6.37 billion annual single dose anti-filarial treatments were offered and 4.45 billion doses were consumed by the target populations. This massive programme is estimated to have prevented or cured 96.71 million LF cases that include 79.20 million microfilaria carriers, 18.73 million hydrocele cases and a minimum of 5.49 million lymphedema cases, a 59% reduction of initial LF levels. It is further estimated that, currently, i.e. after 13 years of the MDA programme, 67.88 million LF cases remain, including 36.45 million microfilaria carriers, 19.43 million hydrocele cases and 16.68 million lymphedema cases. Progressive reduction in this burden is possible as the programme extends to the endemic countries and regions within endemic countries that have not yet been covered by the MDA programme, and if the morbidity management component of the programme can be effectively implemented.
| Lymphatic filariasis (LF) is a disease of the poor that is prevalent in 73 tropical and sub-tropical countries. LF is caused by three species of filarial worms – Wuchereria bancrofti, Brugia malayi and B. timori – and is transmitted by multiple species of mosquitoes. The disease is expressed in a variety of clinical manifestations, the most common being hydrocele and chronic lymphedema/elephantiasis of the legs or arms. People affected by the disease suffer from disability, stigma and associated social and economic consequences. Marginalized people, particularly those living in areas with poor sanitation and housing conditions are more vulnerable and more affected by the disease. Estimates made in 1996 indicated that 119 million people were infected with LF at that time, 43 million of them having the clinical manifestations (principally lymphedema and hydrocele) of chronic LF disease [1].
Earlier severe resource constraints and lack of operationally feasible strategies in the endemic countries left a significant proportion of the LF endemic population living unprotected and exposed to the risk of LF infection. Despite a long-standing and gloomy outlook for these individuals, the situation turned around dramatically in the 1990s for 2 principal reasons: 1) advances made in point-of-care diagnostics and 2) the finding of the long-term effectiveness of anti-filarial drugs given in single doses that permitted development of the strategy of annual two-drug, single-dose mass drug administration (MDA) to control/eliminate LF [2], [3]. As LF had already been postulated to be an eradicable disease [4] and with the success experienced in LF elimination activities in China [5] and elsewhere, the World Health Assembly (WHA) in May 1997 formulated resolution WHA 50.29 urging all endemic countries to increase their efforts and determination to control and eliminate LF. In response, the WHO was able to launch the Global Programme to Eliminate LF (GPELF) in the year 2000, largely because the manufacturers of albendazole (ALB) and ivermectin, two of the principal drugs used in the GPELF MDAs, donated these drugs for as long as needed to eliminate LF [3]. The principal strategy of the programme has been two-fold: 1) to implement MDA programmes in all endemic areas to achieve total interruption of transmission and (2) to provide effective morbidity management in order to alleviate the suffering in people already affected by filarial disease. The GPELF targets elimination of LF, at least as a public health problem, by the year 2020 [6].
The programme to implement MDAs targeting LF (GPELF) completed 13 years of operations in 2012 [7]. With its ambitious goal to eliminate LF by the year 2020, it is essential that progress toward this goal be assessed repeatedly in order to set benchmarks to guide future programmatic planning. How to define and assess this progress remains a challenge, but two strategies have been suggested. The first is to measure reduction in the burden of LF disease (i.e., hydrocele, lymphedema, microfilaraemia and associated subclinical disease) over the past 13 years – i.e., a clinical perspective; the second is to measure reduction in the risk of acquiring infection for populations living in (formerly) endemic areas – i.e., an epidemiologic perspective.
In the present report we have pursued the first alternative – to model the decreased burden of LF (defined for the purposes of our calculations as hydrocele, lymphedema, and microfilaraemia) in order to assess the progress towards LF elimination from inception of the MDA programme through 2012 (i.e., during GPELF's first 13 years). In a parallel study, others have recently modeled the programme's progress from the alternative, risk-of-infection viewpoint (Hooper et al., submitted).
A simple ‘force-of-treatment’ model was formulated to estimate the impact of MDA on LF infection and disease.
The GPELF aims to provide MDA (using ALB+either ivermectin or diethylcarbamazine [DEC]) to entire endemic populations at yearly intervals for 4–6 years. Such a programme, if implemented effectively (i.e. treating at least 65% of the total population during each MDA), is expected to interrupt transmission and eliminate LF [8]. Because the status of MDA activities in all of the 73 endemic countries at the time of this analysis (through 2012) ranged from no MDA at all in some countries to full completion of the MDAs in others, for the present study each country was evaluated separately. First, programme impact was determined for each endemic country; then, the burden of LF remaining in each of the five endemic WHO regions – Southeast Asia (SEAR), Africa (AFR), Western Pacific (WPR), Eastern Mediterranean (EMR) and America (AMR) - was calculated by summing the remaining LF burden for all the endemic countries within each region.
Calculating progress of the MDA programme under GPELF – whether by burden or risk estimates – is affected by a number of important specific factors, namely; (1) the number of countries that have successfully completed implementing the MDA programme, (2) the number of countries currently implementing the programme and the geographical coverage or proportion of the endemic population targeted so far in each country, (3) the treatment coverage of the population targeted for MDA in each country, and (4) the duration of the programme (i.e., the number of rounds of MDA implemented) in each country. For the present analysis, all of these data have been sourced from the WHO PC data bank [9].
There are 3 essential steps to assessing the decrease of LF burden since 2000: first, the establishment of the LF base-line burden (in 2000); then, estimation of the MDA impact for countries or IUs where MDAs have taken place during 2000–2012; and, finally, estimation of current burden for countries or IUs where no MDA has taken place.
Treatment of LF has been shown to be especially effective and beneficial in children. Prevalence and intensity of childhood infections are relatively low [34], [35], and MDA is particularly effective in clearing them [14], [17], [18], [36]. Assessment carried out after two rounds of MDA suggests that the treatment is able to clear infection in 0–5 year age children [14], [18]; children of 1–10 year age were shown to become free from infection after 2–4 rounds of MDA [18], [20]; and, further, single dose treatment can reverse lymphatic pathology in children [36]. Also, since the MDA exerts an impact on transmission from the first treatment round itself, it offers excellent protection to newborns from acquiring LF [14], [15], [17], [18], [20], [37]. Therefore, for all these reasons the present analysis has considered that the children of 0–5 years in the communities that received one or more MDAs will be free from microfilaraemia and disease. In addition, the children of 0–10 year age in the communities with TI of ≥3 (equivalent to receiving about four rounds of MDA) were considered free from microfilaraemia and disease. Therefore, the impact of the MDAs on LF burden has been treated separately for children and adults.
GPELF had a modest start – only 14 of the 81 countries then identified as endemic were able to develop and implement MDA programmes in 2000, the first year of operations, and the target population was 3.2 million. Nevertheless, the programme scaled up progressively, so that by 2005, national programmes were in place in 42 countries with a target population of 610 million [38]. During the subsequent years, further progress has been made. In 2011, 9 countries with a previous history of low prevalence were re-evaluated and declared non-endemic, leaving 73 countries with a combined endemic population of 1,459 million. By 2013, 13 of the 73 endemic countries had completed the MDA phase of the programme and entered into the post-MDA surveillance phase, 42 countries were implementing the programme, but 18 countries still had no programme in place. These 18 countries – 15 of them in the Africa region - account for about 10% of the global endemic population of 1,459 million still living in 73 endemic countries [9].
The status of the programme in terms of number of treatments offered and consumed, as of 2012 in different regions, is summarized in Table 2. Of the 1,459 million endemic population, 975 million individuals (67%) have been targeted by 2012. The 975 million population has been offered a total of 6.37 billion treatments during 2000–2012. The distribution of treatments is noticeably uneven among the two major endemic regions, Africa and South-East Asia. Whereas Africa has 32% of the endemic population, it accounts for only 13% of the total treatments offered, while South-east Asia is home to 62% of the endemic population but accounts for 82% of the treatments offered (Table 3). India alone, with 42% of the endemic population accounts for 71% of the total global treatments offered to date. Of the total 6.37 billion treatments provided, 4.45 billion or 70% of treatments were reported as consumed by the endemic populations.
In addition to the 18 countries that had not yet started the programme by 2012, there were also several regionally major endemic countries that had initially launched their programmes but then progressed slowly, principally because of logistic difficulties, funding challenges, lack of political support, civil strife, or, in the case of many Central African countries, the coexistence of loaisis, a contraindication for treating LF with the standard MDA drug regimens [39]. These large countries (including Nigeria, Tanzania, Kenya, Sudan, Papua New Guinea and Indonesia) have an endemic population of 398 million and account for 27% of the global endemic population. (Many of these countries have accelerated their programmes significantly since that time).
Prior to the GPELF, efforts to control LF met with little success, largely because of the lack of feasible and affordable strategies. Even most of the countries that initiated control programmes in the 1950s could make only marginal progress because of the relatively low priority for LF control and lack of feasible, scalable control strategies. The advent of preventive chemotherapy-based annual MDA programmes and the launching of GPELF provided great stimulus toward the control and elimination of LF and its very significant health and socio-economic consequences. Single dose treatment was shown to be very effective against LF infection [2], and mass administration of such single dose treatment was shown to be both broadly feasible [43] and comparatively inexpensive [44], [45]. Availability of donated drugs [46] and the implementation support by international organizations and aid agencies [3], [47] provided further impetus to launch the MDA programme. These factors have enabled as many as 55 countries to undertake national MDA programmes targeting LF elimination. In these countries, an unprecedented 6.37 billion treatments were made available during 2000–12 period [9], making the preventive chemotherapy for LF elimination one of the largest ever public health interventions. The scale of the programme also highlights not only the positive response of endemic countries to accept the challenge of implementing interventions that are ‘simple’ and feasible but also the ability of these countries – some of them among the least resourced – to implement these very large-scale public health programmes successfully.
Given all of this implementation success, it is now essential that the disease-specific health impact of these programmes be assessed as well. While there are, indeed, many important clinical consequences of LF infection (including renal pathology [48], acute episodic ADL [41], [42], [49] and others [50], because the manifestations most frequently measured are microfilaraemia, hydrocele and lymphedema/elephantiasis, it is these that we have tracked in modeling GPELF's impact on the burden of LF disease.
LF infection in individuals goes through different phases, beginning with pre-patent infection, then progressing to microfilaraemia, acute manifestations and chronic disease. The anti-filarial drug regimens used in the GPELF – ALB+either DEC or ivermectin – exhibit excellent microfilaricidal effect even in single doses at both the individual and community level [12]–[21]. Hence, as expected, thirteen years of an MDA programme that delivered 6.37 billion treatments with an intake of 4.45 billion treatments (Table 2), has prevented or cured an estimated 79.20 million microfilaraemia cases in the endemic countries. Currently, as projected in this study, there are still an estimated 36.45 million Mf cases, a figure that is still high but that would have been an astounding 115.65 million cases, had there not been an MDA programme under GPELF (Table 4). This also means that the consequences of microfilaraemia, which include LF progression to chronic disease in a proportion of those 79.20 million people, were averted as well (see below).
The direct effects of treatment with anti-filarial drugs are less remarkable against chronic disease manifestations than on microfilaraemia. However, several studies have shown that treatment does, indeed, have significant impact on chronic disease manifestations, ranging from reversal of early disease signs and symptoms to actual reversal of some of the chronic lesions. The presence of adult worms alone is sufficient to cause hydrocele [50] and reduction in adult worm burden is understandably able to lead to reduction in hydrocele prevalence. The anti-filarial drugs used in the MDA programme - albendazole plus ivermectin, as well as DEC alone or with ALB - exhibit at least partial adulticidal effect, thereby reducing the adult worm burden [51], [52] and hydrocele prevalence in treated individuals [24]–[28]. When the relationship between treatment doses and the reduction in hydrocele prevalence (Fig. 2) was extrapolated to the MDA programme, a reduction of 18.73 million hydrocele cases was projected (Table 4) - reflecting both the prevention of new hydrocele cases, particularly in the younger population, and the cure of hydrocele in a proportion of those older, already affected individuals.
Relatively fewer studies have examined the impact of single- or repeated, annual single-dose treatment on lymphedema and elephantiasis. In Indonesia and Tahiti very high reduction i.e. 68% to 80% in lymphedema prevalence was observed after 82 mostly weekly doses and 12 monthly doses respectively [24], [30]. However, the impact of typical annual MDA was critically evaluated only in two studies, one each in India and Papua New Guinea. The reductions were 14% after 7 rounds of MDA in the Indian study using DEC alone [29], and 20% after 4 rounds of MDA, using DEC alone in Papua New Guinea [13]. Taking various studies into account, we assumed conservatively that in communities with TI of 3 and above, which is equivalent to nearly four rounds of MDA, a 14% reduction in lymphedema prevalence is achieved. This conservative approach was adopted not only to avoid overestimation of the programme impact but also because most of the MDA implementing countries have not yet established robust national morbidity management programmes, whose benefits on disease-improvement will be substantial from controlling the bacterial superinfection of affected limbs that is essential to the progression of elephantiasis [50]. Our analysis suggests that, even despite this conservative modeling approach, an estimated 5.49 million lymphedema cases were prevented or cured by the MDA programme in its first 13 years (Table 4). While those born during and after transmission has been interrupted will have no risk of lymphedema, from a practical standpoint it will still be essential to institute morbidity management programmes in order to achieve significant relief for those already affected.
The estimated disease-specific impact of 13 years of the GPELF (Table 4) has been calculated on the basis only of microfilaremia, hydrocele and lymphedema/elephantiasis, but it is clear that other very significant effects on reducing LF burden have been achieved as well. For example, 79.2 million cases of microfilaremia were projected to have been averted by the Programme (see above), and since nearly 50% of Mf carriers show renal abnormalities which resolve with treatment [48], several million Mf carriers can be recognized to have benefited from resolution of such renal abnormalities as well. Also, since the transmission of LF is generally proportional to the number of Mf carriers and the intensity of microfilaraemia in communities [53], such a significant reduction in the number of Mf carriers also means considerable decrease in transmission of LF in the treated communities; and, of course, transmission reduction and its ultimate interruption determine the elimination of LF, the principal objective of the MDA programmes. Similarly, the projected reduction in chronic LF cases – 18.73 million hydrocele cases and 5.49 million lymphedema cases– is estimated to have averted 39 million acute ADL episodes in endemic areas. This is expected to result in significant relief to the infected population, as ADL, though transient, inflicts severe suffering, makes affected people bed ridden [41], [42], [54]–[56] and requires recuperation from these episodes often extending for weeks at a time.
In an earlier study [57], it was estimated that eight years of MDA, under which >1.9 billion treatments were delivered, prevented 7.4 million cases of hydrocele and 4.3 million cases of lymphedema. While these estimates on the number of hydrocele cases prevented are similar to the estimates in the present study, there is less agreement on the number of lymphedema cases prevented. The estimated 5.49 million lymphedema cases prevented in this study, after 13 years of MDA and delivery of 6.37 billion treatments, was lower, likely because of both the different strategies for calculating the effects and the conservative approach adopted in assessing the impact of MDA on lymphedema. The estimated 5.49 million lymphedema cases prevented in this study was a minimum number, and the actual reduction may be much higher.
Of the various factors influencing the outcome of MDA programmes, treatment coverage is particularly important [8]. In this study, the impact of MDA was assessed using the reported treatment coverage – i.e. the treatment coverage reported by the country level programme managers and compiled in WHO's PC data bank [9]. There are, however, a number of reports suggesting that the programme-reported treatment coverage in the South-east Asia region, particularly in India, may be higher than the actual treatment coverage in the communities. For example, while programme-reported treatment coverage in India was generally in the range of 58% to 90%, various independent studies showed treatment coverage that varied widely and ranged from <20% to >90% in different parts of the country [58]–[74]. The data from these published studies give rise to an average ‘evaluated’ treatment coverage rate of 51.0%, less than the 71.33% average reported national coverage [9]. Since the TI used to calculate programme impact in our model incorporates programme coverage, it is necessary to understand the effect of this difference between reported and evaluated coverage. For India, the TI based on reported coverage was 5.27, but only 4.21 when based on ‘evaluated’ coverage – a difference of 20%. Interestingly, however, when those different TI's were applied to the model (Figs. 1 & 2), the effect was minimal, because for TI's >4, little or no additional benefit was achieved on the 3 parameters measured (microfilaraemia, hydrocele, lymphedema/elephantiasis). In other words, the initial rounds of MDA will exert greater impact on these manifestations compared to later rounds, a finding already reported empirically and shown in various studies [12], [13], [15], [17]–[20]. However, if the treatment coverage rate is high, a higher TI can be achieved in the early rounds of the programme, and fewer rounds of MDA may be required to maximize both impact and cost-effectiveness.
It is possible that preventive chemotherapy as well as other interventions implemented against other vector-borne diseases have added to the impact of LF MDA and caused further reduction in LF burden in some countries. Principal among these other interventions are the ivemectin distribution under the African Programme for Onchocrciasis Control (APOC) and the malaria control measures of insecticide treated nets (ITN) and indoor residual spraying (IRS). Currently, ivermectin is distributedfor onchocerciasis control in as many as 26 countries in Africa, covering nearly 130 million population [75]. Most of the 26 countries are co-endemic for LF also and while less than half of this LF-endemic population is under specific treatment as part of the GPELF, many are likely receiving benefit from the ivermectin being used for onchocerciasis control, as has been demonstrated specifically in a number of countries in West Africa [76]–[80]. Similarly, the malaria control measures have been shown to reduce LF transmission considerably and remain promising adjuncts to the MDA of the GPELF activities [81]–[83].
While these coincident intervention measures have, and will continue to have, positive impact on the LF elimination efforts, quantification of their impact remains a daunting challenge. The reduction in LF burden achieved during the GPELF's first 13 years is almost certainly higher than shown through our analyses both because of the additional, on-going intervention measures and because of our conservative approach to estimating the impact on chronic disease.
Though, there can be little question that impressive gains in decreasing LF burden have been achieved as a result of 13 years of MDA in the GPELF, still, however, a considerable burden of LF remains – estimated at 36.45 million Mf cases, 16.68 million cases of lymphedema and 19.43 million cases of hydrocele (Table 4). Extension of MDA to all at-risk countries and to all regions within those countries where MDA has not yet started is absolutely necessary to reduce the number of microfilaraemia cases and transmission. Such an extension of MDA will also reduce a proportion of hydrocele and lymphedema cases, but the burden of LF disease needs also to be approached directly. Techniques for effective morbidity management – both medical and surgical – are available but not nearly so widely implemented as they could or should be. The present model's calculations take into consideration only those burden-reducing benefits coming pari passu with MDA implementation. When appropriate morbidity management strategies are finally introduced and accelerated, the burden of LF disease will fall even more dramatically (and the model can be adapted accordingly).
|
10.1371/journal.pgen.1003739 | Conserved Translatome Remodeling in Nematode Species Executing a Shared Developmental Transition | Nematodes of the genus Caenorhabditis enter a developmental diapause state after hatching in the absence of food. To better understand the relative contributions of distinct regulatory modalities to gene expression changes associated with this developmental transition, we characterized genome-wide changes in mRNA abundance and translational efficiency associated with L1 diapause exit in four species using ribosome profiling and mRNA-seq. We found a strong tendency for translational regulation and mRNA abundance processes to act synergistically, together effecting a dramatic remodeling of the gene expression program. While gene-specific differences were observed between species, overall translational dynamics were broadly and functionally conserved. A striking, conserved feature of the response was strong translational suppression of ribosomal protein production during L1 diapause, followed by activation upon resumed development. On a global scale, ribosome footprint abundance changes showed greater similarity between species than changes in mRNA abundance, illustrating a substantial and genome-wide contribution of translational regulation to evolutionary maintenance of stable gene expression.
| Working with a set of four related animal species, we have studied a conserved developmental and metabolic transition at the level of protein production and regulation of RNA levels. Strikingly, regulatory effects at the level of RNA accumulation and protein synthesis act together to achieve the observed metabolic shift. In addition to a general conservation of the underlying basis for the regulation of individual genes, alterations of these two processes—mRNA production and protein synthesis—can compensate for one another during evolution to maintain stable amounts of functional gene products. A salient feature of the observed regulation was the storage of idle mRNAs encoding key members of the protein synthesis machinery during metabolic arrest (diapause). Maintenance of this pool facilitates re-activation upon feeding, with the rapid regeneration of protein synthesis capacity an early and critical function during adaptation to a major metabolic shift.
| Animals of diverse genera react to unfavorable growth conditions by entering developmentally arrested states known as diapause [1], [2]. Nematodes of the genus Caenorhabditis can enter and exit diapause at several developmental time points, allowing populations to reproduce through boom and bust cycles of nutrient availability. At least four specific programs of developmental arrest and resumption have been identified, each accompanied by unique morphological and gene-regulatory responses [3]–[6]. In newly-hatched Caenorhabditids, entry and exit from L1 diapause can be controlled in large and synchronous populations by depriving or providing food. C. elegans L1 diapause responses have been well characterized at the level of mRNA biogenesis, with developmental state changes associated with substantial transcriptional changes, the accumulation of RNA polymerase at gene promoters, and alternative splicing [7], [8].
Translational regulation is also expected to contribute significantly to major developmental transitions. We selected four nematode species for investigation of the translation and mRNA-level gene regulatory program associated with L1 diapause exit: two hermaphroditic species, Caenorhabditis elegans and C. briggsae, and two gonochoristic (male/female) species, C. remanei and C. brenneri (Fig. 1A). These four species exhibit highly similar morphologies and developmental timing despite significant genomic sequence divergence [9]. For each species, we applied mRNA-seq and ribosome profiling [10] to populations of arrested L1 diapause larvae and to populations harvested three hours after the first food encounter that signals the animals to exit diapause and commence development (Fig. 1B; sequencing summary statistics: Table S1). Samples from the two conditions are denoted “diapause” or “developing” throughout. All sequence reads and processed count data are available from the Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo) via accession number GSE48140.
Our data provide gene-by-gene measurements at two levels of expression: (i) mRNA-seq data measures the relative steady-state abundances of mRNAs in the transcriptome (abundances are products of mRNA biogenesis and decay), (ii) ribosome profiling allows the counting of ribosome-protected fragments (RPFs) derived from each mRNA, with each fragment corresponding to one active ribosome and thus to an instance of peptide synthesis [10]. For brevity, we use “translatome” [11] to describe the population of mRNA fragments undergoing translation at a given time point, with the relative abundances captured by RPF counts. We emphasize that the abundance of RPFs reflects the input of two biological parameters: steady-state mRNA abundance and ribosome binding (translation efficiency). As diapause entails a substantial decrease in overall translational activity [12], it is additionally important to note that all measurements we infer from the data are relative. Changes in RPF levels between L1 diapause and developing states thus represent relative changes in the commitment of available translational resources between the two conditions.
As a starting point for our analysis, we sought to examine the general character of the gene expression changes associated with the transition from L1 diapause to development. mRNA-seq measurements indicated that diapause exit triggered substantial remodeling of transcriptome composition in the four species, similar to the response previously described in C. elegans [7], with thousands of differentially expressed transcripts and expression changes spanning three orders of magnitude (Fig. 2A, Fig. S1). As expected, transcriptome changes (changes in steady-state mRNA levels) were accompanied by dramatic shifts in the composition of the translatomes (Fig. 2B, Fig. S1). The combination of transcriptome and translatome data from common samples allowed us to compare of the relative magnitude of the two levels of response. Fig. 2C shows a comparison of the frequency distributions of mRNA and RPF abundance changes for the four species. In each case, we observed a significantly broader distribution for RPF changes, consistent with a regulatory response in which changes in translation efficiencies and mRNA levels taken together constitute a larger magnitude overall response than that seen at the level of mRNA abundance alone (Fig. 2C, p<2e-16 for all comparisons). Overall, between the four species we found that ∼15–30% of well-expressed transcripts showed a >2-fold change in relative mRNA abundance, and ∼30–45% of transcripts changed >2-fold in relative RPF abundance.
The transcriptome-wide tendency for RPF level changes to exceed mRNA changes could in principle have resulted from (i) translational changes that act synergistically with and amplify mRNA abundance changes, or (ii) translational changes of large magnitude whose directions are somewhat or predominantly independent of the direction of mRNA abundance changes. While the first scenario may seem most likely from a cellular energetic perspective, genome-wide studies in a variety of systems have revealed varying degrees of coordination (and lack of coordination) of transcriptome and translational responses to a range of stimuli [11], [13]–[15]. To determine which of these scenarios best match our data, we compared changes in RPF level to changes in mRNA level on a gene-by-gene basis. As RPF levels represent the combined input of processes affecting mRNA abundance (e.g., transcription, decay) and translational regulation, we reasoned that, for a given transcript, a change in RPF level in the same direction and of a greater magnitude than the change in mRNA level represents a case in which translation and mRNA abundance processes act in concert (a “concordant” change; equivalent to “homodirectional” in [15]). Conversely, if the change in RPF level is of lesser magnitude or in the opposite direction to the change in mRNA level, this represents a situation in which translational regulation is acting in opposition to mRNA abundance processes (a “discordant” change; Fig. 3A). Comparing transcript-wise changes in mRNA and RPF levels from our data, we found that concordant changes were overwhelmingly favored over discordant changes in the four species by a ratio of 2.8–3.7∶1 (Fig. 3B–E). Thus during the feeding-induced transition from L1 diapause to active development, these four species apparently utilize a shared regulatory logic: an amplified global gene expression response produced by synergistic changes in mRNA abundance and translational control.
We next asked to what degree expression changes were similar between species at the gene level. Studies of gene expression conservation have largely focused on mRNA levels [16]–[20], though several groups have reported superior conservation of orthologous protein abundances, implying an important role for translational or post-translational mechanisms in maintaining optimal gene expression during evolution [21], [22]. We began by comparing feeding-induced changes in mRNA abundance for ortholog pairs identified between the four species. mRNA abundance changes for well-expressed transcripts correlated strongly in all pair-wise species comparisons (Fig. 4A, Fig. S2). Observed correlation coefficients were highly significant and ranged from 0.63 to 0.74 (Spearman's rho, Fig. 4B, E). We next examined the between-species correlation of changes in RPF levels. For each pair-wise species comparison, we observed significantly stronger correlations for changes in RPF abundance than for changes in mRNA abundance, with correlation coefficients in the range of 0.76 to 0.85 (Fig. 4C–E, Fig. S2). To complement pair-wise correlations, we examined expression changes within ortholog groups for which an ortholog could be assigned in each of the four species (the four genes together constituting a “four-way” ortholog group). We found that overall expression divergence within four-way ortholog groups was significantly greater for mRNA abundance changes than for RPF changes (Fig. 4F, p<2e-16), and that this difference disappeared after randomly shuffling ortholog groupings (Fig. S3). Together, these results demonstrate that, for the L1 diapause program in these nematode species, comparisons accounting for translational regulation reveal a greater level of overall gene expression conservation than is observed at the level of mRNA abundance alone. A significant resulting inference is that alterations in translational control and processes affecting mRNA abundance can compensate for one another during evolution to achieve stable protein expression.
The substantial observed conservation underscored the functional significance of expression changes during L1 diapause exit. We therefore sought to compare general properties of gene expression in the arrested and developing states, and specifically to identify features that distinguished the two states. Principal components analysis (PCA) is a technique that takes data featuring many variables, such as gene expression data, and extracts a series of linear combinations of the individual variables (called a “principal component”) that explains a substantial fraction of the variance between samples, with each successive component explaining less variance than the previous component. We applied PCA separately to ortholog abundance data from mRNA-seq and RPF datasets, including all arrested and developing samples from the four species in the analysis. For mRNA-seq data, the first two principal components provide poor separation of samples by condition or species (Fig. 5A). For RPF data, we observe a clear separation of diapause samples from developing samples on both of the first two principal components (Fig. 5B, Fig. S4). The clean separation achieved with RPF data suggests that translatomes of animals from different species in the same nutritional/developmental state are more similar than translatomes from animals of the same species in opposite states.
For RPF data, the first principle component explains a majority of the between-sample variance (86.4%), and higher values on this component are associated with developing samples. We identified a set of transcripts that were exceptionally highly weighted on the first component. Overlaying these transcripts on RPF fold-change plots revealed that these transcripts largely corresponded, in each species examined, to a cluster of highly-expressed and strongly up-regulated genes independently identified (by visual inspection) as a group of genes of interest (Fig. 5C, Fig. S5). We investigated the identities of the transcripts making up this group and found that the significant majority corresponded to ribosomal protein genes, along with several core translation factors and a small number of additional genes including ubiquitin, heat shock proteins, and the RACK1 homolog (Table S2). These transcripts also formed a readily-identifiable cluster in mRNA-seq data, but with substantially weaker up-regulation (Fig. 5C, Fig. S5). Direct comparison of fold-changes for ribosomal proteins in mRNA and RPF data showed that up-regulation was significantly stronger at the RPF level, with average fold-changes >10 compared to ∼2-fold up-regulation at the mRNA level, indicating that the differential representation of these genes in the translatome was primarily due to translational regulation (Fig. 5D). We also found that non-ribosomal components of the core translation apparatus showed a significant trend towards up-regulation (Fig. 5C, red), with contributions from both increased mRNA abundance and translation efficiency (Fig. 5E).
Genes of the translation apparatus include many of the most highly-expressed transcripts. The coordinated up-regulation of these transcripts thus constitutes an extraordinary re-allocation of cellular energetic resources. From the raw counts for mapped RPF reads, we infer that, in C. elegans, approximately 3% of ribosomes are bound to a ribosomal protein transcript during L1 diapause. Three hours after feeding, more than 20% of all ribosomes are engaged in translating ribosomal proteins (Table S3). For the translation apparatus as a whole, this figure jumps from ∼4.5% in L1 diapause to nearly 30% in fed animals (Table S3). This striking change suggests that a central feature of the gene regulatory response to L1 diapause exit is to prioritize existing translational resources to building up the animal's capacity for protein synthesis.
In addition to translation genes, several categories of functionally related genes were enriched among transcripts whose RPF levels were significantly higher in developing animals. These include genes involved in promoting growth, development, ribosome biogenesis, the proteasome, and mitochondrial genes (Table S4). These enrichments were exceptionally consistent between the four species examined (Fig. S6A). In contrast, functional enrichments among genes with higher RPF levels in the L1 diapause state were generally weaker and less-consistent between species than those seen for transcripts up-regulated after feeding (Table S4, Fig. S6B). Manual inspection revealed a number of interesting genes that showed conserved higher expression during diapause, including nhr-49 (required for adult reproductive diapause [3]–[6]), genes involved in autophagy and dauer formation, superoxide dismutases, and several heat-shock proteins (Table S5).
Ribosomal protein genes were subject to qualitatively similar regulation in each of the four species examined, i.e., modest increases in mRNA abundance and strong increases in translation efficiency. This led us to ask whether other groups of functionally related genes tended to share regulatory patterns across species. To this end, we defined the “translational component” of regulation as the ratio of the change in translation efficiency to the change in mRNA abundance (see Materials and Methods and Text S1). Examining the distribution of log(2)-transformed translational component scores for differentially-expressed transcripts in the four species revealed unimodal distributions in which a majority of transcripts (71–91%) were subject to “mixed” regulation, with mRNA abundance and translational regulation each accounting for at least 25% of the change in RPF level (Fig. 6A). Despite the overall similarities between the species, species-specific differences were evident. Notably, the broader distribution evidenced in C. briggsae indicated a greater tendency for transcripts to be primarily regulated either translationally or at the level of mRNA abundance, while the narrow distribution of C. brenneri suggested a trend toward coordinated regulation (Fig. 6A). While the results indicate a potential quantitative difference in the balance between transcriptional and translational regulation in different species, the coordination between these two regulatory modalities is evident in all four species.
We examined the distribution of translational components for transcripts corresponding to up-regulated functional gene categories and found that these categories showed remarkably similar profiles across the four species (Fig. 6B). For example, transcripts corresponding to proteasome components showed minimal contributions from translational regulation in each species, whereas ribosomal components, as demonstrated previously, exhibited large translational components (Fig. 6B). Likewise, spliceosome components favored strong contributions from mRNA accumulation changes in every species, while transcripts of the non-ribosomal translation machinery showed a broad distribution, indicating varied contributions from mRNA abundance and translation for this category. The between-species similarity of the relative contributions of mRNA abundance and translational control to regulation of functionally related transcripts suggests that the transcriptional and translational control networks underlying these changes may also be conserved in these species.
Our results point to a key role for translational control in the transition from L1 diapause to active development in these Caenorhabditis species. Translational regulation affects the expression of thousands of transcripts, and the patterns of regulation are well conserved between species at the genome-wide, functional, and gene level. Highly conserved translational modulation of certain sets of related transcripts, notably the ribosomal protein genes, implies that translational control programs may remain largely intact despite significant genome sequence divergence. A recent study reported the persistence of a pool of translationally repressed ribosomal protein mRNAs in yeast undergoing glucose starvation [13], suggesting that translational suppression of this class of genes may be a somewhat generalized feature of eukaryotic stress responses. In summary, we describe a system in which evolutionarily diverged species maintain a common program of mRNA abundance and translational efficiency changes that cooperatively drive the dynamic reallocation of gene expression resources to traverse a shared developmental and environmental transition.
Strains were obtained from the Caenorhabditis Genetics Center: C. elegans N2 bristol, C. briggsae AF16, C. remanei PB4641, C. brenneri PB2801. Embryos were hatched in sterile S-complete liquid media, starved for 24 hours, and half were supplied with E. coli HB101. Samples were frozen in liquid nitrogen after 24 hours of starvation and after three hours of feeding. Three full biological replicates were prepared for each species. mRNA-seq and ribosome profiling were carried out as described in [23], [24], with modifications as described in Text S1. Sequencing was performed on Illumina's HiSeq 2000 machine.
Raw sequence reads were trimmed of adaptor sequence and mapped using Bowtie 0.12.7 [25] to the appropriate species' genomes and coding sequence with genomic flanking sequence, and screened for quality. For between-species comparisons, count normalization was performed with the EdgeR package [26] and orthologs were assigned using inParanoid [27]. Differential expression was determined using the DESeq package [28]. All additional analysis was carried out with custom Perl scripts and using the R computing environment [29]. Figures were created with R [30], [31].
Expression divergence for four-way orthologs was calculated by first normalizing log fold changes for each species by mean and standard deviation, then calculating each pair-wise species-species difference and taking the mean of the resulting differences. Principal components analysis was carried out using the prcomp function in R. Ontology analysis was performed using the web-based DAVID knowledge tool [32].
A more extensive description of the methods can be found in Text S1.
|
10.1371/journal.pntd.0006454 | Adverse events following single dose treatment of lymphatic filariasis: Observations from a review of the literature | WHO’s Global Programme to Eliminate Lymphatic Filariasis (LF) uses mass drug administration (MDA) of anthelmintic medications to interrupt LF transmission in endemic areas. Recently, a single dose combination of ivermectin (IVM), diethylcarbamazine (DEC), and albendazole (ALB) was shown to be markedly more effective than the standard two-drug regimens (DEC or IVM, plus ALB) for achieving long-term clearance of microfilaremia.
To provide context for the results of a large-scale, international safety trial of MDA using triple drug therapy, we searched Ovid Medline for studies published from 1985–2017 that reported adverse events (AEs) following treatment of LF with IVM, DEC, ALB, or any combination of these medications. Studies that reported AE rates by treatment group were included.
We reviewed 162 published manuscripts, 55 of which met inclusion criteria. Among these, 34 were clinic or hospital-based clinical trials, and 21 were community-based studies. Reported AE rates varied widely. The median AE rate following DEC or IVM treatment was greater than 60% among microfilaremic participants and less than 10% in persons without microfilaremia. The most common AEs reported were fever, headache, myalgia or arthralgia, fatigue, and malaise.
Mild to moderate systemic AEs related to death of microfilariae are common following LF treatment. Post-treatment AEs are transient and rarely severe or serious. Comparison of AE rates from different community studies is difficult due to inconsistent AE reporting, varied infection rates, and varied intensity of follow-up. A more uniform approach for assessing and reporting AEs in LF community treatment studies would be helpful.
| WHO’s Global Programme to Eliminate Lymphatic Filariais (LF) supports annual mass drug administration to over 400 million people in LF-endemic areas each year. Two drug combinations (either DEC or ivermectin, given with albendazole) have been recommended in most endemic areas. With the exception of well-described serious adverse events (AEs) occurring in patients with high level loiasis, severe AEs due to these medications are extremely rare. Mild to moderate AEs, however, are common, particularly in patients with active filarial infection. In this manuscript we synthesize published data on AEs following single-dose treatment of LF with ivermectin, DEC, or albendazole. This provides a background against which to compare the safety of triple drug therapy (ivermectin, DEC, and albendazole) recently endorsed by WHO, and provides a useful context for evaluating safety of new treatments for LF. The compiled data illustrate that transient, mild to moderate AEs following single-dose LF treatment are common in microfilaremic patients and are much less common in amicrofilaremic patients. They also show that passive surveillance for post-treatment AEs underestimates AE incidence and suggest that adherence to common reporting standards would improve the usefulness of AE reporting in filariasis studies.
| Infection with the filarial nematode parasites Wuchereria bancrofti, Brugia malayi, or Brugia timori is known as lymphatic filariasis (LF). These infections cause severe, disabling conditions including lymphedema, elephantiasis, and hydroceles in tens of millions of people in tropical and subtropical countries. Annual mass drug administration (MDA) coordinated by WHO’s Global Programme to Eliminate LF (GPELF) has significantly reduced LF transmission in many of the 78 initially endemic nations [1–3]. Yet LF remains far too common, with tens of millions infected and 850 million people at risk of acquiring the infection in 53 countries [3]. With approximately 500 million people receiving MDA for LF each year, understanding, anticipating, and preparing the targeted population for MDA-related adverse events (AEs) is important for program success.
Medications used for MDA include diethylcarbamazine (DEC), ivermectin (IVM) and albendazole (ALB). The combination of IVM plus ALB is used in areas of Africa where onchocerciasis (river blindness) is co-endemic with LF. Twice yearly ALB alone is recommended for LF-endemic areas of Africa that are co-endemic for loiasis, and DEC plus ALB is used in the rest of the world. Serious (life-threatening) AEs due to MDA are exceedingly rare [4–7]. However, when they do occur they can profoundly impact the treated community and jeopardize program success [8]. When communities are well-informed about the type and severity of AEs to be expected, they may be less likely to avoid MDA out of fear of AEs. Furthermore, the knowledgeability of community health workers (drug distributors) can be a major determinant of MDA adherence [8]. A clear understanding of the nature of expected AEs should empower program managers and community health workers to prepare their communities to anticipate and accept transient AEs, which may in turn improve compliance with MDA and facilitate LF elimination efforts.
A promising new combination therapy for LF that combines a single dose of IVM, DEC, and ALB (IDA) appears to be highly effective [9], and its safety is has been evaluated in large community-based studies in several locations (ClinicalTrials.gov Identifier: NCT02899936) [10]. This manuscript’s purpose is to provide context for understanding the safety of the new IDA treatment by reviewing published data on the rates and nature of AEs following single-dose treatment for LF with any of the IDA medications. As previously noted by many others, AE reporting in LF treatment trials is highly variable and potentially affected by multiple factors including blood microfilaria (Mf) counts, treatment regimens, filarial species, population demographics, and importantly, the thoroughness of post-treatment surveillance. We therefore sought to review AE data from published LF treatment studies to further understand the effect of these parameters on AE rates and severity. Our objective was to evaluate reports of AEs following single dose LF treatment of children and adults with IVM, DEC, or ALB (either as monotherapy or in multidrug combination regimens), published since 1985. In this report we first present a broad summary of the literature reviewed and then a quantitative synthesis of published AE rates from studies meeting our specified inclusion criteria.
The primary outcome of interest for the quantitative synthesis was the proportion of participants experiencing at least one AE (aggregate AE rate). Rates of individual AEs were a secondary outcome. We did not use a pre-specified AE definition, but rather accepted all AEs reported by the authors of the individual studies. In this manuscript we classify AEs as mild if they do not interfere with normal daily routine (work or school), as moderate if they interfere with daily routines (work or school) but not with activities of daily living, and as severe if they interfere with activities of daily living or cause temporary incapacitation. These designations correspond to Common Terminology Criteria for AEs grades 1 (mild), 2 (moderate), and 3 (severe). Serious AEs are those that are life-threatening or result in hospitalization or permanent injury (grade 4) or are fatal (grade 5) [11].
We reviewed AE data from studies of LF treatment with single-dose regimens that were published between 1985 and 2017. We searched Ovid Medline and Embase for any articles with Medical Subject Headings (MeSH) terms “Elephantiasis, Filarial” and “Drug Therapy” plus any of the following terms: “Adverse Events”, “Poisoning”, or “Toxicity”. We limited our search to English or French language manuscripts dealing with human infections. The most recent search was conducted on 21 Aug, 2017. Two authors (PB and CH) reviewed each publication and gathered additional pertinent publications from articles referenced therein. Publications with sufficient AE data were selected for a quantitative analysis of AEs as described below. We did not pre-specify nor register a review protocol. We did not attempt to contact authors to identify additional studies.
Studies published after 1985 that reported AE rates following single-dose LF treatment with IVM, DEC, or ALB (alone or in combination) were included. Studies dealing with multi-day courses (generally of DEC) were reviewed, but excluded from the quantitative analysis, as were studies that either provided inadequate information on AEs by treatment arm or did not conduct AE surveillance within one week following treatment. Complete inclusion and exclusion criteria are shown in Table 1. We followed the PRISMA Statement for Reporting Systematic Reviews and Meta-Analysis [12]; the completed PRISMA checklist is available as a supplemental file (S1 Table).
From studies meeting inclusion criteria we extracted data including: study location (country), age range and gender of participants, intensity of surveillance, treatment regimen, Mf prevalence, geometric mean Mf counts, presence of co-infections, overall rate of AEs, and rates for any specific AEs reported. For studies that reported AEs following multiple MDA treatment rounds, we included only the AE rates that occurred after the first treatment. For studies in which one but not all treatment arms met inclusion criteria (for example, when single dose IVM or DEC was compared to 12 days of DEC), we included data only from the arm(s) meeting inclusion criteria. The number of participants reported in our analysis is the number for whom AE surveillance was conducted, which was sometimes lower than the total number treated. For example, one study conducted active post-MDA surveillance within a subset of 483 persons among 8 million people treated [18]; in our analysis, the N for this study was 483.
All extracted data were analyzed using Stata version 12.1 (College Station, TX). Because the data were not normally distributed, we report means and interquartile ranges (IQR) and use boxplots for graphical representation. Since AE reporting was insufficiently uniform among included studies, we did not attempt a formal meta-analysis of AE rates, nor did we attempt statistical analyses. Rather, we sought to present a graphical synthesis of data from these disparate studies to illustrate the range of data and an estimate of central tendency (median and interquartile range). To assess for reporting bias in individual studies, we stratified surveillance for AEs in each study as active (individual participants were contacted and asked about AEs) or passive (individuals with AEs had to seek out the study team to report). The quality of active surveillance was further categorized as “high” (at least daily contact during the first 72 hours), “moderate” (at least one contact within first 72 hours), or “low” (participants contacted after 72 hours). Although we hoped to analyze the effect of each extracted variable on reported AE rates, we found that the quality of data reported for most parameters was insufficient. We therefore limited our analysis to an ad hoc comparison of treatment regimens, Mf status, and intensity of AE surveillance.
Many informative articles that reported AEs following treatment for LF could not be included in our quantitative analysis either because they reported composite AE scores rather than rates, or because they did not report AE rates separately by treatment group. We have attempted to review some of the observations from both included and excluded publications in the following paragraphs.
To summarize published rates of AEs following single-dose treatment of LF, and to explore how these might differ by treatment medication and AE surveillance, we compiled data from articles with adequate AE reporting into a combined analysis. Among 162 full text articles reviewed, 55 contained AE data that met criteria for inclusion (Fig 1). There was considerable heterogeneity in the way that AEs were reported in these studies; 34 reported both the aggregate incidence of AEs (i.e. the number of persons experiencing any AE) and the percentage of persons experiencing specific AEs. Seventeen studies reported an aggregate incidence but not specific events, and four reported specific events but not aggregate incidence. Methods of AE ascertainment varied widely between studies, from intensive in-hospital monitoring to passive reporting in community-based trials. For the purposes of our analyses we grouped the studies into two main types: (1) clinical trials with active AE surveillance and (2) community studies with either active or passive surveillance. The former group comprises studies in which 100% of participants were Mf positive, while community studies had varied Mf rates (Table 2).
In this review we initially sought to quantify the effects of various factors on AE rates that occur following MDA for LF. We quickly realized that the heterogeneity in the way AEs have been reported in the literature would not allow a meaningful quantitative multivariate analysis. We nevertheless felt a compilation of reported AE rates would be beneficial. Despite the limitations of combining data from methodologically disparate studies, we believe the compiled data illustrate the following main points: 1) AEs are very common in microfilaremic patients after single-dose treatment of LF with drugs (IVM and DEC) that rapidly reduce Mf counts. 2) AEs are much less common in amicrofilaremic participants, regardless of treatment regimen. 3) Passive surveillance tends to underestimate the occurrence of AEs, and 4) Heterogeneity in the stringency of AE surveillance and format of AE reporting makes comparisons between studies difficult.
The relationship between AE rates and the prevalence of microfilaremia is illustrated by the striking differences between study arms with 100% microfilaremia and those with no microfilaremia. It would have been interesting to compare Mf prevalence to AE rates among the community studies with varied Mf prevalence; this was not attempted because of the variability in AE reporting for these studies and because uncertainty regarding true Mf rates would have made this comparison unreliable.
It is clear that community proclivities for reporting AEs vary from place to place and study to study. This is perhaps most evident in reported AE rates after placebo treatment. Studies with highly active AE surveillance in Haiti and Tahiti reported high AE rates after placebo treatment [54, 84, 93], but AE rates were low after placebo treatment in Ghana and India [86, 120]. This place-to-place AE reporting variability is also evident in the wide range of AE rates reported among different studies with the same treatment regimens (see Figs 3 and 5). Potential reasons for this might include the prevalence of STH or other helminth infections, differing intensities of LF infection, and varying cultural norms. In addition, where populations have been sensitized to expect AEs following MDA, more AEs may be perceived [18].
Nearly all the studies cited in this review reported AE rates in some manner, but we were only able to include 55 in the quantitative synthesis. The primary reason for excluding studies was that they did not present AE data in a way that linked AE rates to treatment regimens. For example, many studies reported AE severity scores rather than rates. Others reported that AE rates did not differ significantly between treatment groups, but did not report the numbers for each group. When AE rates were reported by treatment group, comparisons were often hampered by non-standardized AE reporting procedures. Some authors did not report the timeframe over which AE surveillance was conducted, making it difficult to surmise whether early or late AEs may have been missed. Although most studies included in our analysis described whether AE surveillance was active or passive, many contained insufficient detail to determine how sensitive the study procedures were for detecting AEs. For example, ascertainment rates (the proportion of participants in community-based studies who were actually visited and queried about AEs) were almost never reported.
This review has several strengths and weaknesses. The primary strength is that it compiles data from 30 years of published studies. It also illustrates how variable AE reporting can be, and it provides a context for interpreting AE rates observed in future LF treatment studies. One weakness was our inability to include data from many high quality studies that did not report AEs by treatment arm. In addition, because we restricted our analysis to studies of single-dose therapy, many rich and highly informative studies that used multi-dose treatment regimens were excluded. In general, the pattern of AEs reported in such studies was similar to single dose studies. That is to say, the rate and severity of AEs increased with increasing Mf counts and most AEs occurred during the first 48 hours after the initial treatment dose [38, 82].
The heterogeneity in AE reporting among the reviewed studies highlights the need for a more structured approach to AE reporting in LF treatment studies. Although this problem is not unique to filariasis [128], it can be compounded by the nature of community-based studies. We therefore suggest the following measures for improving AE reporting in community based treatment trials for LF and other neglected tropical diseases (Box 1). 1) Clearly specify the methods for ascertaining AEs. Indicate whether an attempt was made to contact each participant (active surveillance) or whether participants were required to seek out study staff to report AEs (passive surveillance). Indicate when and how often participants were contacted. Avoid ambiguous language such as “followed closely”, or “closely monitored”. Rather, describe what was actually done. For example, “treated individuals were visited daily in their homes for five days after treatment”. 2) If surveillance was active, report the ascertainment rate; that is, the proportion of participants sought during surveillance that was actually found. Knowing what proportion of participants actually contributed to the reported AE rates will help the reader assess the reported findings. For example, one study reported, “All subjects [were] asked to come to the study site on day 2 and day 5….In addition, team members also went door to door.” The door to door contacting was presumably meant to ascertain AEs in subjects not reporting at the study site, since subjects may choose not to present for follow-up either because they feel well and see no need, or because they feel ill and don’t wish to leave their homes. The higher the proportion of participants for whom actual AE status is not ascertained, the greater the uncertainty regarding the reported AE rates. Unfortunately, the study cited—and most other community studies we reviewed—did not report ascertainment rates. 3) Report numerators and denominators. When severity scores are used (for example, 1 = mild, 2 = moderate, 3 = severe) to compare AEs between study groups, the actual number of persons experiencing AEs should also be reported so that rates can be calculated. The difference between one person with a severe AE and three people with one mild AE each is important, and the reporting of AE data should allow the reader to distinguish the difference. In addition to clearly specifying the number of persons experiencing AEs (the numerator), the denominator should be clearly defined. In active studies, we suggest reporting AE rates as the proportion of those experiencing AEs over the number actually assessed. For example, five persons experiencing AEs among 20 patients treated should be reported as 50% (not 25%) if only 10 of those treated were actually assessed. 4) Use standardized grading criteria for reporting AE severity. Examples include the National Cancer Institute’s Common Terminology Criteria for Adverse Events (available at https://evs.nci.nih.gov/ftp1/CTCAE/About.html) or the Division of AIDS Table for Grading the Severity of Adult and Pediatric Adverse Events (available at http://rsc.tech-res.com/clinical-research-sites/safety-reporting/daids-grading-tables). 5) Follow CONSORT guidelines for better reporting of harms in clinical trials [129].
In conclusion, this review has shown that AEs following single dose treatment of LF are common and should be expected in microfilaremic patients. This information provides a useful context for understanding AEs observed with new treatments for LF. Clear and detailed reporting of AEs in community treatment studies is essential to accurately inform elimination program workers and their communities, and to set appropriate expectations. The fear of MDA-associated AEs is often out of proportion to the actual risk, because most post-treatment AEs are mild and transient. A frank explanation of AEs as a marker for treatment efficacy by program managers and community health workers may improve compliance with MDA and facilitate LF elimination efforts.
|
10.1371/journal.pcbi.1005932 | Bayesian inference of phylogenetic networks from bi-allelic genetic markers | Phylogenetic networks are rooted, directed, acyclic graphs that model reticulate evolutionary histories. Recently, statistical methods were devised for inferring such networks from either gene tree estimates or the sequence alignments of multiple unlinked loci. Bi-allelic markers, most notably single nucleotide polymorphisms (SNPs) and amplified fragment length polymorphisms (AFLPs), provide a powerful source of genome-wide data. In a recent paper, a method called SNAPP was introduced for statistical inference of species trees from unlinked bi-allelic markers. The generative process assumed by the method combined both a model of evolution for the bi-allelic markers, as well as the multispecies coalescent. A novel component of the method was a polynomial-time algorithm for exact computation of the likelihood of a fixed species tree via integration over all possible gene trees for a given marker. Here we report on a method for Bayesian inference of phylogenetic networks from bi-allelic markers. Our method significantly extends the algorithm for exact computation of phylogenetic network likelihood via integration over all possible gene trees. Unlike the case of species trees, the algorithm is no longer polynomial-time on all instances of phylogenetic networks. Furthermore, the method utilizes a reversible-jump MCMC technique to sample the posterior of phylogenetic networks given bi-allelic marker data. Our method has a very good performance in terms of accuracy and robustness as we demonstrate on simulated data, as well as a data set of multiple New Zealand species of the plant genus Ourisia (Plantaginaceae). We implemented the method in the publicly available, open-source PhyloNet software package.
| The availability of genomic data has revolutionized the study of evolutionary histories and phylogeny inference. Inferring evolutionary histories from genomic data requires, in most cases, accounting for the fact that different genomic regions could have evolutionary histories that differ from each other as well as from that of the species from which the genomes were sampled. In this paper, we introduce a method for inferring evolutionary histories while accounting for two processes that could give rise to such differences across the genomes, namely incomplete lineage sorting and hybridization. We introduce a novel algorithm for computing the likelihood of phylogenetic networks from bi-allelic genetic markers and use it in a Bayesian inference method. Analyses of synthetic and empirical data sets show a very good performance of the method in terms of the estimates it obtains.
| The availability of genome-wide data from many species and, in some cases, many individuals per species, has transformed the study of evolutionary histories, and given rise to phylogenomics—the inference of gene and species evolutionary histories from genome-wide data. Consider a data set S = {S1, …, Sm} consisting of the molecular sequences of m loci under the assumptions of free recombination between loci and no recombination within a locus. The likelihood of a species phylogeny Ψ (topology and parameters) is given by
L ( Ψ | S ) = ∏ i = 1 m L ( Ψ | S i ) = ∏ i = 1 m ∫ G p ( S i | g ) p ( g | Ψ ) d g (1)
where the integration is taken over all possible gene trees. The term p(Si|g) is the likelihood of gene tree g given the sequence data of locus i [1]. The term p(g|Ψ) is the density function (pdf) of gene trees given the species phylogeny and its parameters. For example, Rannala and Yang [2] derived this pdf under the multispecies coalescent (MSC). This formulation underlies the Bayesian inference methods of [2–4].
Debate has recently ensued regarding the size of genomic regions that would be recombination-free (or almost recombination-free) and could truly have a single underlying evolutionary tree [5, 6]. One way to overcome this issue is to use unlinked single nucleotide polymorphisms (SNPs) or amplified fragment length polymorphisms (AFLPs). Such data provide a powerful signal for inferring species phylogenies and the issue of recombination within a locus becomes irrelevant. Furthermore, as long as those markers are sampled far enough from each other the assumption of free recombination among loci holds. Indeed, this is the basis of the SNAPP method that was recently introduced in [7]. Since a bi-allelic SNP or AFLP marker has no signal by itself to resolve much of the branching patterns of a gene genealogy, a major contribution of Bryant et al. was an algorithm for analytically computing the integration in Eq (1) for bi-allelic markers.
While trees constitute an appropriate model of the evolutionary histories of many groups of species, it is well known that other groups of species have evolutionary histories that are reticulate [8]. Horizontal gene transfer is ubiquitous in prokaryotes [9, 10], and several bodies of work are pointing to much larger extent and role of hybridization in eukaryotic evolution than once thought [8, 11–15]. Not only does hybridization play an important role in the genomic diversification of several eukaryotic groups, but increasing evidence is pointing to the adaptive role it has played, for example, in wild sunflowers [16], humans [17], macaques [18], mice [19], butterflies [20], and mosquitoes [21, 22].
Reticulate evolutionary histories are best modeled by phylogenetic networks. Two statistical methods were recently introduced for inference under the formulation given by Eq (1), when Ψ is a phylogenetic network [23, 24], and other methods were also introduced for statistical inference of phylogenetic networks using gene tree estimates as the input data [25–29].
The methods of [23, 24] assume that the data for each locus consists of a sequence alignment that has no recombination. In this paper, we devise an algorithm that builds on the algorithm of [7] for analytically computing the integral in Eq (1) when Ψ is a phylogenetic network. In other words, our algorithm allows for computing the likelihood of a phylogenetic network from unlinked bi-allelic markers while analytically integrating out the gene trees for the individual markers. We couple this likelihood function with priors on the phylogenetic network and its parameters to obtain a Bayesian formulation, and then employ the reversible-jump MCMC (RJMCMC) kernel from [23] to sample the posterior of the phylogenetic networks and their associated parameters given the bi-allelic data.
We implemented our algorithm and the RJMCMC sampler in PhyloNet [30], which is a publicly available open-source software package for inferring and analyzing reticulate evolutionary histories. We studied the performance of our method on simulated and biological data. For simulations, we extended the framework of [7] so that the evolution of bi-allelic markers could be simulated within the branches of a phylogenetic network. For the biological data, we analyzed two data sets of multiple New Zealand species of the plant genus Ourisia (Plantaginaceae). The results on the simulated data show very good accuracy and robustness as reflected by the method’s ability to recover the true phylogenetic networks and their associated parameters even when the underlying assumptions of the method are violated. For the biological data, the method recovers two established hybrids and their putative parents correctly.
The proposed method and Bayesian sampler provide a new tool for biologists to infer reticulate evolutionary histories, while also account for the complexity arising from incomplete lineage sorting, from bi-allelic markers, thus complementing existing tools that use gene tree estimates or sequence alignments of the individual loci as the input data. The use of such bi-allelic markers, particularly when they are sampled far enough across the genome, completely sidesteps potential problems that could arise due to the presence of recombination within loci.
A phylogenetic X -network, or X -network for short, Ψ is a rooted, directed, acyclic graph (DAG) whose leaves are bijectively labeled by set X of taxa. We denote by V(Ψ) and E(Ψ) the sets of nodes and edges, respectively, of the phylogenetic network Ψ. Every node of the network has in-degree 1, which we call a tree node, or in-degree 2, which we call a reticulation node. The only exception is special node s whose in-degree is 0 and out-degree is 1; the edge (s, r) defines the branch above the root. The edges whose head is a reticulation node are the reticulation edges of the network; all other edges constitute the tree edges of the network. Every edge is directed forward in time. We assume all phylogenies considered here (trees and networks) are binary—no node has out-degree higher than 2.
Here, we use the bottom of a branch to refer to the end of the branch that is farther from the root of the network, and use the top of a branch to refer to the end of the branch that is closer to the root. Given that the coalescent views the evolution of alleles backward in time, we say that a lineage enters a branch to mean a lineage that exists at the bottom of that branch. Similarly, we say a lineage exits a branch to mean a lineage that exists at the top of that branch.
Each node in the network has a species divergence time and each edge b has an associated population mutation rate θb = 4Nbμ. This parameter is typically referred to in the literature as the (rescaled) population size. Given the length τ of a branch in units of expected number of mutations per site, the length of that branch in coalescent units is 2τ/θ, assuming diploid individuals. The branch above the root, (s, r), is infinite in length so that all lineages that enter it coalesce on it eventually.
For every pair of reticulation edges e1 and e2 that share the same reticulation node, we associate an inheritance probability, γ, such that γ e 1 , γ e 2 ∈ [ 0 , 1 ] with γ e 1 + γ e 2 = 1. We denote by Γ the vector of inheritance probabilities corresponding to all the reticulation nodes in the phylogenetic network.
We use Ψ to refer to the topology, species divergence times, population mutation rates, and inheritance probabilities of the phylogenetic network. That is, here we include Γ as part of Ψ.
An X -phylogenetic tree, or X -tree, is an X -network with no reticulation nodes. A gene tree is an X -tree. Each node in the gene tree has an associated coalescence time. In the algorithm below, we make use of a coloring function c: (E(g), t) → {0, 1}, similar to that used in [7], where c(e, t) indicates the color, or allele, at time t along the branch e of gene tree g. We will follow [7] in calling the two colors red and green.
Looking forward in time (from the root toward the leaves), let u and v be the mutation rate from red allele to green allele and the mutation rate from green allele to red allele, respectively. The stationary distribution of the red and green alleles at the root is given by v/(u + v) and u/(u + v), respectively. Observed alleles are indicated by values of the coloring function c at gene tree leaves.
Given a gene history embedded within the branches of the network, the numbers and types of lineages at both ends of each branch of the network are needed to compute the likelihood. Let x be a branch in the phylogenetic network. We denote by n x T and n x B the total numbers of lineages at the top and bottom of x, respectively, and by r x T and r x B the numbers of red lineages at the top and bottom of x, respectively. See Fig 1 for an illustration.
Let x be an arbitrary branch in the phylogenetic network and let R x be the event that for every external branch z that is a descendant of x, the actual number of red alleles in z equals to r z B.
We define two partial likelihoods: F x B is the product of the likelihood of a subtree rooted at the bottom of x and the probability P r [ n x B = n ], and F x T is the product of the likelihood at the top of branch x and the probability P r [ n x T = n ]. In the case of a species tree (i.e., no reticulation nodes in the species phylogeny), the partial likelihood vectors F x B and F x T are given by [7]
F x B ( n , r ) = P r [ R x | n x B = n , r x B = r ] P r [ n x B = n ] (2)
and
F x T ( n , r ) = P r [ R x | n x T = n , r x T = r ] P r [ n x T = n ] . (3)
Here F x B and F x T are indexed by nonnegative integers n and r, where r ≤ n. Let M be the maximum possible value of n x B and n x T over all branches. Then, each of F x B and F x T has at most l = (1 + (M + 1))(M + 1)/2 entries.
In the case of a species tree, the path from a leaf to the root is unique. However, this might not be the case for phylogenetic networks: If there is a reticulation node on a path from a leaf to the root, then multiple paths exist between that leaf and the root. This is the issue that necessitates modifying the algorithm of [7] significantly, and that leads to much larger computational requirements in the case of phylogenetic networks. The key idea behind the modification is as follows. As the algorithm proceeds to compute the likelihood in a bottom-up fashion from the leaves to the root, whenever a reticulation node is encountered, the current set of lineages is bipartitioned in every possible way so that one side of the bipartition tracks one parent of the reticulation node and the other side tracks the other parent. As the network has a unique root, the two sides of each bipartition eventually come back together at an ancestral node. At that point, these two sides are merged properly.
To achieve this proper merger, we introduce “labeled partial likelihoods,” or LPL. Like the case of [7], LPLs are not “real” partial likelihoods. The reason for this is that when partial likelihood vectors are split (described below), those become symbolic terms that do not evaluate to partial likelihoods until they are merged later. This is analogous to the difference between ancestral configurations on species trees [31] and their labeled counterparts on phylogenetic networks [32], where the latter are in many cases just symbolic terms that do not evaluate to true (partial) likelihood values.
Given a phylogenetic network Ψ with k reticulation nodes numbered 0, 1, ⋯, k − 1, an LPL P is an element of [ 0 , 1 ] l × Z k, where the first element of the pair is a partial likelihood as in [7]. The second element is the label to keep track of partial likelihoods that originated from a split of the same partial likelihood at a reticulation node so that these two could be merged. More formally, we say two LPLs P1 = (F1, s1) and P2 = (F2, s2), where |s1| = |s2|, are compatible if and only if for every 0 ≤ i < |s1|, either s1(i) = s2(i) or s1(i) ⋅ s2(i) = 0.
We denote by P x T and P x B the sets of LPLs that are associated with the top and bottom of branch x, respectively. These two quantities are computed in a bottom-up fashion, proceeding from the leaves of the network towards its root. Once the LPLs at the root are computed, the overall likelihood of a given site is computed. As the algorithm proceeds from the leaves towards the root, it needs to compute LPLs at the leaves, the top of a branch, the bottom of reticulation edges, and the bottom of tree edges. We now describe each of those computations; the overall algorithm is simply a bottom-up traversal of the network while applying the appropriate computation as a node is encountered.
Our algorithm computes the likelihood of a phylogenetic network given a set of biallelic markers. This algorithm computes matrix exponential along every branch, and processes the network’s nodes in a post-order traversal. Computation at a leaf takes O(1) time.
At a reticulation node, the time consumption increases after each reticulation node is processed, due to the accumulation of (split) LPLs. In the last processed reticulation node, the number of LPLs in its descendant is at most O(n4(k−1)). There are at most O(n4) new LPLs generated due to decompose-and-split operation for each original LPL. Therefore the time complexity of processing a reticulation node is at most O(n4k). We adopted the same approximation of matrix exponential as in [7], so the time complexity of computing matrix exponentiation is O(n2), and computation along every branch is at most O(n4k+2).
At a tree node, computation is mostly spent on evaluating Eq (13). Let n be the number of individuals present under an internal tree node. Then, this evaluation takes O(n4) time for a pair of compatible LPLs. The total time consumption of processing tree nodes also depends on the number of LPLs. Assuming k reticulation nodes in the phylogenetic network, there are at most O(n4k) pairs of compatible LPLs. Therefore the time complexity of processing a tree node is O(n4k+4).
In total, the time complexity of the algorithm is O(mn4k+4), where m is the number of species, n is the total number of lineages sampled from the species, and k is the number of reticulation nodes. Notice that when k = 0, which means the species phylogeny is a tree, the time complexity is O(mn4), which is the running time of the SNAPP algorithm without fast Fourier transforms.
To speed up computation, and since markers are independent, computations for the individual markers are parallelized by multi-threading. Furthermore, the data is preprocessed so that the unique marker patterns are identified and their probabilities are computed only once and reused for for all markers with the same patterns (states for the taxa).
The prior on the phylogenetic network is the same as that employed in [23], which we review briefly here. The prior is given by
p ( Ψ | ν , δ , η , ζ , α , β ) = p n u m r e t ( Ψ | ν ) × p d i a m ( Ψ | η ) × p d i v ( Ψ | δ ) × p p o p ( Ψ | ζ ) × p i n h ( Ψ | α , β ) . (15)
Here, p(Ψ|ν) is a Poisson prior on the number of reticulation nodes, normalized by the number of networks with the same number of reticulation nodes as Ψ. pdiam(Ψ|η) is an exponential prior on the diameters of reticulation nodes. The diameter of a reticulation node is the sum of the branch lengths on the cycle that contains the reticulation node in the underlying undirected graph of the network. pdiv(Ψ|δ) is an exponential prior on the divergence times. Rannala and Yang used independent Gamma distributions for time intervals (branch lengths) instead of divergence times. However, in the absence of any information on the number of edges of the species network as well as the time intervals, it is computationally intensive to infer the hyperparameters of independent Gamma distributions. Currently, we use a uniform distribution (as in BEST [33]). ppop(Ψ|ζ) is a Gamma prior on the population mutation rate. For ppop, we use the Gamma distribution Γ(2, ζ) with mean value 2ζ and shape parameter 2. pinh(Ψ|α, β) is a Beta prior, with parameters α and β, on the inheritance probabilities. Unless there is some specific knowledge on the inheritance probabilities, a uniform prior on [0, 1] is adopted by setting α = β = 1.
It is important to note here that if the topology of Ψ does not follow the phylogenetic network definition (e.g., has a cycle), then p(Ψ|ν, δ, η, ψ) = 0. This is crucial since, in the MCMC kernels we employ for sampling the posterior distribution, we allow the moves to produce directed graphs that slightly deviate from the definition; in this case, having the prior be 0 guarantees that the proposal is rejected. Using the strategy, rather than defining only “legal” moves simplifies the calculation of the Hastings ratios. However, the sampler always guarantees that the divergence times are consistent; that is, no node has a divergence time smaller than or equal to the divergence time of any of its descendants.
We employed the reversible-jump MCMC, or RJMCMC [34] algorithm implemented in PhyloNet [30] to sample from the posterior distribution given by
p ( Ψ | S ) ∝ L ( Ψ | S ) p ( Ψ ) , (16)
where Ψ here denotes the topology of the network and all its parameters, and p(Ψ) is the prior on the network and its parameters as described above.
We make use of only the 12 proposals designed for sampling phylogenetic networks and their parameters described in [23], but not the proposals aimed at sampling gene trees, as gene trees are integrated out.
We implemented in PhyloNet [30] a program to simulate bi-allelic markers on a given phylogenetic network. Bryant et al. [7] simulated bi-allelic markers by first generating gene trees inside a species tree (under the multispecies coalescent model), and then simulating the markers down the gene trees. In our case, we replaced the first step by generating gene trees inside a phylogenetic network under the multispecies network coalescent [26]; the second step of simulating bi-allelic markers down gene trees remains the same as that employed in [7]. When requiring the data set to contain only polymorphic sites, if the generated site is not polymorphic, we discard both gene tree and markers, and repeat until a polymorphic site is generated.
Two small subsets of a larger AFLP data set of multiple New Zealand species of the plant genus Ourisia (Plantaginaceae) [39] were analyzed, including previously unpublished AFLP profiles from two different hybrid individuals O. × cockayneana and O. × prorepens (herbarium codes follow [40] [continuously updated]). There are both morphological [41] and molecular (Meudt unpubl.) data supporting the hybrid nature of these two individuals. Although other Ourisia hybrid combinations have been reported in New Zealand [41], O. × cockayneana and O. × prorepens are perhaps the most common, both involve O. caespitosa as a putative parent, and both have been formally named. Each data subset comprised five diploid individuals in total, which means ten haploid individuals were effectively analyzed due to the correction for dominant markers. A Poisson distribution with λ = 1.5 as the prior on the number of reticulations, an exponential prior with λ = 2.0 as the prior on the species divergence times, and a Gamma distribution with α = 2.0 and β = 0.05 as the prior on the population mutation rates were adopted. An MCMC chain was run on each data subset for 1.5 × 106 iterations with 2 × 105 burn-in iterations, and a sample was collected every 500 iterations. We used following commands:
Phylogenetic networks allow for representing evolutionary relationships that involve both vertical and horizontal transmission of genetic material. Extensions of the multispecies coalescent process to include hybridization events have facilitated the development of statistical methods for inferring and analyzing phylogenetic networks from gene tree estimates and sequence data. A major challenge with using gene tree estimates as the input to species phylogeny inference methods is the error in these estimates. While using the sequence data directly overcomes this issue, the problem of recombinations within loci can confound inferences. Using bi-allelic markers from individual, independent loci could provide a way to avoid both the gene tree uncertainty and recombination problems (the two are not necessarily independent). Furthermore, it is important to note that many biological studies use data sets that consists of bi-allelic markers and no available sequence alignment data for individual loci.
Bryant et al. recently devised an algorithm for inferring species trees from bi-allelic genetic markers while analytically integrating out the gene trees for the individual loci [7]. In this paper, we extended their algorithm significantly so as the likelihood of a phylogenetic network given bi-allelic markers could be computed while integrating out the gene trees. This method complements existing ones that use gene tree estimates or sequence alignments as input for statistical inference of phylogenetic networks.
We implemented a Bayesian method for sampling the posterior of phylogenetic networks and their associated parameters from bi-allelic data, and studied its performance on both simulated and empirical data. The results indicate a very good performance of the method. This work adds a powerful method to the biologist’s toolbox that allows for estimating reticulate evolutionary histories.
A major bottleneck of the method is its computational requirements. While the SNAPP method is very time consuming on species trees, our method is much more time consuming given that reticulations in the phylogenetic network give rise to an explosion of the number of partial likelihoods that need to be computed and stored. More generally, the number of taxa in a data set has more of an effect on the running time of the method than the number of loci does. In particular, two aspects of the phylogenetic network under consideration affect the computational requirements of the method: The number of leaves under the reticulation nodes and the diameter of each of the reticulation nodes. As discussed above, the set of lineages entering a reticulation node must be bipartitioned in every possible way. This number of lineages is dependent on the number of leaves under that reticulation node. For example, if a single individual is sampled from a single species that exist under the reticulation node, then the number of bipartitions is very small (only two bipartitions exist). However, if n individuals are sampled from a single species that exist under the reticulation node or one individual is sampled per n species that exist under the reticulation node, then a number of bipartitions on the order of 2n arises. This computation becomes much more demanding if there are more reticulation nodes on the path to a lowest articulation node. As for the diameter—which is the number of branches on the paths between the two parents of the reticulation node and a lowest articulation node above them, the larger its value, the more demanding the computation becomes. An important direction for future research is improving the computational requirements of the method to scale up to data sets with many taxa.
|
10.1371/journal.pgen.1006343 | No Reliable Association between Runs of Homozygosity and Schizophrenia in a Well-Powered Replication Study | It is well known that inbreeding increases the risk of recessive monogenic diseases, but it is less certain whether it contributes to the etiology of complex diseases such as schizophrenia. One way to estimate the effects of inbreeding is to examine the association between disease diagnosis and genome-wide autozygosity estimated using runs of homozygosity (ROH) in genome-wide single nucleotide polymorphism arrays. Using data for schizophrenia from the Psychiatric Genomics Consortium (n = 21,868), Keller et al. (2012) estimated that the odds of developing schizophrenia increased by approximately 17% for every additional percent of the genome that is autozygous (β = 16.1, CI(β) = [6.93, 25.7], Z = 3.44, p = 0.0006). Here we describe replication results from 22 independent schizophrenia case-control datasets from the Psychiatric Genomics Consortium (n = 39,830). Using the same ROH calling thresholds and procedures as Keller et al. (2012), we were unable to replicate the significant association between ROH burden and schizophrenia in the independent PGC phase II data, although the effect was in the predicted direction, and the combined (original + replication) dataset yielded an attenuated but significant relationship between Froh and schizophrenia (β = 4.86,CI(β) = [0.90,8.83],Z = 2.40,p = 0.02). Since Keller et al. (2012), several studies reported inconsistent association of ROH burden with complex traits, particularly in case-control data. These conflicting results might suggest that the effects of autozygosity are confounded by various factors, such as socioeconomic status, education, urbanicity, and religiosity, which may be associated with both real inbreeding and the outcome measures of interest.
| It is well known that mating between relatives increases the risk that a child will have a rare recessive genetic disease, but there has also been increasing interest and inconsistent findings on whether inbreeding is a risk factor for common, complex psychiatric disorders such as schizophrenia. The best powered study to date investigating this theory predicted that the odds of developing schizophrenia increase by approximately 17% for every additional percent of the genome that shows evidence of inbreeding. In this replication, we used genome-wide single nucleotide polymorphism data from 18,562 schizophrenia cases and 21,268 controls to quantify the degree to which they were inbred and to test the hypothesis that schizophrenia cases show higher mean levels of inbreeding. Contrary to the original study, we did not find evidence for distant inbreeding to play a role in schizophrenia risk. There are various confounding factors that could explain the discrepancy in results from the original study and our replication, and this should serve as a cautionary note–careful attention should be paid to issues like ascertainment when using the data from genome-wide case-control association studies for secondary analyses for which the data may not have originally been intended.
| Close inbreeding (e.g., cousin-cousin mating) is known to decrease fitness in animals[1] and to increase risk for recessive Mendelian diseases in humans[2], a phenomenon known as inbreeding depression. Inbreeding depression is thought to occur due to evolutionary selection against genetic variants that decrease fitness—e.g., variants that increase risk of disorders[3]. Such fitness-reducing variants should not only be more rare, but also more recessive than expected under a neutral evolution model (i.e., show directional dominance). If so, individuals with a greater proportion of their genome in autozygous stretches (two homologous segments of a chromosome inherited from a common ancestor identical by descent [IBD]) should have higher rates of disorders. This is because autozygous regions reveal the full, harmful effects of any deleterious, recessive alleles that existed on the haplotype of the common ancestor.
Whether inbreeding increases risk for complex disorders like schizophrenia is less clear. Previous studies have found that inbreeding is associated with higher rates of complex disorders[4–9]. However, sample sizes have typically been small and the possibility that confounding factors might explain the results has left the links inconclusive. Moreover, close inbreeding accounts for fewer than 1% of marriages in industrialized countries[10], and information on pedigrees going back many generations is difficult to collect reliably. For these reasons, investigators have recently begun looking at signatures of very distant inbreeding (e.g., common ancestry up to ~100 generations ago) using genome-wide single nucleotide polymorphism (SNP) data in an attempt to understand whether autozygosity increases the risk to schizophrenia and other complex diseases[11]. Autozygosity in SNP data is typically inferred from runs of homozygosity (ROHs): long, contiguous stretches (e.g., > 40) of homozygous SNPs. The proportion of the genome contained in such ROHs, Froh, can then be used to predict complex traits[12–19]. Keller et al.[11] showed that Froh is the optimal method for detecting inbreeding signals that are due to rare, recessive to partially recessive mutations, such as those thought to occur when traits are under directional selection[3]. The low variation in Froh means that large sample sizes (e.g., >12,000) are required to uncover realistic effects of distant inbreeding on complex diseases in samples unselected for inbreeding[11].
In 2012, Keller et al.[20] used the original Psychiatric Genomics Consortium schizophrenia data (17 case-control datasets, total n = 21,831) to investigate whether Froh is associated with increased risk of schizophrenia. The authors estimated that the odds of developing schizophrenia increased by approximately 17% for every additional percent of the genome that is contained in autozygous regions (β = 16.1, CI(β) = [6.93, 25.7], p = 6x10-4.) This was by far the largest study to that date examining the association between Froh and any psychiatric disorder, and the significant relationship between Froh and case-control status remained robust through secondary analyses of various covariate combinations, common vs. rare IBD haplotypes, and SNP thresholds used to define ROHs. These results are consistent with the hypothesis that autozygosity causally increases the risk of schizophrenia. Nevertheless, because various confounding factors may increase likelihood of distant inbreeding as well as the probability of having offspring with schizophrenia, these results do not imply a causal relationship. For example, parents higher on schizophrenia liability may pass their higher liability to offspring and mate with more genetically similar partners (e.g., due to decreased mobility, educational opportunities, etc.).
The current study seeks to provide a well-powered, independent replication of Keller et al.(2012)[20]. In light of the growing concern about publication bias[21,22] and dearth of well-powered replications[23,24], this follow-up analysis is a necessary step in validating the Froh—schizophrenia relationship. The present study used genome-wide SNP data from 22 independent schizophrenia case-control datasets (n = 39,830) from the PGC[25] to further examine the relationship between Froh and schizophrenia. Our replication attempt is an important contribution to the growing body of literature examining autozygosity and psychiatric disorders, and should help verify whether autozygosity estimated from ROHs is robustly related to schizophrenia risk and, by extension, can help elucidate whether schizophrenia risk alleles are biased, on average, toward recessive effects.
SNP data from 28,985 schizophrenia cases and 35,017 controls were collected as detailed in Ripke et al.[25]. Quality control (QC) and analyses were conducted separately for the original and replication datasets. The “original” dataset included subjects from the PGC’s SCZ1[26] samples used by Keller et al[20] (n = 21,868 after QC), and the “replication” dataset contained all subjects (n = 39,830 after QC) in the PGC SCZ2[25] samples not included in the original Keller et al. study, making the replication dataset independent of the original dataset analyzed in Keller et al.
Despite the number of imputed SNPs ranging from ~1.8 million to ~4.2 million in the datasets, there were not enough well imputed SNPs in common across all 22 datasets to conduct a viable ROH analysis in the same way as in the original study (see Methods). Nevertheless, Keller et al. also reported results from ROHs estimated from unimputed SNP data, and these results were highly consistent with imputed SNPs. Therefore, our primary analyses were conducted using post-QC, unimputed genotype data. We also report results on imputed SNPs (see S5–S12 Figs and S1 Table) using slightly different QC procedures than used in the original report (see Methods), which do not change the conclusions below. While ROHs from the imputed data were called from a common SNP set, ROHs from the unimputed data were called on unique sets of SNPs for each dataset.
Keller et al.[20] found that all ROH length thresholds were significantly associated with schizophrenia, but because ROH thresholds are ultimately arbitrary, they focused their discussion on the thresholds (e.g., 110 consecutive homozygous SNPs in the unimputed data) that maximized the schizophrenia-ROH relationship. In an attempt to follow as closely as possible the method used by Keller et al., we report two sets of ROH results. The first approach—a direct replication attempt of Keller et al.—defined ROHs as being ≥ 110 consecutive homozygous SNPs in a row (with median Mb ranging from ~1 to ~3.4 Mb, depending on sample) in the unimputed data. Because using unimputed SNP data introduces large differences in mean ROH length across datasets (when defined by number of consecutive homozygous SNPs) due to varying SNP densities, we also employed a secondary replication approach using a 2.3 Mb minimum length threshold that corresponds to 110 SNPs-in-a-row average length in the original report. As in the original report, we also show results across all thresholds to ensure that no results were missed.
Table 1 gives the descriptive statistics for average ROH lengths and Froh across datasets, where ROHs were defined as ≥ 110 consecutive homozygous SNPs. There was wide variation in average Froh and ROH lengths between datasets, a consequence of using unimputed SNP data, which introduces more between-dataset variability in Froh and mean ROH length[20]. Across datasets, mean Froh was also higher (0.30% vs. 0.14%) and average ROH lengths shorter (1.1–3.4 Mb vs. 2.0–4.7 Mb) in the replication versus original datasets. Part of the reason for the mean Froh discrepancy seemed to be due to replication datasets being genotyped on denser SNP chips, because this discrepancy reduced when we defined ROHs as ≥ 2.3 Mb homozygous SNPs (0.22% vs. 0.13%; Table 1). The remaining higher average Froh in the replication datasets appears to be due to more samples being from countries with higher overall Froh (e.g., Sweden, Estonia, Israel) in the replication datasets; the average Froh levels were very similar across replication vs. original datasets within the same countries.
For each dataset, we regressed case-control status on Froh using mixed effects logistic regression treating dataset as a random factor, and controlled for 20 principal components (PCs) from the genomic relationship matrix[27] and two SNP quality measures (excess heterozygosity and SNP missingness; see Methods). In Keller et al. (2012), the authors used mixed effects models to test the ROH burden association with schizophrenia. However, in the current analysis we used fixed effect logistic regression models, treating dataset as a fixed, because a minority of the mixed effects models failed to converge. When the mixed effects models did converge, the results were highly similar to the respective fixed effect models. Figs 1 and S1 show the predicted change in odds of schizophrenia risk (and 95% confidence intervals) for every 1% increase in average Froh for each logistic regression in the replication data using ROHs defined by either ≥110 consecutive homozygous SNPs (Fig 1) or ROH length ≥ 2.3 Mb (S1 Fig). The overall association between schizophrenia and Froh in the replication data was in the predicted direction but not significant for ROHs defined as at least 110 consecutive homozygous SNPs (β = 0.19, CI(β) = [−4.50,4.88], Z = 0.08, p = 0.94) or for ROHs defined as ≥ 2.3 Mb (β = 0.75, CI(β) = [−4.05,5.56], Z = 0.31, p = 0.76). The results from analyses on ROHs called from imputed rather than raw SNP data were also non-significant (S5 Fig). As in Keller et al., we also explored increasingly long SNP and Mb ROH thresholds to assess the stability of the Froh-schizophrenia relationship (Figs 2 and 3). Across all thresholds, the only thresholds that approached significant associations between Froh and schizophrenia in the replication data were at the upper limits of the Mb-length ROH thresholds; the strongest association was for ROHs defined as ≥ 19 Mb (β = 8.64, CI(β) = [−0.85,18.13], Z = 1.78, p = 0.07).
We conducted a series of follow-up analyses to ensure that the failure to replicate our original report was not due to analytical error, inclusion of outlier individuals or datasets, or suppressing covariates in the replication data. We reran the same analyses described above on SNP data from the “original” report using the exact same quality control and analytic procedures performed on the replication data. Results were virtually identical to those obtained in Keller et al.’s 2012 study (S2–S4 Figs), increasing our confidence that the procedures used in the replication attempt were identical to those used in the original analysis and that the results from the original analysis were not due to analytic or procedural errors. We then reran analyses in the replication data after (a) omitting individuals with very long (>30 Mb) ROHs, (b) omitting only long ROHs, (c) including all combinations of covariates in the model (SNP missingness, average heterozygosity, 10 or 20 principle components), and (d) including only the longest ROH for each individual. The Froh-schizophrenia relationship remained non-significant in these follow-up analyses (results shown in S2 Table).
We noticed that there was greater variability in Froh in the replication datasets and that this greater variability was mostly driven by replication datasets that had n < 300. Under the premise that smaller samples might differ in genotypic or phenotypic quality, we excluded seven samples that contained fewer than 300 cases (“egcu”, “ersw”, “lie2”, “pews”, “top8”, “umes”), reran our baseline analysis (including all covariates mentioned above and using an ROH threshold of ≥ 110 consecutive homozygous SNPs), but still observed a non-significant Froh-schizophrenia relationship (β = 1.04, CI(β) = [−3.88,5.96], Z = 0.42, p = 0.68) in the predicted direction. Therefore, this post-hoc analysis does not lend support to the possibility that small samples in the replication set added noise to our analysis, obscuring an Froh-schizophrenia relationship.
Although results from the replication analysis were not significant, they were in the same direction as the original analysis. It could therefore be argued that the best estimate of the association between ROHs and schizophrenia is obtained by combining the two datasets. When we reran our analyses on the combined original + replication data (n = 61,661), all Froh associations based on ROH thresholds greater than 60 consecutive homozygous SNPs or longer than 1 Mb were significant (Figs 4 and 5). For an ROH threshold of ≥ 110 consecutive homozygous SNPs), we observed a significant Froh-schizophrenia relationship in the combined data (β = 4.86, CI(β) = [0.90,8.83], Z = 2.40, p = 0.02). In this combined dataset, we also used a replication status-by-Froh interaction to conclude that the Froh-schizophrenia association was only marginally higher in the original compared to the replication datasets (interaction β = −3.98, Z = −1.84, p = 0.07) for ROHs defined as at least 110 consecutive homozygous SNPs.
To assess the relative importance of distant versus close inbreeding, we compared the effects of short versus long ROHs. As in the original study, we chose our ROH length threshold based on the Mb length cutoff that resulted in equal Froh variances, calculating Froh_short as the proportion of the genome contained in ROHs < 8 Mb long, and Froh_long as the proportion of the genome contained in ROHs > 8 Mb long. Although neither association was significant, the effect of Froh_short (β = −5.06, CI(β) = [−12.08,1.95], Z = −1.42, p = 0.16), caused by autozygosity arising from more ancient common ancestors, was negative (“protective”) and in the opposite direction of effect of Froh_long (β = 1.23, CI(β) = [−4.78,7.25], Z = 0.40, p = 0.69), caused by autozygosity arising from more recent common ancestors, which predicted increased risk for schizophrenia (Fig 6).
Despite exploring various homozygous SNP length thresholds, Mb thresholds, and combinations of covariates, the findings from this study do not lend much support to the original observation of a highly significant Froh-schizophrenia association[20], and provide only equivocal support, based on combining the original and replication data, for the hypothesis that autozygosity is a risk factor for schizophrenia.
Perhaps the simplest explanation for this pattern of results is that the conclusions about distant inbreeding from the original data represent a type-I error or that the lack of replication in the current report was a type-II error. Despite the fact that the effect in the original study was highly significant (p = 6x10-4) and the statistical power in the replication study to detect the observed effect size in the original study was nearly 100%, it is possible that the estimated effects of the original analysis could have been over-estimated and/or those of the replication analysis under-estimated, due to sampling variability. There is some support for this interpretation, as there was not a significant difference in results between replication versus original datasets (interaction p = 0.07).
An alternative explanation for the overall pattern of results has to do with the potential influence of unmeasured confounding factors in both the original and replication analyses. Unlike genotype frequencies, which change very slowly and are unaffected by inbreeding, ROH levels can change substantially after even a single generation of inbreeding, making ROH analyses highly susceptible to confounding factors associated with both disease risk and the degree of inbreeding/outbreeding. For example, contrary to initial predictions, Abdellaoui et al.[28] identified a significant and negative (“protective”) relationship between Froh and risk for major depressive disorder (MDD) in the Dutch population. However, the authors found that religiosity was significantly associated with both higher autozygosity and lower MDD in this population. When religiosity was accounted for in their regression model, the original association between MDD and Froh disappeared. A similar effect was detected for educational attainment: highly educated individuals were more likely to migrate and mate with highly educated and more diverse partners, making highly educated spouse pairs share less ancestry and leading to their offspring having lower Froh[29]. Thus, assortative mating on variables such as education or religion could subtly influence observed Froh associations, potentially affecting results in ways that can be difficult to account for. For example, an observed Froh-schizophrenia relationship could be due to parents with a higher schizophrenia liability mating with less genetically diverse mates due to, e.g., fewer educational opportunities or lower migration rates. Thus, the causation may be reversed: schizophrenia liability in parents could cause not only higher schizophrenia risk, but also higher Froh, in offspring rather than Froh in offspring increasing their schizophrenia liability. Such reverse and third variable causation possibilities can only be tested if relevant socio-demographic variables in subjects and (optimally) their parents are collected.
The possibility of unmeasured variables confounding Froh-disorder relationships seems particularly likely in analyses conducted on ascertained samples. Ascertainment of cases and controls not perfectly matched on socio-demographic factors that might affect degree of outbreeding (e.g., socioeconomic status, education level, age, religion, urbanicity) can mask any true Froh association and bias the observed association in either direction. Such a scenario might explain otherwise contradictory findings in previous ROH case-control analyses[18,28,30–36]. For example, following two studies showing that genome-wide autozygosity was significantly associated with schizophrenia risk, including the original Keller et al. study[13,20], two newer studies failed to replicate this association[34,35], although both replication sample sizes (n = 3,400 and 11,244 respectively) were substantially smaller than the current one (n = 39,830). (It should be noted that the sample used in the latter study[36] overlapped with the samples in both the original Keller et al.[20] study and the current replication study). Even within the same study, Froh results in ascertained samples have been inconsistent. Using PGC MDD data, Power et al.[36] found a significant positive Froh-MDD relationship in data from three German sites but a significant negative Froh-MDD relationship in six non-German sites. A possible explanation for this and other such examples of heterogeneity across sites they observed is that cases and controls differed on socio-demographic factors that were associated with Froh, and the direction of this ascertainment bias was inconsistent across data collection sites.
We believe that similar ascertainment biases could have affected results in the present study as well as in the original Keller et al.[20] report. Many of the PGC schizophrenia datasets used cases ascertained from hospitals, clinics, health surveys, and advertisements but controls from previous biomedical research volunteers, university students, blood donors, and population registries. While such differences in ascertainment between cases and controls are highly unlikely to lead to allele frequency differences, and thus are of little concern to genome-wide association studies, they could very easily lead to Froh differences due to differences in degree of inbreeding/outbreeding in the populations from which cases and controls were drawn. Controlling for ancestry principal components in this case would only help to the degree that degree of inbreeding/outbreeding is associated with ancestry. Unfortunately, none of the other variables that might statistically control for such biases due to differences in case/control ascertainment are currently available in the PGC data collection. The PGC collection of studies was designed for association analyses; it was not optimally designed for ancillary purposes, such as ROH analyses.
It is important to recognize that even ascertainment biases that differ at random across sites would substantially inflate type-I error rates because the proper degrees of freedom for the test should be closer to the number of independent sites rather than the number of independent cases and controls. To demonstrate this, we permuted data under the null hypothesis of no relationship between Froh and schizophrenia in the 17 datasets from the original 2012 study by randomly flipping case or control status within each dataset for each permutation (e.g., cases and control statuses in a dataset either remained the same or were flipped to the opposite status). We then calculated the overall Froh ~ schizophrenia relationship with the same logistic regression model and using the same covariates as in the original analysis. Across 1,000 permutations, 183 p-values were significant (p < 0.05), implying a type-I error rate of 0.18 and demonstrating how false conclusions about Froh relationships can be reached even when ascertainment biases are random across multiple sites.
Given concerns about the false discovery rate in science[22], there has been increasing emphasis on the need for well-powered, direct replications of novel findings in genetics[23,37,38] and other fields[39–41]. The current study was a well-powered, direct replication attempt that failed to replicate an earlier finding that autozygosity arising from distant common ancestors was significantly associated with schizophrenia. As is typical with null findings, it is difficult to identify the reason for this failure to replicate. However, we have argued that a likely cause is that ROH associations are highly susceptible to confounding, especially in case-control (ascertained) samples. Thus, we believe that the conclusions of the original study were premature and the true causal relationship between schizophrenia and autozygosity could be either stronger/more positive (if the populations from which controls were ascertained were, on average, slightly less outbred than populations from which cases were ascertained) or weaker/more negative (the reverse) than reported here. Unfortunately, we do not have the ability to test these hypotheses directly in the current datasets, and doing so awaits either new samples in which cases and controls are carefully matched or the collection of information that allows potential confounders to be statistically controlled. This creates a dilemma for ROH analyses using existing case-control genome-wide data: GWAS datasets usually do not match cases and controls to the degree necessary to rule out confounding effects on ROH analyses and typically do not collect the relevant socio-demographic information necessary to control for potential confounders. The current study therefore serves as a cautionary tale for analyzing ROHs in existing ascertained GWAS datasets. Such datasets may be perfectly adequate for their designed purpose–GWAS–but may be problematic and even misleading for ROH analyses.
Our study used 37 datasets from the Psychiatric Genomics Consortium’s SCZ2 data–these data included 28,985 schizophrenia cases and 35,017 controls, collected from 37 sites in 13 countries. Data collection and ascertainment details are described elsewhere.[25]
Keller et al.[20] used 17 datasets from the PGC SCZ1[26] data. Several of these original 17 studies recruited additional subjects by the time of our study, necessitating two well-defined, independent datasets: one including all of the individuals analyzed in the original 2012 study (“original” dataset), and one containing only subjects not included in Keller et al.’s 2012 report (the “replication” dataset, comprised of 22 studies and a total sample size of 18,562 cases and 21,268 controls after QC; see Table 1). Three of the original case-control datasets from the PGC’s SCZ1 added more subjects and/or controls in SCZ2, but only two of these datasets had enough subjects to pass QC and merit inclusion in the current study—thus there is a “top8” dataset (N = 180) in this replication study, comprised of the samples that were added to the “top3” dataset (N = 598) from the original 2012 study, and a “boco” dataset (N = 1,870), which includes the new cases and controls that were added to the original “bon” dataset (N = 1,778). For consistency with the original Keller et al. (2012) study[20], we excluded the three family-based datasets of parent-proband trios and three East Asian datasets.
We followed the same QC procedures as Keller et al.[20]. We removed a) one individual from any pair of individuals who were related with π^ >0.2, b) individuals with non-European ancestry as determined by principal components analysis; c) samples with SNP missingness >0.02; or d) samples with genome-wide heterozygosities >6 standard deviations above the mean. SNPs were excluded if they a) deviated from Hardy-Weinberg equilibrium at p<1×10−6; b) had missingness >0.02; or c) had a missingness difference between cases and controls >0.02.
Early in the analysis process, we found that only including SNPs with imputation dosage r2 > .90 across all datasets, as was done in the original study[20], left us with too few SNPs with which to conduct viable ROH analyses in the replication data. Because having ROHs of similar length and SNP density is important for comparing present results to those from the 2012 study, we decided that having a similar number of SNPs to Keller et al.[20] was more important than following the exact same QC procedures. Thus, to arrive at a similar number of genome-wide SNPs in the new and old datasets, some of the QC measures described below were different than in the 2012 investigation.
SNPs were imputed using the 1000 Genomes reference panel[42]; imputation procedures are described elsewhere[25]. Imputation dosages were converted to best-guess (highest posterior probability) SNP calls because ROH detection algorithms require discrete SNP calls, and extremely stringent QC thresholds were employed to achieve accuracy rates similar to those in genotyped SNPs[43]. We excluded any imputed SNPs that were not included in the HapMap3[44] reference panel, as done in the 2012 study. Unlike the original QC procedures, we did not require that the dosage r2 had to be > .90 in each individual datasets. We excluded any imputed SNPs that had a dosage r2<0.98 or >1.02 in the overall sample (calculated using average dosage r2 weighted by sample size) or that had MAF<0.15 within each sample (vs. .05 in original), leaving 340,084 high-quality imputed SNPs (vs. 398,325 in original).
Again, we followed the same ROH calling procedures as in Keller et al[20]. As recommended in a separate investigation[45] by three of the authors of the present study, we chose PLINK software[46] for its computational efficiency and superior detection of autozygous stretches. As in the 2012 study, we pruned for LD using PLINK’s—indep flag, which ensures more uniform SNP coverage across the genome and reduces false autozygosity calls by removing redundant markers. We pruned SNPs for LD using a VIF threshold of 10, which is equivalent to multiple R2 > 0.90 between the focal SNP and the 50 surrounding SNPs.
We called ROHs using PLINK’s—homozyg flags, defining initial ROHs as being ≥40 homozygous SNPs in a row with no heterozygote calls allowed. We required that ROHs have a density greater than 1 SNP per 200 kb, and split an ROH into two if a gap >500 kb existed between consecutive homozygous SNPs. We then post-processed the initial ROH calls by altering the SNPs-in-a-row threshold and the Mb length threshold; specifically, we looked at ROH calls with a minimum of 40 to 200 consecutive homozygous SNPs in increments of 10, and ROH calls with minimum lengths ranging from 1 to 20 Mb by increments of 1 Mb. We varied ROH thresholds this widely to ensure that no potential effects of autozygosity were missed, but the primary results presented here are based on two replication attempts in the unimputed data: (a) using the same SNP thresholds that gave the most straightforward comparison with the original report (this was 110 SNPs-in-a-row for the unimputed data, spanning ~1 to ~2.1 Mb in the replication datasets, and 65 SNPs-in-a-row for the imputed data), and (b) using the physical length threshold (2.3 Mb) that corresponded to the average Mb length for 110 SNPs-in-a row in the original report.
After calling ROHs, we summed the total length of all autosomal ROHs for each individual and divided that by the total SNP-mappable distance (2.77x109 bases) to calculate Froh. Froh, the proportion of the genome contained in long homozygous regions, was used as the predictor of schizophrenia case-control status in analyses described below. As confounding factors such as population stratification, SNP missingness, call quality, and plate effects can influence Froh, we included the first 20 principle components (based on a genome relationship matrix calculated from ~30K LD-pruned SNPs), percentage of missing SNP calls in the raw data, and excess heterozygosity in all regression models[20]. We then regressed case-control status on Froh using a mixed linear effects logistic regression model (available in the lme4 package in R version 3.1.0), treating dataset as a random factor, to assess the overall effect of Froh on schizophrenia across all sites. Some of the models with random effects did not converge; thus, for consistency, we modeled dataset as a fixed factor for all analyses. The results from mixed linear effects models that converged were very similar to fixed effects models, giving us confidence that the fixed effects results of this analysis and the random effect results from the original Keller et al. (2012) study are commensurate. We also ran logistic regressions in each of the 22 datasets separately.
This research was approved by CU Boulder's Institutional Review Board with regard to protocol number 13–0266 on 3/29/2016 in accordance with Federal Regulations at 45 CFR 46. Written patient consent was obtained for each individual study by the study PI, with the exception of the "clm3" and "clo3" datasets, which obtained anonymous samples via a drug monitoring service under ethical approval and in accordance with the UK Human Tissue Act.
|
10.1371/journal.pcbi.1005591 | Solving the influence maximization problem reveals regulatory organization of the yeast cell cycle | The Influence Maximization Problem (IMP) aims to discover the set of nodes with the greatest influence on network dynamics. The problem has previously been applied in epidemiology and social network analysis. Here, we demonstrate the application to cell cycle regulatory network analysis for Saccharomyces cerevisiae. Fundamentally, gene regulation is linked to the flow of information. Therefore, our implementation of the IMP was framed as an information theoretic problem using network diffusion. Utilizing more than 26,000 regulatory edges from YeastMine, gene expression dynamics were encoded as edge weights using time lagged transfer entropy, a method for quantifying information transfer between variables. By picking a set of source nodes, a diffusion process covers a portion of the network. The size of the network cover relates to the influence of the source nodes. The set of nodes that maximizes influence is the solution to the IMP. By solving the IMP over different numbers of source nodes, an influence ranking on genes was produced. The influence ranking was compared to other metrics of network centrality. Although the top genes from each centrality ranking contained well-known cell cycle regulators, there was little agreement and no clear winner. However, it was found that influential genes tend to directly regulate or sit upstream of genes ranked by other centrality measures. The influential nodes act as critical sources of information flow, potentially having a large impact on the state of the network. Biological events that affect influential nodes and thereby affect information flow could have a strong effect on network dynamics, potentially leading to disease. Code and data can be found at: https://github.com/gibbsdavidl/miergolf.
| The Influence Maximization Problem (IMP) has been applied in fields such as epidemiology and social network analysis. Here, we apply the method to biological networks, aiming to discover the set of regulatory genes with the greatest influence on network dynamics. Fundamentally, since gene regulation is linked to the flow of information, we framed the IMP as an information theoretic problem. Dynamics were encoded as edge weights using time lagged transfer entropy, a quantity that attempts to quantify information transfer across variables. The influential nodes act as critical sources of information flow, potentially affecting the global network state. Biological events that impact the influential nodes and thereby affecting normal information flow could have a strong effect on the network, potentially leading to disease.
| In order to respond to messages and environmental changes, cells dynamically process information arriving from cell surface receptors [1,2]. Information is transferred, stored, and processed in the cell via molecular mechanisms, often triggering a response in the regulatory program. These types of dynamic genetic regulatory processes can be modeled and analyzed using networks.
The cell cycle process in Saccharomyces cerevisiae is well studied, but not completely characterized [3]. The dynamic regulatory process is controlled by a network that processes signals. To gain further understanding of the regulatory structure, we used publicly available time series data and regulatory databases to solve the influence maximization problem (IMP) (Fig 1) [4,5].
Recently, the influence maximization problem (IMP) has received a great deal of interest in social network analysis and epidemiology as a general method for determining the relative importance of nodes in a dynamic process [6,7]. Use case examples are found in modeling the spread of infectious disease in social networks and in identifying optimal targets for vaccination (or advertisements) [8]. The IMP is a search over sets of nodes that, when acting like sources in a diffusion process, cover as much of the network as possible [9,10].
Diffusion on graphs is part of a general class of problems where some quantity flows from source nodes, across the edges of a graph, draining in sink nodes. Various forms of network flow methodologies have found success in algorithms such as Hotnet, ResponseNet, resistor networks, and others [11,12,13]. Diffusion, like the propagation of infection, does not follow algorithmically defined paths on graphs, such as shortest paths, but instead flows on all possible paths. In this work, we use a diffusion algorithm that is modeled using a random walk, where transition probabilities are proportional to edge weights. The random walk produces an expected number of visits to each node. If the expected number of visits is greater than a given threshold (here 0.0001), the node is considered to be ‘covered’, and the network cover is a count of ‘covered’ nodes. The goal of the IMP is to maximize this network cover with a fixed number of nodes.
In our application of the IMP to genetic regulatory networks, the diffusion process represents a flow of information on the network, which opens up many applications in biology [14,15,16]. Directional information flow can be described quantitatively using the model free method, transfer entropy (TE) [15]. Since processes in biology are not instantaneous, time lags are introduced, representing a lag between the transmission and reception of information. As an example, the expression of transcription factors, their subsequent binding to promoter regions, and ultimately, the induction of transcription can take substantial amounts of time.
In this case, we use ant optimization to search for sets of source nodes that lead to diffusion generated network covers that score highly [17]. Typically, ant optimization is used for path finding, but it can also be applied to combinatorial, subset selection problems [18,19]. In ant optimization, ants construct potential solutions as sets, which are scored and reinforced, encouraging good solutions in later iterations. In this work, the result of the optimization procedure is an optimal, or nearly optimal, set of nodes that maximizes network cover after applying the diffusion [15]. In application to biological networks, the IMP essentially remains an unexplored area of research [20].
Each run of the IMP returns a solution set of size K. Using both ‘fast’ and ‘slow’ parameter sets for the ant optimization, we have run the IMP for values of K from 1 to 50, resulting in 50 solutions, one set for each value of K. Genes were ranked by counting the number of times a given gene appeared in a solution set. A highly influential gene would appear in the solution for many values of K, regardless of the solution set size, implying that topologically, the gene is in an optimal position as a source of information, enabling contact to a large portion of the network. Optimization can proceed at different rates; more restarts, more ants, a slow pheromone evaporation rate, and a high number of local optimization steps may result in more robust and repeatable results, but more iterations might be needed and the run time can be longer. On the other hand, few restarts with a small number of ants and a fast evaporation rate, plus fewer local optimization steps, leads to more stochastic results and a shorter run time. The slow-and-steady approach can consistently get stuck in non-optimal minima, whereas the highly stochastic results can sometimes 'jump' out of non-optimal minima. In order to explore results and convergence behavior, both fast and slow parameter sets were used. Our results from either parameter set were in excellent agreement regarding influence rankings, reducing concerns about the stochastic nature of ant optimization.
To better understand topologically where the influential genes are situated, we compare the IMP solution sets to gene sets derived from other centrality metrics, such as degree centrality [21], betweenness-centrality [22], where shortest paths are considered, and PageRank [23], the algorithm used in web search.
This analysis produced a ranked list of genes that agrees with previous studies of cell cycle regulators and models, giving credence to the method as a fairly general approach to analyzing large scale biological network dynamics.
The yeast genetic regulatory network was constructed starting with 26,827 genetic regulatory edges from YeastMine and statistically filtering out edges [4]. Regulatory processes in biology are not instantaneous, so time lags are introduced to account for propagation time (S1 Fig) [24]. Further, genetic regulatory interactions are directional; transcription factors act on genes, and not the other way around. So, although correlation is easy to compute and is sometimes used to estimate the activity of regulatory edges, there are more appropriate metrics to use with time series data, such as transfer entropy. Transfer entropy (TE) is a model-free method that attempts to quantify information transfer between two variables in a directional manner. At present, computing transfer entropy is not trivial, and there is active research in comparing and deriving methods for approximating the value. In this work, we used a Gaussian kernel density based approach, which has been previously shown to be relatively accurate [25,26].
Using time series data for 5,080 measured genes and 26,827 genetic regulatory edges from YeastMine, both time lagged Spearman's correlation and transfer entropy were computed for all regulatory edges. Permutation-based statistics were applied to assess the significance of TE. Edges were accepted if empirical p-values was less than or equal to 1(pn+1) where pn is the number of permutations (pn = 50,000).
Spearman's correlation tests were performed on each time lag (0–5 time steps). The maximum ρ was kept, and at FDR 1%, this resulted in 12,555 edges, containing 3,939 nodes. Significant edge weights had a median correlation of 0.58. Most of the edges (52%) showed a maximum correlation when using a time lag of zero.
The metric of interest, time lagged TE, resulted in 2,084 significant edges containing 1,409 nodes with median values of 0.499 (Fig 2).
The overlap between the correlation and TE networks is moderate; only 16% of the edges in the correlation network are shared with the TE network (1,988 of 2,084 edges in the TE network or 97%), and while most TE nodes are found in the correlation network (95%), only 35% of the correlation nodes are found in the TE network. When comparing Spearman's and TE weights on matched edges, the correlation between matched edge weights was moderately weak (Spearman's correlation 0.43). Additionally, the mean node degree distribution in the correlation network is much higher than that of the TE network. For example, SFP1 has degree 923 in the correlation network, compared to 76 in the TE network (summing both in- and out-edges). The high node degree in the correlation network suggests that correlation testing may be overly permissive, with less informative edge weights.
Clauset, Shalizi, and Newman’s method for statistically determining whether a network is ‘scale-free’ showed that the TE network is not [27]. Using the TE network, the result showed alpha = 2.17, which is consistent with power law networks. However, the goodness of fit test using the Kolmogorov-Smirnov statistic produced a p-value of 0.011, indicating that only a small fraction of the simulated scale-free distributions are "close" to the observed degree distribution.
In the rest of the analysis, only the transfer entropy network is used, since it is clear that the correlation-based network is not a super-set of the transfer entropy network, does not agree in the weighting, and is likely overly permissive with regard to active interactions.
Using transfer entropy to quantify information flow, if an upstream node transfers information to a downstream node, respecting edge directions, the downstream node is said to be 'influenced'. The area of influence can be found by application of a diffusion process, where the flow follows edges with greater information transfer (edges with greater weights), ‘visiting’ nodes and resulting in a cover on the network. The maximization problem involves finding a set of nodes with size K, that when treated as sources, influences the largest proportion of the network, which is to say, that after the diffusion process is applied, no other set would lead to a greater network cover.
The Influence Maximization Problem (IMP) was solved over a range of set sizes, K = 1 to 50. Since ant optimization is stochastic and can result in variable solutions, two different parameter sets were used (S1 Text). First a ‘slow’ parameter set was used (best of 8 restarts, 64 ants, 32 local optimization steps, evaporation rate 0.2). The range of K was run three times, for a total of 150 ant-optimization runs. A count was made on the number of times genes were selected across solutions. As an example, if a gene appeared in 46 solutions, on average, for K = 1 to 50, it would be considered a high-ranking gene. The influence score, representing a network cover, increased quickly for small values of K, gradually leveling out. With K = 44 source nodes (3% of the network), a maximum network cover of 1,308 nodes (93%) was produced. Beyond K = 44, the score increased by single digits through the addition of single nodes (see S2 Fig). Regarding the rate of change in network cover, from K = 1 to K = 2, the total network cover increased 12%. However, after that, the rate of increase drops quickly. Between K = 14 to K = 15, the network cover increased at a rate of less than 1%, and after K = 24, for each additional node added to the set of sources, the increase in network cover dropped to less than 0.5%. The top ranked gene FKH1, was selected on average 49 (out of 50 possible) times, followed by two genes, SFP1 and TFC7, that were selected on average 47 and 46 times respectively. Overall, 52 genes were selected in at least one run.
A second parameter set, the ‘fast’ set, used 4 restarts, 16 ants, 8 local optimization steps, evaporation rate 0.2. For each value of K, 49 optimizations were run, for a total of 2,450 result sets. We found that faster optimization runs lead to more variation in the results. However, using the same ranking method, counting the number of times a gene was selected, resulted in excellent agreement with the ‘slow’ parameter set (S1 Text, S3 Fig). The set of genes in the top 15 ranked influencers are identical across parameter sets. The top 15 influencers from both parameter sets are found in Table 1.
To provide a basis for comparison to the ranked influencers, 13 different centrality measures were computed on the TE network. Brief descriptions of each centrality metric can be found in supplementary text (S1 Table).
As stated earlier, after K = 24, the increase on network cover had dropped below 0.5%, making this a reasonable stopping point in selecting the most influential genes. To compare with other metrics, the top 24 genes were selected for each centrality measure. A Jaccard index was computed for each pair of centrality measures (Fig 3), and although some clustering is observed among centrality metrics, especially among node-degree related measures, there remains substantial disagreement in top ranked genes.
The top ranked influential genes are not found among highly ranked genes in eigenvector based centrality measures including authority, eigenvector centrality, and alpha centrality. However, eigenvector related measures of centrality revealed important genes that are not found in other lists. For example, the well-known cell cycle regulatory gene CLB2 was selected by alpha centrality and authority, while it was not found using influence ranking or betweenness. Overall, no ranked list contained a definitive set of cell cycle related regulators. Across measures, gene set enrichment showed a wide variety of associations with biological processes, illustrating differences in the gene rankings (S2 Text).
We have found that within the regulatory network structure, the influential genes tend to be situated upstream of genes selected by other centrality measures (Fig 4, S4 Fig).
For example, the influencer genes act as regulators for genes selected by alpha centrality, while no genes selected by alpha centrality regulate the influencer genes. The same is found for the eigenvecteor centrality and betweenness sets. In some cases, there is a fair amount of overlap in the top-level regulators, such as among the high degree nodes and the articulation set. But, overall, we see the influencers stay as top-level regulators to genes selected by other centrality measures. This can be quantified by computing the fraction of reachable genes, starting at a given measure, and excluding overlapping genes (Fig 5).
For example, starting at the set of influential genes, 79% of the betweenness selected genes can be reached, while starting at the betweenness genes, only 12% of influencers can be reached. Starting at the influencer genes, 41% of degree central nodes can be reached, while only 12% of influencers can be reached from the degree central nodes. Starting from every centrality measure, the fraction of reachable nodes is fewer, compared to starting from the influential genes. On average, 54% of “central genes” (excluding subgraph centrality) can be reached when starting at the influential genes, compared to 8% of reachable influential genes, after starting from “central genes” of other measures. Subgraph centrality forms a strong intersection with the influential genes, resulting in no connections between sets. These influential genes are, in a sense, topologically central and connect to important genes found by other centrality measures.
Since the yeast cell cycle has been the subject of many studies, we have data and results from other projects which we can use in the evaluation of the algorithm.
First, we examined the experimental outcomes for yeast genetic experiments found in the SGD [28]. In order of influence ranking, large-scale genetic survey phenotypes are listed in Table 2, as well as PubMed Central IDs for papers showing evidence of cell cycle regulation. If a direct cell cycle related phenotype was found, it was reported in Table 2. But given the close connection between lifespan, metabolism and the cell cycle, if no direct cell cycle phenotype was found, then a related phenotype was reported. It should be noted that even MBP1, which is clearly involved in the G1/S transition, does not have a phenotype listed that directly mentions the cell cycle. Nearly all ranked genes have phenotypes that are in some way related to cell cycle, metabolism, or longevity.
To compute gene set enrichment, over-representation testing was performed using the ConsensusPathDB service, which utilizes a hypergeometric test over a large collection of pathways and gene ontology (GO) terms [29]. P-value adjustment is done using FDR correction and a background of 4,766 genes was used relating to the array used. Gene set enrichment showed that the influence ranked genes were significantly associated with cell cycle related pathways and cell cycle related GO categories. The “regulation of transcription involved in G1/S phase of mitotic cell cycle” GO term (GO:0000083) had a q-value of 1.1e-4, the "regulation of transcription involved in G2/M-phase of mitotic cell cycle" GO term (GO:0000117) had a q-value of 1.08e-3 and the cell cycle phase (GO:0022403) had a q-value of 0.008. The KEGG pathway “Cell cycle—yeast—Saccharomyces cerevisiae (budding yeast)” had a q-value of 0.02.
In Eser et al., the source of the data, 32 hypothesized cell cycle regulators were named [5], five of which are found in the 24 top ranked influencer list. Comparing the top ranked influential genes, we see again that the influential genes are immediately upstream of the Eser TFs (Fig 6), where in total, out of 27 TFs in the network, 15 cell cycle regulators overlap with the influential list, or are regulated by influential genes. Two more, SWI6 and BAS1 were selected as low ranking influential genes (ranks 33 & 34). Therefore, the influential ranked list contained or regulated 63% of the available Eser genes.
Recently a cell cycle model by Tyson et al. that successfully accounts for 257 of 263 phenotypes [30] was published. In total, 29 genes were extracted from the model where complexed genes were considered separately (e.g. SWI6 and SWI4 were used instead of SBF). The full YeastMine network scaffold contained 28 of the 29 genes (CDC55 was not present), and 20 genes were in the TE network. Three genes from the model were ranked as influencers (MBP1, SWI4 and SWI6).
While most of the Tyson model genes are not ranked influencers, they are immediately regulated by influential genes. SWE1 is regulated by 4 ranked genes. CDC20 is regulated by 2 ranked genes. CLB5 is regulated by 2 ranked genes. SIC1 is regulated by 1 ranked gene. So, in almost all cases, the Tyson model genes are not regulated by a single influencer, but by multiple influencers. This shows that even though the mechanistic modelers have different goals–the derivation of small models consisting of well-known elements on multiple levels (protein level and others) that produce a desired behavior, such as cell cycle timing, and timing changes with given mutations–there is a clear relation to the influential genes.
Transfer entropy has been shown to be useful in quantifying information transfer. Here, we showed that using time lagged transfer entropy, along with a permutation testing framework, leads to biologically salient network structures. Even though the network was constructed by considering all possible regulatory edges, it recovers much of the structure and functional enrichment that one would expect, as demonstrated by the lists of genes returned by commonly used centrality metrics, such as betweenness and degree.
Edges with the highest weights, implying greatest information transfer, include (SWI4 → SPT21, TE = 1.57), (TFC7 → MSL1, TE = 1.36), (FKH2 → ALK1, TE = 1.34), (TFC7 → CHL1, TE = 1.27) and (SWI4 → RNR1, TE = 1.27). The source nodes are well-known, multi-functional transcription factors, while the target nodes have more focused functions. SPT21 has a role in regulating transcription through chromatin silencing. MSL1 is involved in mRNA splicing through interactions with the U2 small nuclear RNA. ALK1 is involved in proper spindle positioning and nuclear segregation following mitotic arrest. CHL1 is related to the cohesion of sister chromatids during mitosis. Finally, RNR1 plays an essential role in the cell cycle, assisting with DNA replication and repair. More well-known cell cycle interactions also have high TE edge weights. These include SWI4-SWE1 (TE ranked 7th highest out of 2,084), NDD1-SWI5 (ranked 17/2084), RAP1-FKH2 (ranked 20/2084), and SWI4-YHP1 (ranked 30 / 2084).
Yeast is often used as a model organism in the study of aging. Interestingly, the top two most influential genes, FKH1 and SFP1 have both been related to lifespan [31–34]. The close ties of sources and edge weights to the cell cycle process show that the general dynamics of the cell cycle were captured, reinforcing the usefulness of transfer entropy in biological investigations.
Some well-known cell cycle regulators, such as NDD1, were not selected by influence maximization. In cases such as this, it can often be explained by exploring the immediate neighborhood. In the TE network, NDD1 has upstream regulators FHL1, STB1, SWI4 and SWI6 (three of which are ranked influencers). NDD1 itself targets 18 other genes, all with no influence ranking. Among the targets, we found ALK1, which is also a target from FKH2 as mentioned earlier, as well as CLN1, which is also targeted by three influencers FKH2, SWI4, and SWI6. So, although NDD1 is famous as a cell cycle regulator, when solving the IMP, there are more optimal sources that target the same downstream genes.
When we considered the ranking of influential genes, we saw that high-ranking genes were also more likely to be ranked by other centrality metrics. But there are several notable exceptions. SWI4 and SWI6 were relatively low ranked influencers, but were highly ranked by other metrics. These examples are notable due to their established role in the cell cycle and regular inclusion in models. Proteins SWI4 and SWI6 are members of the SBF complex, interacting with the MBF complex (SWI6-MBP1) to regulate late G1 events. The “low” influence ranking was due to higher ranked influencers being upstream in the regulatory network. Therefore, they were only selected as K, the set of requested influencers, grew large enough.
Network control is one goal in the study of dynamic networks [35,36]. Given that influential nodes seem to have a topologically advantageous position, one could speculate that influential genes might be useful selections for network control. Biological events that impact the influential nodes, thereby affecting normal information flow, could have a strong effect on the network, potentially leading to disease states. Discovering the minimum sets of biological entities that hold the greatest influence in the network context could lead to further understanding of how network dynamics is associated with disease.
The work in this paper can be summarized in a few important steps that are discussed in more detail below: 1) time lagged variants of Spearman's correlation and transfer entropy are described, which were used in constructing the genetic regulatory network; 2) the diffusion model is described, which forms the basis of the score function; and 3) the ant optimization method is described, which was used to maximize the score function, thereby solving the IMP.
The methods described here have been implemented in python and are freely available. Run times are kept low by computing the diffusion using sparse matrix linear solvers, and using a multicore-parallel strategy for performing ant optimization. The network weighting, optimization, and diffusion methods are independent, allowing researchers to "mix-and-match" their favorite modules.
Eser et al. [5] generated time series expression data from two replicates of synchronized yeast producing metabolically labeled RNA levels every five minutes over 41 time points. The expression series spans three cell cycles, which progressively dampen in wave amplitude, as yeast synchrony is lost. Using a model for detecting periodicity in gene expression, 479 genes were labeled as statistically periodic. Additionally, 32 transcription factors were predicted to be cell cycle regulators.
YeastMine, the database of genetic regulatory interactions in yeast (May 2015) [4] provided regulatory edges. Using 6,417 yeast genes, 26,827 genetic regulatory edges were collected. Edge weights were computed using a variation of transfer entropy, as described below.
The Saccharomyces Genome Database (SGD) was used to reference experimental phenotypes and gene annotations [28].
Given two genes connected by an edge, the edge weight was computed in two ways. First, time lagged Spearman's correlation was used with time lags of 0 to 5 steps (0 to 25 mins.), keeping the maximum. Second, time lagged transfer entropy (TE) was used, similar to what is described in [37,38]. TE is computed at each time lag along with a robust distance comparing the observed TE to TEs generated from permuted data. The TE and lag time is returned that maximizes this distance.
Time lagged Spearman's correlation is computed by taking two time series, or numeric vectors x = {x1,x2,…,xn} and y = {y1,y2,…,yn}, and computing the correlation on sub- sequences {x1+k,…xn−1,xn} and {y1 y2,…yn−k}, where k is some integer representing the time lag between variables.
Transfer entropy (TE) is an information theoretic quantity that uses sequence or time series data to measure the magnitude of information transfer between variables [25,38]. Transfer entropy is model-free, directional, and shown to be related to Granger causality [39]. In TE, given two random variables variables X and Y, where X is directionally connected to Y (or X → Y), we would like to know if prior states X help in the prediction of Y, beyond knowing the prior states of Y.
Given two sequences x and y, we describe transfer entropy as
Tx→y(k)=∑yt,yt−1,xt−kP(yt,yt−1,xt−k)logP(yt,yt−1,xt−k)P(yt−1)P(yt−1,xt−k)P(yt,yt−1),
where xt−k indicates value of the sequence at time step t − k.
To perform the computation, first x and y are mean-centered and scaled to be within the range [−1,1]. A Gaussian kernel density estimate (KDE) is fit with a bandwidth given by “Scott’s rule”. Then, a three-dimensional grid is generated by equally spacing some number of points between −1 and 1 in each dimension. Using the grid, points are sampled from the KDE, creating a joint probability distribution, which is normalized in order to sum to 1. The required distributions are marginalized from the joint distribution by summing across the grid. Smaller grid sizes provide a finer grained probability distribution, but slow the computation without changing the values substantially. A three-dimensional grid of 103 points was found to be a good compromise between computation time and accuracy.
A permutation test was performed to assess statistical significance of the transfer entropy, Tx→y. The sequence x was split into a list of subsequences with length 3 and permuted 50,000 times. A robust distance, (Tx→y−Median(Tx→yperm))/MAD(Tx→yperm), was computed where Tx→y is the observed transfer entropy and Tx→yperm is the set of TEs resulting from permuted sequences, and the MAD is the median absolute deviation. The time lag maximizing the robust distance is selected and a p-value is computed by taking a count on the number of times the permuted TE was greater than the observed TE, giving an empirical p-value. Edges were accepted if empirical p-values were less than or equal to 1/(pn + 1), where pn is the number of permutations (pn = 50,000).
The IMP maximizes a network cover based on diffusion. The diffusion model, and most of the nomenclature, is described in [15]. The diffusion models are Markov chains with absorbing states [40]. In the model, vertices are first partitioned into sets S ⊆ V and T ⊆ V, where V is the set of all vertices. The set S contains sources, which in the model are generating information flowing through the rest of the network (nodes in T) until reaching a dead end or absorbing back into S.
The stochastic matrix, defining the probability of moving from one vertex to another, is defined as
pij=wij∑jwij,
where edge weights wij are the weights on outgoing edges. Sets S and T partition the stochastic matrix as
P=[PSSPSTPTSPTT],
where PSS defines the transition probabilities from nodes in S to S, and PST defines transition probabilities from S to T, and so on. Although the matrix is square, it is not symmetric, given the directed edges.
Ultimately, we wish to compute the expected number of visits from a node vi ∈ S, to a node vj ∈ T, defined as matrix H. At time step t, information can travel from vi ∈ S to vj ∈ T directly, or it would already be at adjacent node vk, and would travel from vk ∈ T to vj ∈ T in the next time step. So, at time point t, the estimated number of visits from vi ∈ S to vj ∈ T is given as
hij(t)=pij+∑k∈Thik(t−1)pkj,
where pij is the transition probability of vi ∈ S to vj ∈ T, hik(t−1) is the expected number of visits that have already taken place at time (t − 1), from vi ∈ S to vk ∈ T, and pkj is the probability of the transition from vk ∈ T to vj ∈ T. The matrix form of the equation is
H(t)=PST+H(t−1)PTT.
In the long run, at steady state, when H(t) ∼ H(t−1), the equation reduces to H(I − PTT) = PST, where I is the identity matrix. By taking the transpose of both sides, we have (I−PTT)′H′=PST′. This form lets us avoid the matrix inverse when solving for H, which can be expensive or impossible to compute given that the directed network is represented as an asymmetric matrix. Fortunately, the appropriate iterative solvers are available in the Python SciPy sparse linear algebra library and are robust enough to handle singular matrices.
To compute a measure of influence on the network, after solving for H the expected number of visits on nodes, the influence is summarized as the “influence-score”,
Ωs=∑i∈S{∑j∈TΙ(hij>θ)}
where hij is the number of visitations (using matrix H) from node vi ∈ S to connected nodes vj ∈ T. Indicator function I (hij > θ) is equal to 1 if the number visitations is greater than a threshold θ. The sum of edge weights, ∑i∈S wi, is used as a tie-breaker in the case of degenerate solutions. Degenerate solutions refer to the situation where different solution sets produce an identical cover on the network. In that case, we would like to give preference to the solution that contains nodes with higher overall edge weights, indicating greater degree of information transfer to the network, and potentially greater influence. This influence score is equivalent to computing the cover on nodes in T. In this work, θ = 0.0001 is used, which was selected after observing values in H.
An implementation of the hypercube min-max ant optimization algorithm was used to search for solutions to the Influence Maximization Problem [41,42]. Ant optimization is based on the idea of probabilistically constructing potential solutions to a given problem, in this case a subset selection problem, and reinforcing good solutions with a "pheromone" weight deposited on solution components, ensuring that good solutions become increasingly likely in later iterations.
Since the algorithm is stochastic, and results can vary, the optimization is repeated for a defined number of runs. The main results were produced using a ‘slow’ parameter set, using 8 restarts per value of K, 64 ants, and 16 local optimization steps (full parameterization is given in S1 Text). Each convergence (before restarting) takes a number of iterations where ants construct solutions, perform a local search, score the solutions using the influence score, and reinforce the components in that order. As a run progresses, the pheromone values move to either one or zero, indicating whether the component was selected. The goal of the optimization is to find the subset S ⊆ V of vertices such that
Sopt=argmax{S⊆V:|S|=K}Ωs.
At the start of each iteration, ants construct potential solutions, a subset of vertices, by sampling from nodes using probability distribution
qi=uiαriβ∑uiαriβ,
where qi is the probability for sampling any node vi, with the sum of outgoing edges giving node weight ui and pheromone weight ri. The α and β parameters are used to give importance to either node weights or pheromones. Solutions are constructed by sampling one node at a time. After each sample, the probabilities are renormalized. Here, α and β are set to 1.
Local search is performed by stochastic hill climbing, where we try alternative solutions produced by random single bit flips. If a better score is found, the solution is replaced, and carried forward. Local search has a fairly strong effect on the quality of the solutions, and even a small number of hill climbing steps tends reduce the time required for convergence.
Next, using the influence score function, each potential solution is scored, with the best solution kept and compared to solutions found in earlier runs. As part of the Min-Max algorithm, three solutions are kept throughout the run: the iteration-best, the restart-best and the overall-best. The pheromone updates use a weighted average over the three solutions. At the beginning of the run, the pheromone updates are entirely from the iteration-best solution, but gradually, the updates are increasingly influenced by the restart and overall-best solutions, which is done to avoid local minima. The weighted average pheromone would be ravg = f1bi + f2br + f3bb where bi is the iteration best, br is the restart best, bb is the best overall, and fractions f1 + f2 + f3 = 1. The pheromone updates are defined as r(t+1) = r(t) + d (ravg − r(t)), where r(t) is the pheromone weights at time t, d is the learning rate, and ravg is the average over the three solutions. Eventually, the pheromone weights become sufficiently close to zero or one, and the rate of change among the weights slows. When the difference in sums over the last solution (all r) and the next solution is less than 0.0001, the solution is returned along with the influence score.
BioFabric, R and the R packages igraph, pheatmap and ggplot2 were used for visualization and analysis [43,44,45,46]. Cytoscape 3.5.1 was used for visualizing graphs [47,48]. Pathway and GO term enrichment was generated using the CPDB from The Max Planck Institute for Molecular Genetics [49]. SciPy was used in the software implementation [50].
|
10.1371/journal.pcbi.1003805 | Spatial Heterogeneity and Peptide Availability Determine CTL Killing Efficiency In Vivo | The rate at which a cytotoxic T lymphocyte (CTL) can survey for infected cells is a key ingredient of models of vertebrate immune responses to intracellular pathogens. Estimates have been obtained using in vivo cytotoxicity assays in which peptide-pulsed splenocytes are killed by CTL in the spleens of immunised mice. However the spleen is a heterogeneous environment and splenocytes comprise multiple cell types. Are some cell types intrinsically more susceptible to lysis than others? Quantitatively, what impacts are made by the spatial distribution of targets and effectors, and the level of peptide-MHC on the target cell surface? To address these questions we revisited the splenocyte killing assay, using CTL specific for an epitope of influenza virus. We found that at the cell population level T cell targets were killed more rapidly than B cells. Using modeling, quantitative imaging and in vitro killing assays we conclude that this difference in vivo likely reflects different migratory patterns of targets within the spleen and a heterogeneous distribution of CTL, with no detectable difference in the intrinsic susceptibilities of the two populations to lysis. Modeling of the stages involved in the detection and killing of peptide-pulsed targets in vitro revealed that peptide dose influenced the ability of CTL to form conjugates with targets but had no detectable effect on the probability that conjugation resulted in lysis, and that T cell targets took longer to lyse than B cells. We also infer that incomplete killing in vivo of cells pulsed with low doses of peptide may be due to a combination of heterogeneity in peptide uptake and the dissociation, but not internalisation, of peptide-MHC complexes. Our analyses demonstrate how population-averaged parameters in models of immune responses can be dissected to account for both spatial and cellular heterogeneity.
| Measurements of the rates at which a single cytotoxic T lymphocyte (CTL) can survey for infected cells, and kill them upon encounter, are important for constructing predictive models of vertebrate immune responses to intracellular pathogens. The surveillance rate has been estimated previously using combinations of modeling and experiment, making the assumption that CTL and target cells are well-mixed and that all cell types are killed with equal efficiency. In this study we take an iterative approach with theory and experiment to go beyond such models and detail the effects of cellular heterogeneity, the spatial organisation of the tissue within which killing is taking place, and the influence of the level of expression of peptides on the target cell surface. We demonstrate that determining the degree of co-localisation of effector and target cells, and the level of peptide expression on targets, are most important for improving estimates of CTL killing rates. Further, while the probabilities of killing upon conjugation of CTL with T and B cell targets are similar, T cells take substantially longer to kill than B cells, an effect that may be important when CTL numbers are limiting.
| Cytotoxic T lymphocytes (CTL) prevent the spread of intracellular pathogens through T cell receptor (TCR) recognition of pathogen-derived peptides presented on MHC class I molecules on the surface of infected cells. CTL may have several modes of action but their canonically understood role is to kill cells recognised as infected, either through delivery of lytic mediators through the target cell membrane or engaging ligands on the cell surface that induce apoptosis.
Quantifying the kinetics of CTL killing has been of interest for many years [1]–[19] (see ref. [20] for a review) and is important for at least two reasons. First, knowledge of the rate at which individual CTL can survey and kill cells allows us to derive estimates of the numbers or tissue densities of CTL required to contain an infection. Second, developing tools to measure the kinetics of the different processes involved in lytic activity (locating cells, forming stable conjugates, lysing the infected cell and dissociating from it) may help us to understand how ineffective or exhausted CTL are functionally impaired or to identify bottlenecks in the lytic process that may be potential targets for augmenting CTL responses.
Early studies of CTL-target dynamics were performed almost exclusively in vitro but more recently there has been some focus on data from splenic killing assays, using variants and generalizations of the experimental and modeling approach taken by Barchet et al. [21] and Regoes et al. [11]. There, mice are challenged with a pathogen and following the clearance of infection, a mixture of isogenic splenocytes either pulsed with pathogen-derived peptides or left as unpulsed controls is injected intravenously. Proportions of both populations accumulate in the spleen, where the peptide-pulsed cells can be killed by resident epitope-specific CTL. Models of the kinetics of the transferred cell populations in the spleens in the hours following transfer have yielded estimates of the rate at which single spleen-resident CTL are able to survey and kill.
These models assume that CTL and targets are interacting in a well-mixed environment and have provided reasonable descriptions of the data with the assumption that peptide-pulsed targets are lost with first-order kinetics. When CTL are present in excess – that is, at high effector:target (E:T) ratios – one can assume that the total rate at which targets are killed is not limited by the time each CTL takes to lyse its target [22]. The simplest model of CTL activity then assumes that the per-capita rate of loss of targets is , where is a measure of CTL density or numbers in the spleen. The units and magnitude of dictate the interpretation of the constant , but if is measured as a proportion of all surveyable cells in the spleen, is the rate at which one CTL can move between cells of any type, multiplied by the probability of lysis upon engagement with a peptide-pulsed or infected cell (Text S1, section A). We term the ‘effective surveillance rate’. If killing is assumed to occur with 100% efficiency, is simply the rate of CTL surveillance, and has been estimated to be in the range 1–35 cells per minute in a variety of experimental infection systems [11], [12], [16], [18]. In simple models of well-mixed CTL and targets, knowledge of this parameter, the time taken for CTL to lyse infected cells, and the uncontrolled pathogen growth rate are sufficient to calculate the critical CTL density required to control an infection [22].
These studies took a rather coarse-grained view of CTL-target dynamics that may mask several potential sources of heterogeneity and CTL biology. First, the assays are performed with mixed splenocyte populations but these models assumed all cells are detected and killed with equal efficiencies. Are some cell populations intrinsically easier to kill than others? Second, although CTL have been shown to be able to respond in a dose-dependent manner to very low levels of peptide-MHC (pMHC) ligands on a cell surface [23], [24], it is not clear whether this susceptibility varies across cell types, or at what stage in the killing process any effect of pMHC availability is manifest. Third, the models assume the spleen to be a single compartment, but it is a heterogeneous environment with areas enriched for T cells, B cells and red blood cells, raising the possibility that cell populations are not well-mixed and cells of different types may be exposed to different CTL densities. To explore these issues we used a combination of in vivo and in vitro killing assays to examine the influence of target cell heterogeneity, peptide dose and spatial heterogeneity on the kinetics of CTL activity in vivo.
We revisited the splenic killing assay in the setting of influenza infection (Figure 1). The experimental system is detailed in Materials and Methods, but briefly, TCR transgenic F5 T cells specific for the NP68 influenza epitope were transferred to congenic mice and challenged with systemic administration of live influenza virus. Our calculations depended on being able to enumerate the CTL capable of participating in lysis of pulsed targets. Tetramer staining one week after challenge revealed that transferred F5 cells outnumbered endogenous CTL specific for the NP68 influenza epitope roughly 500-fold in the spleen (Text S1, section B). We therefore assumed that F5 numbers were a reasonable approximation to the number of antigen-specific effector cells in the spleen. To assess the cytotoxic activity of these CTL, total splenocytes from donors were pulsed with four different doses of peptide, with each dose associated with a different level of cell dye, then transferred intravenously to influenza-immunised hosts. T and B cell targets were identified by the expression of either TCR or B220, yielding 8 different target cell populations.
The original model of the kinetics of transferred cells in blood and spleen [11] assumed pulsed and unpulsed cells flow from blood to spleen in the same ratio as they are present in the inoculum, and that transferred, unpulsed cells are not lost from the spleen after entry over the course of the assay. We found direct evidence challenging both of these assumptions (Text S1, section C), so as a starting point we extended the model to allow for (1) a rate of enrichment of unpulsed cells relative to pulsed cells in the blood over time, and (2) a rate of loss of the transferred cell populations from the spleen through egress and/or death due to non-CTL-related mechanisms. The basic model is then represented with the following:(1)(2)(3)(4)where in equations 3 and 4 we have inserted the solutions for the time-dependent densities of transferred cells in blood (equations 1 and 2). All populations are assumed to enter the spleen from the blood at per-capita rate . Unpulsed and pulsed cells die in the blood or migrate into locations other than the spleen at rates and , respectively. is the per-capita rate of loss of pulsed targets in the spleen due to lysis by CTL, and and (0) are the initial concentrations of pulsed and unpulsed cells in the blood, respectively.
The quantity of interest is the ‘fractional killing’ in the spleen or the extent of loss of pulsed relative to unpulsed cells, for each transferred cell population, in which the quantity corrects for departures from a 1∶1 ratio of pulsed to unpulsed cells in the inoculum (see Materials and Methods). The fractional killing therefore lies in the range [0,1]. Using this quantity controls to some extent for variation across animals in the number of cells recovered from the spleen. Because it is dimensionless, units of measurement of cell numbers in blood and spleen do not need to be specified in this calculation. The constant in equations 3 and 4 relates the units of cell numbers in the blood to those in the spleen, and appears linearly in the solutions to equations 1 – 4. In combination with the spleen influx rate it therefore disappears from the ratio (Text S1, section D).
Before using equations 3 and 4 to estimate the killing rate from the fractional killing, the total rate that unpulsed cells leave the blood () and the splenic loss/egress rate () were estimated by fitting equation 3 to the timecourse of unpulsed cells in the spleen (Text S1, section C). We chose to measure cell population sizes as proportions of total splenocytes. This measure exhibited a much smaller coefficient of variation than total numbers in the spleen (Data S1), possibly due to variation in spleen size across animals and associated differences in the total rate of ingress of lymphocytes. The excess rate of loss of pulsed over unpulsed targets from the blood, , was estimated independently from observations of the ratio of pulsed to unpulsed transferred cells in the blood up to 18 hours post-transfer (Text S1, section C). Using these estimates (shown in Table 1) left only the target cell death rate due to CTL, , to be estimated from the timecourse of the fractional killing, . The further assumption of mass-action at the whole spleen level corresponds to where is the effective surveillance rate and is the measured number of F5 CTL in the spleen as a proportion of total splenocytes.
Equations 1-4 predict that the fractional killing should asymptotically approach 1, simply because the per-capita rates of loss of all populations are constant over time and the loss rate is higher for pulsed than unpulsed cells. However the loss of pulsed targets stopped short of 100%, an effect more pronounced for T than B cells and increasingly apparent at lower peptide doses. To account for this saturation, we explored three extensions of the basic model, illustrated in Figure 2 and detailed in Text S1, section D.
We combine egress and non-specific cell death in the spleen in a single loss rate . Similarly the rate combines both death in the blood and the net rate of extra-splenic sequestration of unpulsed targets. An integrated model of spleen-blood kinetics would allow for a proportion of targets lost from the spleen to re-enter the circulation, a process that would contribute to the enrichment for unpulsed cells in the blood over time. Instead we represent the population dynamics of cells in the blood semi-empirically with equations 1 and 2, which recapitulate the kinetics of the pulsed:unpulsed ratio in blood and of unpulsed cells in the spleen (net influx slowing to zero by 6h, followed by a net loss). Killing also slows to near zero by 6h (Figure 3A) and our estimates of parameters related to CTL activity are insensitive to the splenic loss rate , suggesting it is reasonable to use our description of blood kinetics. This assumption also eliminates the need to estimate the potentially different non-specific death rates in blood and spleen.
We fitted these three models to the data, estimating the CTL-mediated killing rate and the additional parameter for each model (, or ) separately for the two target cell populations. To complete the analysis we found that the addition of a third parameter, a time lag , was required to account for the fact that the fractional killing curves did not extrapolate cleanly to zero at the beginning of the assay (Figure 3A). represents a delay between transfer and the first evidence of death of pulsed cells. It might comprise the mean time taken to migrate from blood into the spleen and across the marginal zone into white pulp where we would expect specific CTL to be resident in the greatest numbers. It may also comprise the handling time, the time taken for a target to be killed following conjugation with a CTL. Because splenic effector:target ratios were high in these assays (Text S1, section E), few CTL are expected to kill more than once and so the handling time is not expected to limit the total rate of killing, but it may act to delay the appearance of the first apoptotic cells. If the times to complete any or all of these process are non-exponentially distributed, or if more than one process is operating, there may be a discernable delay before killing is observed. We model the effect of migration and/or handling time by setting the killing rate for times and for , where is time-dependent for the decay model, , and a constant for the hidden-target and hybrid models. We found that our estimates of the time lag did not vary significantly across peptide doses and so assumed that it took values specific to the T or B cell populations only.
Due to the non-nested nature of the models we compared their abilities to describe the data using the Akaike Information Criterion (AIC) [25]. Since the three models contained equal numbers of parameters, the AIC could be calculated simply as n log(RSS) where RSS is the residual sum-of-squares and is the number of datapoints. This quantity is the negative of twice the (maximum) log-likelihood up to an additive constant in which disappears when comparing models. We found strongest support for the hybrid model and the weakest for the decay model, for both T cell and B cell targets (Hidden-target vs. hybrid model, for T cells, for B cells; Decay vs hybrid model, for T cells, 17 for B cells). The interpretation of these differences is that the relative probability of the hidden-target and not the hybrid model minimising the information lost in describing the data is for T cells and for B cells. However the fits were visually indistinguishable on both the absolute scale and the logit-transformed scale on which fitting was performed, and parameter estimates were comparable. Graphs of fits and parameter estimates for the Hybrid model are shown in Figure 3 and parameter estimates for all three models are shown in Table 1. Reduced models with no pre-killing lag time yielded substantially inferior fits (for the hybrid model, for B and T cells respectively).
Our key observations are that every model indicated that within the susceptible populations B cells were killed significantly more slowly per-capita than T cells at all doses except the lowest, and for both populations the rate of killing of susceptible targets fell with peptide dose (Figure 3B and Table 1). While we will argue that spleen-averaged estimates may be misleading, to allow comparison with estimates of the effective surveillance rate from other studies and pathogens, we quote here the spleen-averaged where was the average number of F5 cells as a proportion of splenocytes across all animals and all timepoints. The effective surveillance rates of F5 effectors for T cell targets (with 95% CIs) ranged from 0.14 (0.12–0.16)/min at the lowest peptide dose to 0.44 (0.40–0.50)/min at the highest dose, and from 0.15 (0.13–0.17)/min to 0.27 (0.24–0.30)/min at low/high doses for B cells. Interpreting these figures in terms of numbers of targets surveyed per unit time is complicated by the fact that combines both the base cell-cell surveillance rate and the probability that a CTL kills a pulsed target on encounter, , which may be ; so the estimates of at higher peptide doses where one might expect killing efficiency to be greatest [26] are likely closest to the base surveillance rate but still a lower bound. Mempel et al. [27] used intravital imaging to study the lysis of peptide-pulsed B cell targets by specific CTL in a popliteal lymph node, and estimated that CTL take minutes in vivo to survey cells not expressing cognate antigen. Our figures are consistent with this if one assumes that in their system the time for a CTL to migrate between surveyable cells was short.
Estimating from the target cell death rate also requires the assumption that all target cell populations are exposed to the same spleen-averaged density . This assumption is questionable, as we see below. Another caveat is that is the rate of killing of susceptible targets only; if any targets are refractory or inaccessible to CTL, then models that assume all targets are susceptible such as in [11] will underestimate . Estimates of derived from these models may then also be lower bounds. Nevertheless the effective surveillance rates in the setting of influenza infection and F5 transgenic CTL are slower than those estimated using LCMV or polyomavirus infections [11], [12], [16], [18], an issue we return to in the Discussion.
Modeling splenic egress and enrichment of unpulsed targets in the blood was required to explain the kinetics of unpulsed targets in the spleen and of the pulsed:unpulsed ratio in the blood, respectively (Text S1, section C). Including these processes in the models of killing improved their qualities of fit slightly (for the hybrid model, neglecting splenic egress and blood enrichment increased the AIC by 1.6 for B cells and 1.9 for T cells) but had no substantial impact on parameter estimates or the relative support for the three models. This is likely due in part to a disparity in timescales. Our estimates of the time for the pulsed:unpulsed ratio to halve in the blood were 4.3 h for T cells and 11.4 h for B cells (, Table 1). In contrast, the total rate of influx of unpulsed cells into the spleen is proportional to in these models and was estimated to halve in 1.3 h for T cells and 1.5 h for B cells (, Table 1). Thus, the differential influx of the pulsed and unpulsed cells into the spleen became apparent only after the majority of targets had accumulated. Further, we assumed that the splenic egress or non-specific loss rate applied to pulsed and unpulsed populations equally, and so including this process only weakly influences the timecourse of the pulsed:unpulsed ratio in the spleen. We note that other estimates of times to transit the spleen are lower, and that T cells may egress more rapidly than B cells [28], [29], suggesting that killing in the spleen may indeed contribute to the enrichment for unpulsed cells in the blood.
The delay before killing, , was comparable for T and B cells at approximately 75 minutes. T and B cells take 20–30 minutes following intravenous transfer to cross the marginal zone of the spleen and enter the white pulp [30]. A handling time between encounter and breakup of the target cell has been observed in many studies. A study of killing of B cells in vivo found CTL took between 9–25 minutes to lyse targets after conjugation [27]. Note that the models predicted an increase in the fractional killing in the spleen during the lag period , most strongly for T cells (Figure 3A and Table 1). However by neglecting recirculation we predict this is due to the slow but significant enrichent for unpulsed over pulsed T cell targets from the blood into other anatomical locations, and not killing in the spleen itself.
To investigate whether we could detect evidence of mass-action killing operating, following ref. [12] we allowed for the base killing rate to be specific to each mouse, , and assumed it was linearly proportional to the density of effector CTL in the spleen of that animal, . Here is the measured number of F5 CTL as a proportion of all splenocytes, and is the effective surveillance rate. We assumed that was constant across animals but took values specific to each target cell type (T or B cell) and peptide dose.
Effector:target ratios were greater than 3 in these assays (Text S1, section E), so if CTL and targets were well-mixed in the spleen and moving randomly, mass-action might be expected to hold. However, we found that across all peptide doses and target cell types the models using a simple population-average killing rate described the data significantly better than those with mouse-specific, mass-action killing rates . Indeed we saw no significant positive correlation between fractional killing and F5 CTL, either in total numbers or as a fraction of splenocytes, at any peptide dose or timepoint, after correcting for multiple comparisons (Text S1, section F). We saw roughly 40% variability across animals in the total number of F5 CTL recovered from the spleens, . However the dependence of the rate of loss of targets on the size of the CTL population might be expected to be an increasing function of their density, rather than of total numbers alone; the rate of encounter of targets with a given number of randomly dispersed CTL would be expected to vary with spleen volume if the populations are well-mixed. F5 CTL numbers as a proportion of all splenocytes were highly consistent across animals and timepoints (, or variation). We conclude that our data do not provide sufficient power to support or refute the mass-action hypothesis at the whole-spleen level, nor do they allow us to quantify any functional dependence of killing rates on CTL densities. Ganusov et al. [19] performed these assays in the context of LCMV infection, using adoptive transfer of specific CTL and varying their number over three orders of magnitude, and did indeed find evidence for a linear dependence of splenocyte death rates on CTL numbers or frequencies, and so in the analyses below we retain the possibility that mass-action operates.
Another key observation derived from the hybrid model is that the average rate at which pulsed target cells lost susceptibility increased by an order of magnitude as peptide dose fell from 10−6 M to 10−9 M, most rapidly for T cells. This effect was reflected in the hidden-target model by the susceptible fraction declining with dose, and generally being smaller for T cells than B cells (Figure 3B and Table 1). These dose-dependencies suggest that heterogeneity in susceptibility derives from properties intrinsic to target cells rather than global effects such as cells migrating into regions in the spleen inaccessible to CTL.
In the light of these observations, the following is a mechanism of loss of susceptibility compatible with the hybrid model:
Peptide-pulsed targets would be expected to exhibit a unimodal distribution of pMHC densities at each dose, and if peptide is lost this distribution will shift towards lower pMHC densities with time. The lower the dose of peptide, the closer to any threshold of detection the initial distribution will be, and so the average rate of loss of susceptibility across the entire target population (the rate in the hybrid model) will increase, as observed.
To explore the peptide-loss hypothesis, we quantified the expression and turnover of MHC class I on T and B cells in vitro by blocking transport of MHC class I to the cell surface and observing the kinetics of its internalisation (see Materials and Methods). We found T cells expressed MHC class I on the cell surface at 2–3 fold lower levels than B cells, and MHC was lost approximately twice as rapidly (Figure 4; half-lives of approximately 11 and 21 h for T and B cells respectively). If these peptide-MHC turnover rates in vitro reflect the rate of turnover in vivo, then in our splenic cytotoxicity assays pMHC densities on the T cell targets fall only approximately 2-fold before the bulk of surviving pulsed cells are below a threshold of detection and killing has stopped. This means that for loss of MHC to underlie arrest in the hybrid model, peptide doses covering four orders of magnitude result in four target cell populations clustered closely in pMHC expression just above the limit of detection. Parsimony then suggests that it is unlikely that MHC internalisation alone explains the arrest of killing, as suggested previously [13].
MHC turnover imposes only a lower limit on the rate of loss of visibility of targets, as peptide may dissociate from the MHC class I to which it is bound. While we do not have an estimate of the lifetime of the NP68/H-2Db complex used in our assays, peptide-MHC class I dissociation half-lives of 1–4 h have been reported in other systems [31]–[33] and so loss of peptide from targets remains a potential explanation for loss of susceptibility. However we cannot exclude a contribution from heterogeneity in peptide uptake by targets such that at decreasing peptide doses, an increasing proportion of pulsed targets are already below a threshold of detectability by CTL at the beginning of the assay. The hidden-target model, which describes the data with fidelity comparable to the hybrid model, at least visually if not statistically, is an expression of this heterogeneity in initial conditions.
Finally, the dependence of the B cell killing rate on peptide dose was weaker than that of T cells (Figure 3B and Table 1). One explanation for this is that the higher level of MHC expression on B cells means at higher peptide doses the population lies more completely within a saturating region of the curve relating dose to susceptibility to CTL.
All three models indicated that the per-capita rate of killing by CTL was lower for B cell targets than for T cells. This may be because B cells are intrinsically less susceptible to lysis at a given peptide dose, that B and T cells encountered different local densities of CTL, or that the CTL were less motile in B cell areas than in T cell areas. To begin to discriminate between these (non-exclusive) possibilities, we used microscopy to quantify the distribution of effector CTL within the spleen (Figure 5A). F5 CTL were indeed distributed heterogeneously, with the majority (60%) in T cell areas, roughly 4-fold fewer (13%) in B cell follicles, and more than a quarter (27%) in red pulp (Figure 5B).
Since at least a proportion of the intravenously injected T and B cell targets may migrate to their respective areas in the spleen, it then seems likely that the different target cell populations are exposed to different local densities of CTL. To further explore the relative contributions of this spatial heterogeneity and potential differences in susceptibility to lysis, we used this spatial data together with the T and B cell death rates to estimate the relative ability of F5 CTL to kill peptide-pulsed B and T cells. To separate the effects of CTL numbers and susceptibility we began by assuming that mass-action held within the T and B cell areas and killing of each population was restricted to its relevant area. Targets and CTL might reasonably be assumed to be randomly distributed within each.
In the hybrid and hidden-target models the per-capita rate of killing of the susceptible population of peptide-pulsed T cells in the spleen is . If these T cells are restricted to T cell areas and are exposed to specific CTL at local density , this killing rate must be equal to where is the effective surveillance rate in T cell areas. So , and similarly for B cells. Note that this calculation does not depend on the relative volumes of B and T cell regions in the spleen, which we estimated to be (mean s.e.m.) by raising the ratio of their areas in the imaged sections to the power . Different degrees of crowding of T and B cell targets in the spleen are represented by different values of the parameter , which relates density units in blood and spleen but disappears from the estimation of for each population. Then(5)Here and are decomposed into the probabilities of lysis following encounter with a CTL, and , and the base CTL surveillance rates in B and T cell areas, and .
In this calculation we have replaced the local densities of CTL in each of the T and B cell areas (as fractions of total splenocytes within each region) with and , the measured local CTL densities in units of cells per of spleen section. Because these sections were of the order one cell width deep, the CTL densities measured per unit volume are then approximately these cell numbers per unit area divided by the section depth. We assume the densities of total surveyable cells are the same in T and B cell areas, and so . Using the target cell death rates derived from the hybrid model, we estimate that on a per-CTL basis B cells are killed 3-5 times more rapidly than T cell targets, with the lower boundary of the 95% confidence interval lying above 1, at all peptide doses (Figure 5C, solid circles). The hidden target model yielded very similar estimates (Figure 5C, open circles).
In summary, under the assumptions of mass-action and restriction of the populations to their respective areas in the spleen, the difference in local densities of CTL was too large to explain the difference in killing rates of T and B cells pulsed with the same dose of peptide, and so the relative paucity of CTL in B cell areas is compensated to a degree by a higher effective surveillance rate ().
This difference in might stem from susceptibility to lysis; for instance, the 2–3 fold difference in MHC expression by B cells might contribute to a higher probability of detection by CTL upon encounter, , at a given peptide dose. It might also arise from differences in the motility of CTL within T and B cell areas, and . To narrow down the possibilities even further, we wanted to estimate the intrinsic susceptibilities of T and B cells to lysis by CTL and assess their dependence on peptide dose, while minimising any effects of spatial heterogeneity or CTL motility.
Lysis is a multi-stage process. The CTL must encounter and survey the cell, detect that it bears the relevant peptide, form a stable conjugate, initiate lysis and eventually disengage from the apoptotic cell. We wanted to identify at which stage(s) of the killing process any differences between T and B cells or across peptide doses were manifest most strongly. To do this we performed an in vitro cytotoxicity assay using T and B cell targets pulsed as before with different doses of the peptide, and co-localised with F5 CTL activated in vitro (see Materials and Methods). These populations were co-cultured for between 0 and 120 minutes, allowing us to follow the kinetics of free targets (S), the number of CTL-target conjugates in which the CTL had not degranulated (stained negative for LAMP1a at the cell surface, ), and the number of conjugates in which the CTL had degranulated (LAMP1a detected at the cell surface, ), assumed to indicate lysis.
The following generalisation of the hybrid or decay models described the kinetics of these populations well (Figure 6):(6)(7)(8)
The term proportional to is the total rate of formation of conjugates between targets and LAMP1a− CTL. At the beginning of the assay the expected time for a given target to become conjugated is . is the initial rate at which a conjugate dissociates without lysis; is the initial rate at which a CTL in a conjugate becomes LAMP1a+ through degranulation; and is the rate at which a LAMP1a+ conjugate dissociates. CTL were in excess in this assay and so we assumed LAMP1a+ cells did not kill again. With this assumption, LAMP1a+ conjugates arose directly from LAMP1a− conjugates only.
Inspection of the data revealed that the formation of conjugates and killing slowed considerably over the course of the assay, appearing to stop completely after roughly an hour (Figure 7). This was much earlier than the timescale of arrest of killing in vivo (Figure 3A), and seems unlikely to be the result of loss of peptide-MHC from the target cells. We observed a roughly 50% loss of effector CTL numbers over the 2h timecourse, accounting for both free CTL and those in conjugates. We speculate that as well as dying, the CTL became increasingly functionally impaired, perhaps related to the release of cytotoxic factors into the culture medium. To capture this behaviour we allowed rate constants to change with time such intially they reflect the interactions between targets and fully-functional CTL, but by , conjugate formation had stopped, and the efficiency of progression to degranulation was zero;(9)
Parameter estimates are shown in Figure 8 and rate constants are quoted as their inverses (timescales, in minutes). CTL and targets formed conjugates at similar rates () for T and B cells at each peptide dose, but conjugates were slower to form at lower doses (Figure 8A). CTL-B cell conjugates progressed to degranulation (LAMP1a+, which we assumed led to lysis) after roughly 15 minutes, independent of dose, while the mean time to degranulation for CTL-T cell conjugates was roughly 30 minutes slowing to 45 minutes at the lowest peptide dose (Figure 8B). We saw considerable uncertainty in the rate of dissociation without lysis, the failure rate (Figure 8C), but this process was relatively slow and the mean lifetime of LAMP1a− conjugates was determined largely by the time to degranulation (Figure 8D). This led to high (50–90%) efficiencies of lysis, , at the beginning of the assay , but uncertainty in obscured any potential variation in efficiency with peptide dose (Figure 8E). Similarly we detected no significant differences in the rate of change of parameters, indicating that a progressive loss of CTL functionality affected the killing of all cell populations equally (Figure 8F). Lastly, we estimated that degranulated (LAMP1a+) conjugates took between 100 and 200 minutes either to break up or for the target cell to disintegrate (Figure 8G), again with no significant T-B differences.
Multiple CTL bound to single targets may shorten the time taken to kill [4]. We saw evidence for formation of conjugates comprising more than two cells at approximately 10% of doublet numbers after one hour, and stable over time (Text S1, section G). The data were not sufficient to parameterise the dynamics of theses multiples. However, one could reasonably assume that triplets form by a second CTL joining a CTL-target doublet, particularly because CTL were in excess (Text S1, section G). The estimated pre-lytic doublet loss rate will then comprise both breakup into singlets and formation of LAMP1a− triplets, then quadruplets, etc., within which targets may or may not have an increased probability of being killed. Therefore, by neglecting multiples we may overestimate the true conjugate breakup rate , and so our efficiency of killing is a lower bound, with an error of 10% or less. Again, CTL are in excess in these assays and so estimates of the doublet conjugate formation rate and the doublet lysis rate will be unchanged.
In summary, we found that while the rate of conjugate formation fell with peptide dose, there were no detectable differences in the ability of CTL to conjugate with T or B cells; and while T cells subsequently progressed to lysis more slowly than B cells, there were no detectable differences in the efficiency of lysis across cell type or peptide dose. A similar conclusion regarding the effect of peptide dose on conjugation was reached by Jenkins et al. [34], who measured the impact of the avidity of TCR-pMHC interations on lysis using transgenic OT-I CTL specific for the OVA257–264 peptide. There, the avidity of TCR interactions, assumed positively correlated with peptide dose, impacted the rate of formation of conjugates but had no significant effect on the proportion of conjugates exhibiting clustering of tyrosine kinases at the contact site, an early indicator of TCR signaling and progression to lysis. In contrast, lytic efficiency was found to vary with dose in an in vitro tissue model of killing of HIV-derived peptide-pulsed targets [26]. Measurements of rates of conjugate formation and lytic efficiency are somewhat definition-dependent and correlated, however. We may be overestimating lytic efficiency and underestimating the rate of conjugate formation since what we define as a conjugate has remained stable for long enough to be detected by flow cytometry.
The in vivo killing assay and the imaging indicated that if killing of T and B cells was restricted to their respective areas in the spleen and rates were locally linear in CTL numbers, differences in local CTL densities were too great to explain the differences in killing rates of the two populations. This suggested that the effect of excess CTL in T cell areas may be partly compensated by more efficient CTL surveillance in B cell areas – either by an increased rate of encounter with cells of all types, or by B cells being killed with a higher probability than T cells following conjugation (Figure 5C). However the rate of formation of conjugates with CTL, and the probability of progression to lysis, were indistinguishable for T and B cells in vitro, and so together these assumptions and observations prompt the conclusion that the rate at which CTL survey cells of any type is higher in B cell follicles than in T cell areas.
While that remains to be tested, the assumption of killing of each population being entirely restricted to their respective areas in the white pulp is questionable. Both target cell populations enter the spleen through the circulation and enter the white pulp via the marginal zone. At least a proportion of B cells then migrate through CTL-rich T cell areas en route to B cell follicles [30]. Lymphocytes egressing from the spleen do so by transiting the red pulp [29], where more than a quarter of splenic F5 CTL resided in our assay (Figure 5B). Bajénoff et al.[30] observed that between 3 h and 8 h after intravenous transfer of isolated splenocytes, B cells were continuing to accumulate in the white pulp from the marginal zone that separates the red and white pulp, roughly a third were resident in B cell follicles, and the remainder were co-localised with T cells. If similar migration patterns and kinetics apply in our assays, the difference in the average CTL densities encountered by the two target cell populations over the assay may be smaller than that inferred simply from the CTL densities in T cell areas and B cell follicles alone. It is possible that this effect alone may account for the differences in T-B killing rates. It is also not critically dependent on the mass-action assumption, requiring only that death rates are increasing functions of CTL density over the conditions found here.
While we saw no significant differences in the ability of CTL to conjugate with T or B cells in vitro, or in the probability of conjugation resulting in lysis, CTL-T cell pairs took more than twice as long to either break up or progress to lysis. This difference in handling time will not affect the ability of CTL to control an infection when they are in excess, but will become important at lower E:T ratios when an increasing proportion of CTL will be sequestered in conjugates at any time, and so may become limiting [22]. This difference may make growing populations of infected T cells intrinsically more difficult to control than B cells, in the absence of spatial or peptide-dose effects. Note here we are referring to the rate of progression to lysis following encounter, or single-cell behaviour. This is distinct from the population-level killing rate which depends only the rate at which CTL can encounter and identify targets, and not on the handling time, when CTL are in excess or the mean time to locate the next pulsed target us much longer than the handling time.
We predicted a delay of more than an hour before killing of targets within the spleen was evident. However in an LCMV infection model, Barber et al. [35] observed substantial loss of peptide-pulsed cells in the spleen within 15 minutes, relative to target numbers in uninfected control animals. A similarly rapid decline of pulsed relative to unpulsed targets was observed in a polyoma virus infection model [16], [36]. It is possible that this faster loss derives in part from the systemic nature of those infections, which might lead to greater extra-splenic sequestration or killing of pulsed targets in rapidly perfused organs such as liver and lung. Indeed in an LCMV model, Graw et al. [18] saw a roughly four-fold enrichment for unpulsed transferred cells in the blood by 4 hours, compared to the two-fold enrichment in our assay (Text S1, section C). However Barber et al. [35] saw that the early loss of pulsed targets in the spleen was attenuated in mice with CTL lacking Perforin, a membrane pore-forming protein involved in the delivery of cytolytic molecules to the target cell. We might expect filtering of pulsed targets from the blood by 15 minutes to be similar in these and WT mice, since it is initially TCR- and not Perforin-dependent. This strongly suggests that lysis was indeed occurring in the spleen within 15 minutes of cell transfer.
Estimates of the time CTL take to kill targets have varied widely across systems, from minutes [26], [27] to hours [37], and so a discrepancy of this magnitude is perhaps not surprising. F5 CTL may simply take longer to kill; we found mean handling times of 30 minutes or longer for T cell targets in vitro, although handling times with B cell targets were shorter (Figure 8B). The longer delay before killing is apparent may also derive from the time taken for CTL and targets to encounter each other. By day 7 the influenza infection is well controlled and so levels of inflammation are likely lower than in the LCMV system, which we speculate may result in reduced CTL motility; and the spatial distribution of specific CTL that we found in our system may differ from those in LCMV infection models [38], which might result in differences in the mean time for ingressing targets to enter CTL-rich areas of the spleen.
Our results recapitulate previous findings that peptide dose influences susceptibility to lysis by CTL (see, for example, [16], [24], [34]). Threshold effects have also been observed. Purbhoo et al. [24] demonstrated a sigmoid relation between peptide dose and the extent of lysis at one timepoint in an in vitro cytotoxicity assay, with a location and steepness that varied with the particular TCR and peptide but over ranges of peptide doses comparable to ours. They showed that as few as two pMHC within the interface between the T cell and its target were sufficient to induce lysis at least in a proportion of contacts, an effect saturating at between 4-200 pMHC, consistent with other studies [39], [40]. Along similar lines, Henrickson et al. [33] showed in an LCMV model that a sharp threshold of peptide dose given to dendritic cells (DC) exists for activation of specific CD8 T cells, corresponding to between 30 and 60 pMHC complexes per DC.
If dissociation of peptide from MHC generates the refractory or ‘invisible’ targets in the in vivo assay, these results suggest that these targets have reached very low surface densities of specific pMHC. It is then possible that the greater proportion of refractory cells among the T cell targets derives in part from their 2–3 fold lower levels of MHC class I expression (Figure 4). Further, the fact that we and others [24] observe incomplete killing even in the populations receiving high doses of peptide suggests heterogeneity in peptide uptake can be substantial.
Our work builds on other studies that used the splenic killing assay and exposes different sources of heterogeneity that need to be considered when estimating rates of CTL surveillance. However, the issues that we raise highlight the need for measurements of CTL efficacy performed with live replicating pathogens in relevant tissues, for several reasons. First, there appears to be considerable variation across in vivo cytotoxicity assays in the parameters defining CTL activity, likely deriving from differences in microenvironment, TCR specificity, the mode of CD8 T cell priming and hence effector quality, and target cell susceptibility. Second, both the in vitro and in vivo analyses confirmed earlier findings that peptide dose influences the ability of CTL to detect pulsed targets, but it is not known what peptide doses yield physiologically relevant levels of cognate pMHC on target cells. Third, the influenza infection is well-controlled by the time of the assay 7 days post-challenge; inflammation in the spleen even while killing of peptide-pulsed targets is occurring is presumably low, and so CTL motility and any ability to home to targets may differ between this scenario and one in which an infection is ongoing. Finally, while we have focused on the lytic mode of CTL action, they may also control the spread of intracellular pathogens by non-lytic mechanisms [41], [42] that will presumably not be manifest in assays using peptide-pulsed targets.
The parameters defining how CTL survey and kill infected cells are key elements of models of the within-host dynamics of intracellular pathogens. Deterministic models assuming homogeneous mixing of components of the immune system and infected cells have been used widely and have provided many mechanistic insights into the progression and control of viral infections (for a review, see for example Ref. [43]). While the functional forms of the terms in these models may be appropriate for describing the dynamics of an infection, the parameters they contain are usually compound quantities and may implicitly average over spatial and cellular heterogeneity. Characterising this heterogeneity is important when attempting to make more detailed quantitative statements regarding host-pathogen interactions. For example, a substantial number of CTL in our in vivo assay resided in the red pulp, and would only have been encountered by the proportion of splenocytes that egress from the white pulp over the course of the assay. Depending on the time take to transit the red pulp, it may be that these CTL contribute very little to killing of targets. Estimates of per-CTL killing rates will then be too low if these CTL (enumerated following the homogenisation of whole spleens) are assumed to be co-localised with splenocytes only. With increasing availability of in-vivo imaging data, quantitative immunologists will be able to characterise the within-host ecology of infections in more detail, and specifically the critical sizes of effector cell populations needed for immunity.
The UK Home Office Project Licence 80/2506 (Development and function of innate and adaptive immune responses) covers all animal experiments conducted at the NIMR.
Ly5.1 C57BL/6J, Ly5.2 C57BL/6J, and F5.Rag1-/- mice were bred and maintained in a conventional pathogen-free colony at the National Institute for Medical Research, London, UK. All lines were of H-2b haplotype. Animal experiments were performed in accordance with UK Home Office regulations.
The following monoclonal antibodies and cell dyes were used: CD45.2 PE-Cy7, CD45.1 FITC, TCRβ APC, B220 PE-TexasRed (all eBioscience), H-2Db PE (BioLegend), LiveDead nearIR and CellTrace Violet (both Invitrogen), and H-2Db-ASNENMDAM dextramer-PE (Immudex). Samples were acquired on CyAn ADP (Dako Cytomation), Canto-II (BD) or Fortessa X20 (BD) flow cytometers, and analysis was performed with FlowJo software (Treestar).
Cell culture medium was RPMI supplemented with 10% FCS, 2mM glutamine, 1% penicillin/streptomycin and 50 µM β-mercaptoethanol (all Sigma).
Splenocytes from naive Ly5.1 or Ly5.2 C57BL/6J mice were cultured with NP68 peptide (influenza NP366-374, strain A/NT/60/68, ASNENMDAM, Mimotopes) at 10−6 M, 10−7 M, 10−8 M, 10−9 M, or in culture medium alone for unpulsed cells, for 2 hours at 37°C. These cells were then labelled with CellTrace Violet (CTV) at either 10 µM, 2.5 µM, 625 nM, 156 nM or 40 nM, respectively. Following peptide pulse and CTV labelling, target cells were mixed together in equal ratios.
Ly5.1 C57BL/6J mice were injected IV with 2 million lymph node cells from Ly5.2 F5.Rag1-/- mice and A/NT/60-68 influenza virus, to generate a spleen-resident population of NP68 specific CTL. Seven days later, 10 million Ly5.2+ target cells were injected per recipient mouse. At indicated timepoints from 0.5–24 hours after injection of targets, mice were sacrificed and spleen and blood were harvested for analysis by flow cytometry. Care was taken in the timing of both injection and sacrifice for individual mice, and organs were harvested directly into ice cold media, to ensure an error of no more than 5 minutes in the reported timepoints. Target and effector cells were distinguished from host cells by expression of Ly5.2; CTV fluorescence was used to identify target cells that had been pulsed with different doses of peptide, while effector cells were CTV-unlabelled. Staining for TCRβ and B220 was used to identify T and B cell targets.
To generate effector CTL, lymph node cells from Ly5.2 F5.Rag1-/- mice were activated in vitro for three days in the presence of NP68 peptide (10−8 M). Activated blasts were purified by Ficoll (GE Healthcare) density-gradient centrifugation and expanded for a further four days in the presence of 10 nM IL-2 (Peprotech). Ly5.1+ target cells were prepared as described above. CTL and target cells were added to wells at an E:T ratio of at least 5∶1, and briefly centrifuged to initiate cell contact. Cells were co-cultured at 37°C for the indicated period of time (10 minutes – 2 hours) in the presence of anti-LAMP1a (eBioscience) to detect degranulation of CTL during the culture period. At the end of the culture period, cells were immediately fixed with IC fixation buffer (eBioscience) to preserve E:T conjugates. Samples were then stained and analysed by flow cytometry, with the addition of a known number of AccuCount fluorescent particles (Spherotech) to determine cell counts. Target and effector cells were identified by expression of Ly5.1 or Ly5.2 respectively, and E:T conjugates by dual fluorescence for these markers along with forward scatter area and width characteristics to identify doublets. Staining for TCRβ and B220 was used to identify T and B cell targets, and CTV fluorescence to identify cells that had been pulsed with different doses of peptide.
Ly5.1 C57BL/6J mice were injected IV with 2 million lymph node cells from Ly5.2 F5.Rag1-/- mice and A/NT/60-68 influenza virus, to generate a spleen-resident population of NP68 specific CTL. Seven days later, at the time when in vivo cytotoxicity assays were performed, mice were sacrificed and spleens harvested for analysis. Each spleen was cut in two, and the weight of each segment recorded. One segment was processed for cell counting and analysis by flow cytometry; the other segment was immediately frozen in liquid nitrogen. Frozen spleen segments were then embedded in OCT compound (VWR International). At least three non-consecutive sections 7 µM thick were cut from each spleen, and stained with antibodies to Ly5.2, IgD and CD4 (all eBioscience). Separate images for each fluorescence channel were collected at 20x magnification on a Leica SP5 confocal microscope, and analysed using ImageJ software (NIH). IgD and CD4 fluorescence was used to manually identify regions of interest corresponding to B cell zones, T cell zones and red pulp, and Ly5.2 fluorescence was subsequently used to enumerate CTL within each of these regions.
Single cell suspensions were prepared from the spleen of C57BL/6J mice and incubated at 37°C in culture medium for the indicated periods of time in the presence of 5 µg/mL Brefeldin A (Sigma) or vehicle control (DMSO, Sigma). Cells were then washed with PBS and stained for TCRβ, B220 and H-2Db for analysis by flow cytometry.
Ordinary differential equation models, described in Results, were used to simulate the flux of peptide-pulsed and unpulsed splenocytes from blood into the spleen and the killing of pulsed targets within the spleen. Parameters were estimated separately for T and B cell target populations at each peptide dose by fitting these models to the logit-transformed fractional killing , where and are the numbers of pulsed and unpulsed transferred cell populations recovered from the spleen. The correction factors were close to unity and were the ratio of each peptide-pulsed population to the unpulsed population in the inoculum. Estimates of were obtained from the transfer of targets taken from the prepared splenocyte population into naive animals, and observing the proportions of the different target cell populations as they flowed into the spleen. Closed-form solutions to the models were obtained using Mathematica [44] and fitted to the data using the nls function in [45]. Data were logit-transformed to ensure the normality and heteroscedasticity of the distribution of residuals. Arcsin square root, complementary log-log and probit transforms yielded similar parameter estimates and qualities of fit.
All data used in this manuscript are provided as Supporting Information (Data S1).
|
10.1371/journal.pgen.1000751 | Dosage Regulation of the Active X Chromosome in Human Triploid Cells | In mammals, dosage compensation is achieved by doubling expression of X-linked genes in both sexes, together with X inactivation in females. Up-regulation of the active X chromosome may be controlled by DNA sequence–based and/or epigenetic mechanisms that double the X output potentially in response to autosomal factor(s). To determine whether X expression is adjusted depending on ploidy, we used expression arrays to compare X-linked and autosomal gene expression in human triploid cells. While the average X:autosome expression ratio was about 1 in normal diploid cells, this ratio was lower (0.81–0.84) in triploid cells with one active X and higher (1.32–1.4) in triploid cells with two active X's. Thus, overall X-linked gene expression in triploid cells does not strictly respond to an autosomal factor, nor is it adjusted to achieve a perfect balance. The unbalanced X:autosome expression ratios that we observed could contribute to the abnormal phenotypes associated with triploidy. Absolute autosomal expression levels per gene copy were similar in triploid versus diploid cells, indicating no apparent global effect on autosomal expression. In triploid cells with two active X's our data support a basic doubling of X-linked gene expression. However, in triploid cells with a single active X, X-linked gene expression is adjusted upward presumably by an epigenetic mechanism that senses the ratio between the number of active X chromosomes and autosomal sets. Such a mechanism may act on a subset of genes whose expression dosage in relation to autosomal expression may be critical. Indeed, we found that there was a range of individual X-linked gene expression in relation to ploidy and that a small subset (∼7%) of genes had expression levels apparently proportional to the number of autosomal sets.
| Many organisms have a single X chromosome in males and two in females, leading to a chromosome imbalance between autosomes and sex chromosomes and between the sexes. In mammals, this dosage imbalance is adjusted by doubling expression of X-linked genes in both sexes and by silencing one X chromosome in females. We used expression array analyses of human triploid cultures to test X chromosome expression in the presence of three sets of autosomes and address the question of an autosomal counting factor. We found that overall X-linked gene expression is not tripled in the presence of three sets of autosomes. However, in triploid cells with a single active X chromosome, its expression is adjusted upward presumably by an epigenetic mechanism that senses the active X/autosome ratio. Based on the range of individual gene expression we identified a subset of dosage-sensitive genes whose expression is apparently proportional to the ploidy. Our findings are important for understanding the regulation of the X chromosome and the role of ploidy in the balance of gene expression and associated phenotypes.
| Dosage compensation restores a balanced network of gene expression between autosomes and sex chromosomes in males (XY) and females (XX) [1]. Strategies to achieve this vary among species [2]. In Drosophila, a male-specific ribonucleoprotein complex binds to the X chromosome and modifies chromatin structure to increase expression of most X-linked genes by two fold. X up-regulation also occurs in C. elegans and in mammals but in both sexes [3],[4]. Silencing of one X chromosome in mammalian females and repression of both X chromosomes in C. elegans hermaphrodites have been adapted to avoid hyper-expression in the homogametic sex. While these repressive processes have been well studied, the mechanisms of X up-regulation in mammals and worms remain to be determined.
Human triploids occur in about 1% of conceptions. Depending on their sex chromosome composition (XXX, XXY, XYY), triploid cells show a variety of X inactivation patterns [5]–[8]. Female triploid (XXX) fibroblast clones with either one or two active X chromosomes (Xa) have been successfully established [5]. These cloned lines with a defined number of Xa's as well as XYY triploid cell cultures provide a mean to test X chromosome expression in the presence of three sets of autosomes, which could help understand the underlying mechanisms of X up-regulation and test the potential existence of an autosomal counting factor. If DNA sequence changes that affect promoters, enhancers, and/or regions that alter mRNA elongation and stability were sufficient to mediate the two-fold elevated expression of X-linked genes in normal diploid cells, we would predict that the X:autosome (X:A) expression ratio would be 0.67 (i.e. 2/3) in triploid cells with a single Xa and 1.33 with two Xa's, assuming that autosomal genes have a steady expression level per gene copy and thus produce approximately 1.5-fold more products in triploid versus diploid cells. A second possibility is that epigenetic chromatin remodeling of the X chromosome and/or of the autosomes may modulate the balance of gene expression, in which case the X:A expression ratio in triploid cells with one Xa or two Xa's would differ from 0.67 and 1.33, respectively. If this epigenetic regulation of the active X chromosome in triploid cells was based on a counting mechanism mediated by autosomal factor(s) that would triple the output from the Xa, the predicted X:A expression ratio in triploid cells with one Xa would be 1 (i.e. 3/3), similar to the situation in diploid cells [4], whereas triploid cells with two Xa's would have a ratio of 2 (i.e. 6/3). Alternatively, both genetic and epigenetic mechanisms could be involved in regulation of X up-regulation, in which case the predicted ratios would be further adjusted. A perfect adjustment to a ratio of 1 could possibly be observed if X-linked and/or autosomal gene expression was adjusted in response to the number of active X chromosomes in relation to autosomes. Studies in Drosophila have shown that both X-linked and autosomal gene expression can respond to abnormal genotypes [9].
In the present study, we used expression array analyses to determine the global expression of X-linked genes versus autosomal genes in triploid cell cultures with either one or two Xa's. We found no evidence of a global change in absolute autosomal expression levels per gene copy in triploid versus diploid cells. Our results are consistent with a doubling of expression of most X-linked genes in triploid cells with two Xa's. However, X expression was further increased in triploid cells with one Xa, suggesting an epigenetic adjustment in response to the Xa/autosome ratio. The expression of genes with copies on both sex chromosomes and of genes that escape X inactivation was investigated in triploid cells. We determined that XYY triploid cells had a expression deficiency in genes that escape X inactivation compared to XXX triploid cells. We also found that expression of a small subset of X-linked genes apparently responds to the cell ploidy, suggesting that dosage of these genes is critical.
Six XXX triploid fibroblast clones with either one or two Xa's (XaXiXi or XaXaXi, Xa and Xi being the active and inactive X, respectively) [5], and three XYY triploid fibroblast cultures were analyzed along with two male (XY) and two female (XaXi) control diploid fibroblast cultures using Affymetrix whole genome expression arrays. As expected, while XIST expression was absent in XYY cultures, it was two-fold higher in XaXiXi versus XaXaXi cultures, consistent with the number of Xi's (Figure S1) [5]. For each cell line the X:A expression ratio was calculated by dividing the mean expression of 362 X-linked genes (corresponding to 550 probe-sets) by that of 10,267 autosomal genes (corresponding to 16,984 probe-sets) expressed in all cultures. The average X:A expression ratios were 1.04±0.02 and 1.14±0.03, for control male and female diploid fibroblasts, respectively (Figure 1 and Table S1). Analysis of additional published microarray data for 11 male and 12 female adult fibroblasts yielded X:A expression ratios of 1.09±0.04 and 1.11±0.04, respectively, based on 397 X-linked and 8,832 autosomal genes [10]. These analyses are consistent with functional dosage compensation in normal diploid somatic cells and slightly higher X expression in females [4],[11].
XYY and XaXiXi triploid fibroblasts with a single Xa had average X:A expression ratios of 0.81±0.02 and 0.84±0.04, respectively (Figure 1 and Table S1). These X:A expression ratios were significantly above the theoretical value of 0.67 (2/3) (p<0.0005, n = 550, independent one-sample t-test) expected if X expression was strictly doubled while autosomal genes presumably produced 1.5-fold more products in triploid cells, which was confirmed as described below. For this statistical analysis, the expression signal of each of 550 X-linked probe-sets was divided by the mean of autosomal expression to calculate the X:A expression ratio for each X-linked gene in each triploid genotype. For each genotype an independent one-sample t-test was performed to test and reject the null hypothesis that the mean X:A expression ratio was not significantly different from the expected mean (see Materials and Methods). Even when considering that the mean X:A expression ratio was slightly higher than 1 (1.04–1.14) in control diploid cultures the ratio in triploid cells with one Xa remained significantly higher than expected (p = 0.001 and 0.03, respectively, n = 550). Thus, X expression was more than doubled in triploid cultures with one Xa, but it was not tripled since the X:A expression ratios were significantly below the value of 1 expected if expression strictly responded to an autosomal counting factor (p<0.001, n = 550). The average X:A expression ratio in triploid cultures with two active X (XaXaXi) was 1.36±0.04, which was similar to the theoretical value of 1.33 (4/3) expected based on a simple doubling of X expression on active X allele in these cells. In contrast to triploid cells with one Xa, there was no evidence of higher than doubled X expression in cells with two Xa's, i.e. the observed ratio of 1.36 was significantly lower than 1.68 (twice the ratio observed in XaXiXi cells, p = 0.00007, n = 550; Figure 1 and Table S1). We conclude that, on average, expression of X-linked genes is not tripled in triploid cultures; rather, the X chromosome transcriptional output appears to remain basically doubled like in diploid cells, with a further adjustment in triploid cells with a single active X.
Expression array data must be normalized to genome or autosomal expression in order to be comparable between experiments. To address the possibility that the presence of three sets of autosomes might lead to an adjustment in autosomal rather than X-linked gene expression, we verified that absolute expression levels per gene copy were the same for autosomal genes in triploid and diploid cell cultures by quantifying transcripts per copy number of eight autosomal housekeeping genes. Total nucleic acids (XNA) were prepared from triploid and control diploid cultures under conditions that avoided distortion of the relative ratio of mRNA to DNA [12]. The average fold change in absolute expression levels per gene copy determined by quantitative PCR of the autosomal genes was 0.90±0.35 between triploid and diploid cultures, indicating no adjustment in autosomal expression levels per gene copy relative to ploidy (Figure S2). However, individual gene expression was variable and thus we turned to a global approach to measure absolute gene expression.
To compare global absolute expression levels per gene copy NimbleGen DNA tiling arrays were hybridized with Cy3-labeled gDNA and Cy5-labeled cDNA from triploid or diploid XNA. To obtain sufficient materials XNA was prepared from an uncloned XXX triploid culture in which the proportion of cells with one Xa or two Xa's was determined by RNA- and DNA-FISH using a probe for XIST to mark the inactive X (Figure S3). The X:A expression ratio calculated from tiling array data was 0.98 (compared to 0.99 for the diploid culture), as predicted based on a mixture of 57% XaXiXi (ratio = 0.84) and 38% XaXaXi cells (ratio = 1.36) (Figure S3). As expected, probes with a high cDNA/gDNA signal ratio after hybridization to the tiling array were associated with exons, especially at the 3′ end of genes (Figure 2A). Scatter plots of absolute expression values in log2 scale for 5,473 autosomal genes showed similar profiles between triploid and diploid cultures, with a high R-square value (0.90) (Figure 2B); this concordance was confirmed by calculating an overall 1.07 fold change between average absolute expression levels per gene copy of 5,473 autosomal genes in triploid and diploid cultures. Although the distributions of absolute expression levels in log2 scale were similar between cultures, they were unexpectedly bimodal. We investigated them by separating genes into two groups: those with a level <1 (group 1: 1,682 genes) or ≥1 (group 2: 3,478 genes) (Figure 2C). Within each group, the absolute autosomal expression levels had a normal distribution and the fold change between triploid and diploid cultures remained 1.07 (Figure S4A, S4B). Increases in total exon length, length of the last exon, and number of exons for genes in group 2 versus group 1 probably contributed to a shift to higher signals and an apparently bimodal distribution (Figure S4C, S4D, S4E). We conclude that the absolute level of autosomal gene expression, i.e., expression level per gene copy, does not apparently differ between triploid and diploid cells. This suggests that expression adjustment of the X chromosome but not autosomes prevails in triploid cell cultures.
Based on expression array data (after normalization to autosomal expression) the distributions of expression of X-linked genes differed between triploid and diploid cultures (Figure 3). Compared to diploid cultures, triploid cultures with one Xa showed a clear shift towards lower expression values for X-linked genes (single factor ANOVA test p = 0.001 and 0.0003 for XaXiXi and XYY, respectively; Figure 3A). In contrast, triploid cultures with two Xa's showed a shift towards higher expression values for X-linked genes (single factor ANOVA test p = 0.0002). Comparisons of individual gene expression levels (normalized to autosomal expression) between triploid cultures with one or two Xa's using scatter plots showed that most X-linked genes (307/362 or 85%; spots above a central diagonal in Figure 4A) had higher expression in cultures with two Xa's versus one Xa. A subset of X-linked genes (38%; red spots in Figure 4A) had an increase of 2-fold or higher, suggesting that their expression was proportional to the number of Xa's in triploid cultures. However, the rest of genes (62%; Figure 4A) did not show this two-fold increase in cultures with two Xa's versus one Xa, consistent with lower expression from each Xa in XaXaXi cultures compared to expression from the single Xa in XaXiXi cultures.
Scatter plots of X-linked gene expression levels (normalized to autosomal expression) to compare sex-matched triploid to diploid cultures showed a consistent proportion (69–73%) of X-linked genes with higher cumulative expression relative to autosomes in all three XaXaXi cultures compared to XaXi cultures (Figure 4C and Figure S5A, S5B, S5C). The proportion of X-linked genes with lower expression relative to autosomes in all three XaXiXi cultures versus XaXi cultures and all three XYY cultures versus XY cultures was 71–81% and 68–71%, respectively (Figure 4D and 4E, Figure S5D, S5E, S5F, S5G, S5H, S5I). In both of these sex-matched comparisons a subset of genes (30% and 24%, respectively) was identified as having similar expression levels relative to autosomes in triploid and diploid cells (<1.2-fold change, p>0.2, two-tail unpaired student t-test; Figure 5A and 5B, Table 1, and Table S2). This subset of genes, whose expression was nearly tripled in triploid cultures with a single Xa, could represent genes for which a balanced expression was critical. Twenty-six such genes were identified in both female and male triploid cultures (black spots in Figure 5A and 5B and Table 1 and Table S2). These genes consistently showed similar expression levels relative to autosomes, with average fold changes close to 1 between triploid cultures and between triploid and diploid cultures, suggesting that they are regulated in a dosage-sensitive manner (Table 1). Their map position revealed a slightly more prevalent location on the short arm of the human X chromosome (2.2 genes/10 Mb) compared to the long arm (1.4 genes/10 Mb) (Table S2). GO analysis showed no specific functional classification [13].
Specific categories of X-linked genes such as genes located within the pseudoautosomal region (PAR) are not subject to dosage compensation mechanisms and thus were used as controls in our study. Analyses of nine PAR1 genes showed similar expression levels between triploid cultures. The average fold changes were 1.02 (XaXaXi/XaXiXi), 1.08 (XaXaXi/XYY), and 1.07 (XaXiXi/XYY), indicating that PAR1 genes behave like autosomal genes, independent of the sex complement and of the number of Xa's (Figure 6A and 6B). This supports the notion that PAR1 genes escape from X inactivation [14], and are not up-regulated on the active X (Nguyen and Disteche, unpublished data).
Expression of non-pseudoautosomal X/Y gene pairs, in which the X paralogs escaped X inactivation [15],[16], was also examined to determine whether Y expression and/or escape from X inactivation could help balance expression in triploid cultures (Figure 6C). For 8/10 pairs (UTX/UTY, CXorf15/CYorf15A & B, EIF1AX/EIF1AY, ZFX/ZFY, DDX3X/DDX3Y, RPS4X/RPS4Y1, USP9X/USP9Y, JARID1C/JARID1D) significant Y expression was detected that was not due to cross-hybridization with the X paralog (Table S1). Cumulative expression of most X and Y paralogs in XYY cultures was not significantly different from total X expression in XaXiXi cultures (Figure 6C), suggesting that Y expression could potentially help compensate, would the paralogs provide similar functions as is the case for some gene pairs [17], but not others [18],[19]. An exceptional Y-linked gene was JARID1D, which had very high expression compared to JARID1C. For most X paralogs, cumulative expression was generally lower in XaXiXi versus XaXaXi cultures, consistent with lower expression from alleles on the Xi (Figure 6C) [14].
To examine the effects of escape from X inactivation on the expression of genes in triploid cultures non-pseudoautosomal X-linked genes (with or without a Y paralog) were grouped into three categories based on a previous survey [14]: (1) genes subject to X inactivation (not expressed in rodent x human hybrid cells that retain an inactive human X, group 4 in Table 1), (2) genes with variable escape (expressed in 1-6/9 hybrid lines, group 5 in Table 1), and (3) genes that consistently escape (expressed in 7-9/9 hybrid lines, group 6 in Table 1) [14]. Predictably, the expression fold change between XaXaXi and XaXiXi cultures was the lowest for genes that escape X inactivation as expected for genes whose expression depends on the total number of X chromosomes (Figure 5C and Table 1). However, this fold change was higher than 1, consistent with lower expression from alleles on the Xi. Genes that consistently escape X inactivation had higher expression in XaXiXi versus XYY cultures compared to genes subject to X inactivation, (Figure 5D and Table 1). We conclude that XYY triploid cultures have a deficiency in expression of escape genes compared to XXX triploid cultures with one or two active X chromosomes.
Our results indicate that X-linked gene expression is up-regulated approximately two-fold in triploid cultures with two active X chromosomes like in diploid cultures [3],[4]. However, in triploid cultures with a single active X, X expression is apparently adjusted further upward, although not tripled. Thus, global X-linked gene expression does not strictly respond to an autosomal counting factor in triploid cells. However, comparisons between triploid and diploid cultures showed a range of individual gene expression, suggesting that not all genes respond to a change in ploidy in the same manner. We previously proposed two mechanisms for X up-regulation [4]: (1) permanent DNA sequence changes during evolution that may affect promoters, enhancers, and/or regions that alter mRNA elongation and stability; (2) epigenetic mechanisms that may control X expression similar to the situation in Drosophila [1]. A basic doubling of X expression in triploid cells with two Xa's could be mediated by DNA-sequence modifications and/or by epigenetic mechanisms that normally operate in diploid cells. The notion that the doubling cannot default even when X expression is too high in cells with two active X chromosomes strongly suggests that DNA sequence changes or highly stable epigenetic modifications are involved in X up-regulation. We found no evidence of a global change in absolute expression level per autosomal gene copy in triploid versus diploid cultures, suggesting that gene expression is strictly proportional to the number of autosomal sets. Hence, adjustments in gene expression in triploid cultures mostly affect X-linked genes.
In triploid cells with one Xa, X expression is adjusted upward from a basic doubling. This further adjustment is presumably mediated by an epigenetic mechanism that senses X expression but does not respond to the simple X:A genomic ratio since this occurs both in XYY and XaXiXi cultures. Our data suggest that such a mechanism may act on a subset of genes (∼7% of X-linked genes) whose expression dosage in relation to autosomal expression may be critical. Preliminary investigation in their function did not reveal a specific role for these genes. It is also possible that the triploid cultures with one Xa we examined survived because X expression was adjusted upward, thus reflecting a bias of ascertainment. Interestingly, a previous study in Drosophila showed that the normalized enzymatic activity of a protein encoded by an X-linked gene was 0.87 in triploid flies with a single X relative to diploid flies, which suggests a more than doubled, but not tripled expression, similar to our results in human [20].
Previous studies have suggested that fold-changes in expression are often considerably less than differences in gene copy number [3]. It should be noted that we measured steady-state RNA and thus cannot rule out adjustments in protein levels at translation. We previously reported low X:A expression ratios (0.84–0.89) in a subset of normal human diploid tissues [4], suggesting that ratios within this low range could be tolerated in a tissue-specific manner. Conversely, undifferentiated mouse female ES cells with two Xa's have an X:A expression ratio of 1.39 [21], indicating that such a high ratio is tolerated in normal diploid cells during a short time window prior to X inactivation in blastocysts. However, later stage mouse embryos with two Xa's due to failure of X inactivation have a severely abnormal phenotype similar to that of embryos with tetrasomy for an autosome [22]. A previous study of female mouse ES cells with two active X chromosomes showed genome-wide DNA methylation changes in repeated DNA sequences, perhaps in response to the high X expression [23]. However, in our study we did not detect an adjustment in global autosomal expression in triploid cells.
Triploidy, which predominantly results from dispermy, causes lethality in human embryos due to multiple phenotypic abnormalities [8]. The unbalanced X:A expression ratios that we observed could contribute to this lethality. XYY fetuses are much rarer than XXY or XXX fetuses, in which the majority of cells often have two Xa's [5],[24]. We found that XYY cultures had a deficiency in expression of genes that escape X inactivation compared to XaXiXi cultures, which could contribute to the early lethality of XYY triploid fetuses. Relevant to this observation is the prevalent lethality in Turner syndrome fetuses with a single X chromosome, associated with haplo-insufficiency of genes that escape X inactivation. In XYY triploid fetus, Y paralog expression could partially compensate, provided that the function of the X and Y paralogs are similar.
Deficiency in gene expression due to chromosomal abnormalities such as monosomy is less well tolerated than trisomy. Thus, lower than normal X expression may be less compatible with embryonic development than higher expression. A recent study of tetraploid mouse ES cells with a different number of Xa's also supports the notion that an excess of Xa's is more favorable to cell survival than a deficiency [25]. Live-born triploid fetuses have a mixture of cells with one or two Xa's, while early aborted fetuses tend to have a majority of cells with two Xa's [5],[24]. Interestingly, the uncloned culture we studied that consisted of a mixture of cells with one or two Xa's had a balanced overall X:A expression ratio close to 1.0. Thus, mosaicism for cells with different X inactivation patterns may help achieve a more balanced proteome, excluding cell-autonomous protein products. Mechanisms involved in X up-regulation in mammals will need to be investigated at early embryonic stages to follow changes in the level of X-linked gene expression during normal development.
The triploid fibroblast cultures consisted of six XXX clones with previously defined X inactivation patterns (75-29-H2, GM04939-2X-1, GM04939-2X-2, 75-29-E4, 75-29-F3 and 75-29-F9) [5], two XXX uncloned cultures (69,XXXCD2 and 69,XXXCD3), three XYY cultures (69,XYY1, 69,XYY2, 69,XYY4), two male and two female diploid cultures, all derived from aborted fetuses and previously characterized by cytogenetic analysis (Table S1). For 69,XXXCD2 we carried out X enumeration using FISH with CEPX alpha satellite and X inactivation analysis using RNA FISH for XIST (X inactive-specific transcript) using probes from Vysis (Downers Grove, IL) and standard protocols. The study was approved by the IRB at the University of Washington and at the Mayo Clinic College of Medicine.
Total RNA was prepared using RNeasy kits (Qiagen) with on-column DNaseI digestion prior to quality assay, labeling, and hybridization to Affymetrix human HG-U133 2.0 plus chip (Santa Clara, CA). Array hybridizations were done at the Microarray Center (University of Washington, Seattle WA). The raw data files (DAT file) were analyzed by the Affymetrix software (GCOS 1.4) to produce the data in .CHP format (Excel). Expressed “probe-sets” were selected as those showing signals above background levels, which were determined by the mean signal value of Y-linked probe-sets identified as absent using Affymetrix MAS Present/Absent calls based on a PM-MM (Perfect match-Mismatch) mode (Table S1). The mean X:A expression ratios calculated for 16,984 expressed autosomal probe-sets (corresponding to 10,267 autosomal genes) and 550 X-linked probe-sets (corresponding to 362 X-linked genes) were similar to ratios obtained using a method previously described (Table S1) [4]. For statistical analysis the signal value of each X-linked probe-set was divided by the autosomal array mean and averaged from three arrays for each triploid genotype to calculate one X:A expression ratio for each X-linked probe-set for each triploid genotype. Independent one-sample t-test (Hypothesis test) was then used to compare the mean of X:A expression ratios in one genotype to expected values. The X:A expression ratios (log2 scale) had a normal distribution and thus were used to calculate z scores and corresponding p-values (z2 = (log2M–log2E)/(s2/n) (s: standard deviation; n: number of probe-sets) [21]. Gene expression values were normalized to the median value of expressed autosomal probe-sets for each array to compare profiles (distribution and scatter plots) between cell types. Single factor ANOVA test was used to compare distribution profiles. Expression of individual genes was calculated as the mean value of all normalized probe-sets for a given gene. Two-tail unpaired student t-test was used to evaluate the significance of expression changes of individual genes between different triploid cultures and between triploid and diploid cultures. The microarray data are available in NCBI's Gene Expression Omnibus under GEO Series accession number GSE18877 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE18877).
Total nucleic acids (XNA, i.e. DNA and RNA) were isolated [12]. Two µg of XNA was reverse-transcribed using random hexamers and SuperScript II reverse transcriptase (Invitrogen). The intergenic region ∼3.5kb upstream of ACTG1 (actin, gamma 1) was used as the internal reference for gene copy normalization. cDNA-specific primers designed to span large intron(s) of eight autosomal housekeeping genes abundantly and similarly expressed in triploid and diploid cultures were tested for PCR efficiency and R2 (square of correlation coefficient) using standard curves (Table S3). After quantitative PCR absolute gene expression was calculated as the ratio between cDNA signals and genomic DNA signals.
Three µg of XNA from diploid (46;XX4) or triploid cells (69;XXXCD2) was fragmented by sonication (100–600bp), digested by RNase A (Promega), and purified (Qiagen PCR purification kit). Genomic DNA was Cy3- labeled using a standard NimbleGen (Madison, WI) sample labeling protocol. 25 µg of XNA from the same preparation was completely digested by DNaseI (Qiagen) and treated with an RNeasy kit (Qiagen). cDNA was produced using oligo-dT and a Superscript double-strand cDNA Synthesis kit (Invitrogen). The cDNA was Cy5-labeled and co-hybridized with Cy3-labeled gDNA to NimbleGen human HG18 economy tiling array set4 that contains 2.1M probes (50–75 mer; median probe spacing: 205 bp) and covers chromosomes 14–22, X and Y.
Array hybridization and scanning were done at the Genome Resources Center (Fred Hutchinson Cancer Research Center, Seattle, WA). The raw data files (pair file) were processed by NimbleScan software using a CGH algorithm without any normalization between two channels. The probes were mapped to exons according NimbleGen annotation files based on build HG18 (UCSC Genome browser) [26]. The absolute expression level for a given gene was calculated by averaging cDNA/gDNA ratios of five exon-associated probes in the 3′ end of that gene.
|
10.1371/journal.pntd.0005166 | A Phosphorylcholine-Containing Glycolipid-like Antigen Present on the Surface of Infective Stage Larvae of Ascaris spp. Is a Major Antibody Target in Infected Pigs and Humans | The pig parasite Ascaris suum plays and important role in veterinary medicine and represents a suitable model for A. lumbricoides, which infects over 800 million people. In pigs, continued exposure to Ascaris induces immunity at the level of the gut, protecting the host against migrating larvae. The objective of this study was to identify and characterize parasite antigens targeted by this local immune response that may be crucial for parasite invasion and establishment and to evaluate their protective and diagnostic potential.
Pigs were immunized by trickle infection for 30 weeks, challenged with 2,000 eggs at week 32 and euthanized two weeks after challenge. At necropsy, there was a 100% reduction in worms recovered from the intestine and a 97.2% reduction in liver white spots in comparison with challenged non-immune control animals. Antibodies purified from the intestinal mucus or from the supernatant of cultured antibody secreting cells from mesenteric lymph nodes of immune pigs were used to probe L3 extracts to identify antibody targets. This resulted in the recognition of a 12kDa antigen (As12) that is actively shed from infective Ascaris L3. As12 was characterized as a phosphorylcholine-containing glycolipid-like antigen that is highly resistant to different enzymatic and chemical treatments. Vaccinating pigs with an As12 fraction did not induce protective immunity to challenge infection. However, serological analysis using sera or plasma from experimentally infected pigs or naturally infected humans demonstrated that the As12 ELISA was able to detect long-term exposure to Ascaris with a high diagnostic sensitivity (98.4% and 92%, respectively) and specificity (95.5% and 90.0%) in pigs and humans, respectively.
These findings show the presence of a highly stage specific, glycolipid-like component (As12) that is actively secreted by infectious Ascaris larvae and which acts as a major antibody target in infected humans and pigs.
| Roundworms infect millions of humans and pigs throughout the world. The pig roundworm A. suum is a good model for A. lumbricoides infection in humans due to similar host physiology and the close genetic relationship between the worms. The aim of this study was to identify and characterize early larval antigens that are targeted by antibodies at the level of the intestine in immune pigs and to evaluate their protective and diagnostic potential. In order to do so, we generated highly immune pigs by repeatedly infecting them with A. suum for a long time (32 weeks). After necropsy, locally harvested antibodies from the gut were used to screen larval extracts. Hereby one particular antigen, named As12, was detected. It was characterized as a molecule of glycolipid nature that is presented on, and actively secreted from, the surface of infective larvae. Pigs immunized with this antigen are not protected from subsequent challenge infection. Experimentally infected pigs or naturally infected humans do however mount a significant serological antibody response to the antigen. These findings shed light on a glycolipid-like antigen (As12) that is secreted by infectious Ascaris larvae and is targeted by the immune system of infected humans and pigs.
| Ascaris lumbricoides is the most prevalent intestinal parasitic nematode of man, infecting approximately 819 million people worldwide in developing countries [1]. Due to the high degree of morphological and genetic similarity, it is still debated as to whether A. lumbricoides from humans is a different species than A. suum from pigs [2–4]. Moreover, recent studies have shown that pig Ascaris is a zoonosis [5–8]. Even though anthelmintic treatment remains highly effective against A. lumbricoides, there is increased concern about the development of anthelmintic resistance. In addition, the high risk of reinfections after treatment calls for the development of new, long-acting solutions like vaccination. In addition, the development of more rapid and sensitive diagnostic techniques that adequately reflect the level of Ascaris exposure in a population could greatly improve our knowledge on infection dynamics and prevalence. Consequently, it would thus allow for a more precise estimate of the impact of infection and a better evaluation of a given intervention.
Vaccination has proven to be the most efficient and cost-effective way of disease control [9]. Vaccination against ascariasis should in theory be feasible since pigs, repeatedly infected with A. suum, develop immunological responses at the level of the liver, lungs and intestine that stop migrating larvae from reaching adulthood. Furthermore, repeated exposure to the parasite induces an immunological response at the level of the intestine that is called the 'pre-hepatic barrier', eventually preventing newly acquired larvae from migrating to the liver [10, 11]. Recently, it has been shown that this immunity was associated with eosinophilia, mastocytosis and goblet cell hyperplasia in the caecum, the place where the infective stage 3 larvae (L3) penetrate the intestine and start their hepatopulmonary migration [12]. However, it is still unclear which parasite products induce these immune responses or what the targets of these responses are. An increased understanding on this matter could offer important information for the development of protection by vaccination against this parasite.
The need for improved methods to diagnose Ascaris infections in pigs and humans has recently been extensively discussed [13, 14]. It was suggested that diagnostic tools detecting eggs in the stool are not useful for accurate evaluation of the level of exposure in pig farms [15] or sensitive enough for the detection of infection in humans where prevalence was low [16]. Serological tools detecting exposure to Ascaris might be more sensitive than egg based diagnostics for measuring prevalence or intensity of exposure in a human community [17]. Until now, only a handful of studies report the evaluation of antibody-based tests for ascariasis [18–23]. Recently, Vlaminck et al., [15, 17] showed that an ELISA detecting antibodies to Ascaris haemoglobin in plasma or serum samples appears to reflect general exposure to Ascaris on a community or herd level in humans and pigs, respectively. However, more species-specific antigens from early larval stages might increase the sensitivity and specificity of serological assays or recognize infections at an earlier stage.
Hence, the main objective of this study was to use intestinal antibodies from pigs with a proven pre-hepatic barrier to identify immunogenic proteins of the infective stage larvae of A. suum and subsequently evaluate their protective and diagnostic potential.
The piglets used in this study were female and castrated male Rattlerow Seghers hybrid pigs of the local stock of the animal facility (Ghent University). They were approximately 10 weeks old and weighed between 20 and 30 kg at the start of the trials. The pigs were raised indoors in a helminth free environment and had free access to a commercial feed and water.
Adult female A. suum worms were collected from the intestines of naturally infected pigs from commercial farms that were being processed as part of the normal work at a local abattoir in Ghent, Belgium. Consent was acquired from abattoir management to collect the worms. Adult female A. lumbricoides worms were collected from the stool of individuals after treatment with anthelmintics in Jimma Town, Ethiopia [24].
Ascaris eggs were obtained by dissection of worm uteri and suspended in a 0.5% (w/v) potassium dichromate solution to a volume of 50 ml and placed in a culture flask at a concentration of 50 eggs/μl. The eggs were incubated at 27°C in the dark until third stage larvae were present and were subsequently used for infection trials or to extract infective stage-three larvae (L3).
Fresh L3 were obtained from Ascaris eggs using the method described by Urban et al. [25]. In brief, Ascaris eggs, cultured in vitro were treated for 1 hr with commercial bleach and subsequently washed 3 times with phosphate buffered saline (PBS) of 37°C after which the eggs were transferred into an Erlenmeyer flask containing glass beads and a magnetic stir bar and stirred very slowly (60 rpm) to induce hatching. After 15 min the suspension was poured onto a layer of cotton wool placed on top a Baermann apparatus with PBS at 37°C and left overnight. The next day, the larvae in the neck of the funnel were collected and washed 3 times in PBS. L3 extracts were prepared by grinding the collected larvae with a mortar and pestle that was placed in a bath of liquid nitrogen. The larval homogenate was transferred to a 15 ml tube and mixed with PBS and proteinase inhibitor cocktail (1:100) (Sigma, Diegem, Belgium). The homogenate was then inverted at 4°C for 2 hrs followed by centrifugation for 30 min at 10,000g at 4°C. The supernatant (L3 PBS) was removed and kept on ice. The pellet was resuspended in PBS with 0.05% Tween-20 solution and fresh proteinase inhibitor cocktail (1:100) was added. The mixture was inverted at 4°C for 2 hrs and the supernatant (L3 PBST) was removed after centrifugation and stored on ice. Finally, the remaining pellet was resuspended in PBS containing Triton X-100 (2%) and proteinase inhibitor cocktail (1:100) and subsequently inverted at 4°C for 2 hrs followed by centrifugation. The supernatant (L3 Triton) was collected and stored on ice. Subsequently, all extracts were sterilised by filtration (0.22μm) and the filtrate concentrated at 4°C using a Centriprep centrifugal filter with YM-3 membranes (Millipore, Overijse, Belgium). Protein concentration was determined by the BCA method (Pierce, Rockford, USA) and the extracts were stored at -80°C until use.
In order to obtain L3 excretory-secretory (E/S) products, freshly obtained A. suum L3 were incubated at 37°C and 5% CO2 at a concentration of 5,000 larvae/ml in DMEM with 4,5 g/L Glucose, L-Glutamine and Pyruvate (Thermo Fisher, Erembodegem, Belgium) containing 1% Penicillin-Streptomycin (P/S) (5,000 u/ml Penicillin and 5,000 μg/ml Streptomycin, Thermo Fisher), 1% Kanamycin (10000 μg/ml, Thermo Fisher), 1% Amphotericin B (250 μg/ml, Sigma) and 0,5% Gentamicin (10 mg/ml, Thermo Fisher). The culture fluid was collected daily and filtered using 0.2 μm membrane disc filters (Supor 200, Zaventem, Belgium), concentrated and dialysed against PBS in an Ultrafiltration Stirred Cell (Millipore) using a 10kDa cut-off filter membrane (Millipore). The L3 E/S material was then stored in aliquots at -80°C.
In order to obtain lung L3 and intestinal L4 and L5 stages, piglets were infected with approximately 100,000 infective A. suum eggs and sacrificed 7, 14 and 28 days post-infection, respectively. The lung L3 and intestinal L4 were collected from minced lung tissue or intestinal content that was placed on a modified Baermann device as described by Slotved et al. [26]. L5 larvae were collected from the intestinal content by hand. Larvae were washed excessively using PBS 4°C and stored at -80°C. Adult worms were collected from the intestines of infected pigs from commercial farms that were being processed as part of the normal work at a local abattoir in Ghent, Belgium. Protein extracts from all life stages were produced as described above for the L3 stage. Finally, extracts from A. lumbricoides L3, were obtained after sonication as previously described by Vlaminck et al., [27].
Eight pigs were divided into 2 groups of 4 pigs. Pigs of group B were trickle infected 5 times a week with approximately 100 infective A. suum eggs in the feed for a period of 30 weeks. Pigs of group A were used as challenge controls. Two other pigs were euthanized before the start of the infection trial and used as negative controls. After the 30-week infection period, all pigs of group A and B were treated with a single dose of 5mg/kg fenbendazole (MSD, Brussels, Belgium). One week after treatment, all pigs were infected with 2,000 infective A. suum eggs. All pigs were euthanized 14 days post challenge infection. During necropsy, the number of white spots, characteristic lesions on the liver caused by migrating A. suum larvae, was recorded and mesenteric lymph nodes from the small intestine, caecum and colon were collected and immediately processed as described below. Pieces of small intestine of approximately 1 meter were flushed three times with 50ml of PBS to rinse out all larvae in the intestinal lumen. The rinsing solution was collected and passed over a 200 μm mesh sieve. The remaining debris on top of the sieve was collected and examined for intestinal L4 larvae. After rinsing, the pieces of small intestine and the caecum and colon were cut open longitudinally and washed gently with excessive volumes of lukewarm tap water to remove any remaining intestinal content. Subsequently, mucus was collected by gently scraping the luminal side of the pieces of intestine with a microscope slide. During the experiment, blood samples for serum were collected from all pigs every 2 weeks.
Single-cell suspensions of mesenteric lymph node cells (LNC) were prepared by mechanical disaggregation through a sterile stainless steel gauze. The LNC in ice cold PBS were centrifuged at 150 g for 10 min at 4°C, the supernatant removed and LNC resuspended in DMEM growth medium supplemented with 4,5 g/L Glucose, L-Glutamine and Pyruvate (Thermo Fisher) containing 1% P/S (Thermo Fisher). The mononuclear cells (monocytes and lymphocytes (MNC)) were separated by the addition of LymphoPrep and centrifugation at 800g (without brake) for 30 min at 4°C. The MNC were collected from the interphase and washed twice with DMEM + 1% P/S + 2% heat-inactivated foetal bovine serum (Moregate, Australia & New Zealand) and the cells were subsequently suspended in DMEMcomplete. (DMEM + 1% P/S + 1% Non-Essential Amino Acids (Thermo Fisher), 1% Kanamycin (10 mg/ml, Thermo Fisher), 1% Amphotericin B (250 μg/ml, Sigma) and 0.1% of a 0.35% β-mercaptoethanol solution). Viable MNC were counted, the concentration adjusted to 5.0 106 cells/ml in DMEMcomplete and cultured for 4 days in tissue culture flasks at 38°C and 5% C02. Finally, the culture supernatant (SN) was collected, filtered using 0.2 μm membrane disc filters (Supor 200), concentrated using Centriprep centrifugal filter units with YM-3 membranes (Millipore) and stored in aliquots at -80°C.
An equal volume of ice-cold PBS was added to the mucus collected from small intestine, caecum and colon and subsequently homogenised using an Ultrathurax mixer (2 min at 15,000 rpm). This mixture was then centrifuged for 15 min at 10,000g at 4°C, the supernatant was collected and centrifuged again as before. After this second centrifugation step, the supernatant was stored at -80°C. At a later time, antibodies were purified from this supernatant using Protein-A agarose beads (Sigma) following the manufacturer’s protocol. Purified mucosal antibodies were stored at -20°C until used.
Protein extracts (5 μg) were mixed with 5x sample buffer (60 mM Tris-Cl pH6.8, 2% SDS, 10% glycerol, 5% β-mercaptoethanol, 0.01% bromophenol blue) and put into a boiling water bath for 5 min. Afterwards, the samples were applied to 15% SDS-PAGE gels and separated by electrophoresis in Tris-Glycine buffer (Tris 250mM, Glycine 200mM, SDS 1% w/v). Protein bands were visualized using SimplyBlue Safestain (Thermo Fisher) or SilverStain kit (Thermo Fisher). Glycoproteins were stained by the use of the ProQ Emerald 300 gel stain kit (Thermo Fisher).
For Western blotting, SDS-PAGE gels were blot transferred to PVDF membranes (Millipore) or nitrocellulose membranes (Thermo Fisher) and blocked in PBS + 0.2% Tween80 (PBSt80) or in PBS + 0.2% Tween20 (PBSt20) + 5% Blotting-Grade Blocker (BioRad). The blots were probed for 2 hrs (1ml/lane PBSt80 containing approximately 5 μg/ml antibodies purified from mucus or the concentrated culture supernatant of MNCs). The following conjugates and dilutions were used: goat anti-pig IgG-HRP conjugated (Sigma) (1/10,000), goat anti-pig IgA-HRP (Abcam, Cambridge, UK) (1/5,000). The immunoreactive antigens were visualised by chemiluminescent substrate (5ml 0.1M Tris pH 8.6 + 11μl 90mM p-coumaric acid (Sigma) + 25μl 250mM luminol (Sigma) + 15μl of 10% (v/v) H2O2).
L3 PBS extract was subjected to a Folch method [28] for the extraction of total lipids. In short, L3 PBS extract was mixed with chloroform/methanol (2/1) to a final volume 20 times the volume of the original extract and incubated for 20 min. The homogenate was centrifuged and the liquid phase recovered and washed with 0.2 volume of 0.9% NaCl solution. The mixture was centrifuged at 2,000 rpm to separate the two phases. Both upper and lower phases were collected separately and evaporated at 90°C under a fume hood. The dried lipid pellets were stored at -80°C until further use.
To investigate the composition of the As12 antigen, total L3 PBS extract was incubated with different enzymes and submitted to different chemical treatments. For the enzymatic degradation, the L3 PBS extract was treated overnight at 37°C with pronase, lipase and trypsin at pH 8.0 and pepsin at pH 3.0 (all from Sigma). Additionally, L3 PBS extract was treated overnight at 37°C or 60°C in 20mM periodic acid, 1M NaOH, 1M Trifluoracetic acid and 1M HCl or in PBS at 37°C, 60°C or 90°C. Purified As12 was treated in 48% aqueous HF at 4°C for two days, then lyophilized and used for Western blotting. The presence of phosphorylcholine (PC) on the antigen was confirmed by screening with TEPC-15 monoclonal IgA antibodies (Sigma) on western blot. The deglycosylation of The L3 PBS extract with PNGase F was performed under denaturing conditions according to the manufacturers’s protocol (Promega). RNAse B was used for treatment control. The As12 antigen was also incubated in 50mM NaPO4, pH 5.0 for 2 days at 37°C with 0.1 unit ß-N-acetylglucosaminidase from jack bean (Sigma). After incubation, samples were boiled in 5x sample buffer, run on SDS-page, blotted onto PVDF membranes (Millipore) and recognition of the As12 antigen checked by incubation with mucosal antibodies from pigs of group B with 5μg/ml of antibodies/lane.
For glycan analysis, aliquots of As12 were treated with PNGase-F, -A and chemical β-elimination to release N- and O-glycans, respectively. Released glycans were analysed by MALDI-TOF-MS after derivatisation with 2-aminobenzoic acid or permethylation, as described previously [29, 30].
Approximately 2 μg As12 was hydrolysed in a glass vial with 50 μl 4 M TFA at 100°C for 4 hrs, dried under nitrogen, then monosaccharides labelled with 10 μl 2-aminobenzoic acid (2-AA) labelling mix (48 mg/ml 2-AA, 1 M 2-picoline-borane dissolved in 30% acetic acid / DMSO) followed by 2 hrs incubation at 65°C. For HPLC, 10 μl labelled sugars was added to 90 μl 0.6% sodium acetate, and 25 μl applied to a Superspher 100 RP-18 column 250 x 4 mm (Merck). Buffers were: A = 0.1% butylamine, 0.5% phosphoric acid, 1% tetrahydrofuran; B = 0.05% butylamine, 0.25% phosphoric acid, 0.5% tetrahydrofuran, 50% acetonitrile. Run conditions were 0.5 ml/min, starting at 8% buffer B for 5 min, 8–25% buffer B over 25 min, 25–100% B over 2 min, maintained for 10 min. Monosaccharide standards (500 pmol Glc, Gal, Man, Fuc, Xyl, GlcNAc, GalNAc) were treated with TFA, labelled, and ran as above.
A total of 40,000 L3 were incubated for 1 hr at 37°C in 1ml RPMI (Thermo Fisher) with 20μg purified mucosal antibodies from pigs from group A, group B and from negative controls or with TEPC-15 antibodies (Sigma) at a concentration of 1/500. After the incubation, larvae were washed three times with 1ml RPMI 37°C before being resuspended for 30 min in 1ml RPMI at 37°C containing FITC labelled anti-pig IgG (Bethyl laboratories, Montgomery, TX, USA) at a concentration of 1/1,000 or FITC labelled anti Mouse IgA at a concentration of 1/2,000 (Thermo Fisher). All incubations with FITC-conjugated antibodies were conducted in the dark. Finally, the larvae were washed four times with 1ml of PBS 37°C to remove any remaining secondary antibodies and put on a glass slide for fluorescent microscopy analysis. All labelling experiments were performed in triplicate.
To determine whether the stained L3 actively shed the As12 antigen, larvae were stained with 20μg of purified mucosal antibodies from pigs from group B as described above. After staining and washing, half of the larvae were killed by freezing them for 10 min at -80°C. The other half was kept in RPMI at 37°C. After this, the number of stained larvae in 3 aliquots of both groups was counted by fluorescence microscopy. This assessment was repeated after 1, 2, 3, 4 and 24 hrs.
Twelve piglets (approximately 10 weeks old) were divided into two groups of 6 pigs. Each pig in Group A was vaccinated with total lipid fraction purified from L3 PBS extract that was obtained from 200,000 L3 and dissolved in 0.5 ml sterile PBS + 0.5 ml Alhydrogel (AlOH). The control pigs of group B were injected with 0.5 ml sterile PBS + 0.5 ml AlOH. Pigs were immunized three times by intramuscular injection at day 0, 14 and 28 of the experiment. One week after the final immunization, all pigs were experimentally infected with 1,000 infective A. suum eggs in 5ml of tap water by oral intubation. Two weeks after infection, at day 49 of the experiment, all pigs were euthanized and L4 larvae were recovered from the small intestine as described above and counted. Blood samples were taken at the start of the trial, one week after the third immunization and at the time of necropsy to evaluate seroconversion against the As12 antigen following vaccination.
Indirect ELISA determined antibody recognition of the As12 antigen by pig sera or human plasma samples. ELISA plates were coated with antigen overnight in carbonate buffer (pH 9.6) at 4°C. Plates were coated with 1μl/ml of total lipid extract purified from 200,000 L3 which was dissolved in 50μl UPW. After three washes with PBSt, the plates were blocked with 100μl/well blocking buffer (5% milk powder (w/v) or 5% heat treated fetal calf serum in PBS) for 2hrs at 4°C. Sera or plasma samples were added in duplicate at a dilution of 1/250 in PBSt for 2hrs at 4°C. The plates were washed again as before and incubated with the conjugate (goat anti-pig IgG-HRP (Sigma) (1/10,000), goat anti-human IgG4-HRP (Southern biotech) (1/2,000)) in blocking buffer. Plates were incubated for 1hr at 37°C. O-phenylenediamine 0.1% in citrate buffer (pH 5.0) served as substrate and after a 10 min incubation period in the dark, the development reaction was stopped by adding 50μl of 4M H2SO4 to all wells and optical density (OD) was measured at 490 nm.
Sera from the trickle infected pigs described above were used to evaluate antibody response over time. Additional pig sera were obtained from an experimental infection trial performed by Nejsum et al., [31] where piglets 10 weeks of age were infected twice a week in the feed for a total of 14 weeks with A. suum and Trichuris suis (25 and 5 embryonated eggs kg−1 day−1, respectively). Serum and faecal samples were collected at the start of the trial (W0) and 7 (W7) and 14 (W14) weeks after the first infection. Pigs were euthanized 14 weeks after the start of the experiment and the number of macroscopic worms present in the small intestine counted. The negative control was a pooled serum sample from 4 10-week old piglets without previous exposure to A. suum. The positive control was a pooled serum sample from 4 pigs after 14 weeks of daily infection with 100 A. suum eggs [12]. Reactivity of pig sera to the antigen is shown in ODr (Optical Density ratio). (ODr sample = (OD sample−OD negative control) / (OD positive control−OD negative control)).
Plasma samples from A. lumbricoides infected humans were collected as previously described [32] from individuals living in Mainang village on Alor Island (Province of East Nusa Tenggara, Timor, Indonesia), an area with high Ascaris prevalence (>30%). A subset of 25 plasma samples of which all individuals had A. lumbricoides eggs in their stool was evaluated for anti-As12 IgG4 antibodies. A total of 24 plasma samples from individuals with hookworm infection used in this study were from people living in the East Sepik region of Papua New Guinea where A. lumbricoides was not present. Non-endemic control sera were from American subjects in St. Louis, MO, USA. Reactivity of human sera to the As12 antigen is shown in OD.
TEPC-15 antibodies (Sigma) or pooled mucosal antibodies of pigs from group B of the trickle infection experiment were dissolved in PBSt at a dilution of 1/250 or 10μg/ml respectively and subsequently pre-incubated with PC-Cl salt (Sigma) at different concentrations (0, 10, 25 μg/ml final concentration) for 30 min at room temperature. Following pre-incubation, the antibody mixtures were added to an As12-coated ELISA plate for 2hrs at 4°C. After washing the plates three times with PBSt, the TEPC-15 antibodies were detected by HRP labelled anti Mouse IgA at a concentration of 1/2,000 (Bethyl laboratories) and the mucosal antibodies were detected by HRP conjugated goat anti-pig IgG (Sigma) at a 1/5,000 dilution in blocking buffer for 1hr at 37°C. Finally, the plates were washed, developed and read as described above.
All statistical analyses were performed using Graphpad Prism 6.0e for MacOSx. The Mann-Whitney U-test for pairwise comparison was used to compare the group means of the immune pigs and the challenge controls during the vaccination trial or to compare anti As12 antibody levels between different groups of human or pig plasma or serum samples. T-tests were performed to test for differences in the percentage of stained larvae or antibody reactivity of the two different experimental groups during the vaccination experiment. A paired t-test was used to detect the first time point at which infected pigs showed a significant higher ELISA reactivity to As12 antigen compared to the start of infection. Possible correlations between antibody responses and parasitological data (EPG or worm counts) were assessed using the Spearman's rank correlation coefficient. The use of computed Receiver Operating Characteristic (ROC) curves allowed for the determination of test sensitivity and specificity and to select an appropriate cut-off. Probability (P) values < 0.05 were considered to indicate significant differences.
All animal experiments were conducted in accordance with the E.U. Animal Welfare Directives and VICH Guidelines for Good Clinical Practice. Ethical approval to conduct the studies was obtained from the Ethical Committee of the Faculty of Veterinary Medicine, Ghent University. The collection of adult A. lumbricoides worms was performed during a trial performed by Mekonnen et al., in 2013 [24]. This study was approved by the ethical committee of Jimma University, Ethiopia, Ghent University and Antwerp University, Belgium. The Ethical Committee of the University of Indonesia, Jakarta approved the Alor Island study as previously described [32]. The Human Investigations Institutional Review Boards of Case Western Reserve University and the Papua New Guinea Medical Research Advisory Committee approved all protocols. The Institutional Review Board at Washington University School of Medicine in St. Louis, MO, USA approved our use of anonymized patient samples for the development of serological tests for helminth infections. Since written consent is not consistent with cultural norms on Alor Island, oral informed consent was obtained from all adults or, in the case of children, from their parents. The participant’s oral consent was noted on a survey questionnaire. The ethical board of the University of Indonesia and the institutional review boards in Germany and the USA approved the use of oral consent.
Immunity was induced in 4 pigs (Group B) trough trickle infection with 100 A. suum eggs 5 times a week for a period of 30 weeks. After a subsequent challenge infection with 2,000 infective A. suum eggs, the number of larvae and liver white spots were significantly reduced in these pigs when compared to control pigs who were not trickle infected (Group A) (Table 1). Challenge control pigs from Group A showed an average of 72.5 ± 43.3 liver white spots and 105 ± 81.3 L4’s in the intestine. In contrast, immunized pigs showed a 99% reduction in number of white spots on the liver (2.0 ± 1.8) and a 100% reduction in the number of L4 stage larvae in the intestine 2 weeks post-challenge, indicating the presence of intestinal immunity against infectious A. suum L3 in these pigs.
Antibodies from the intestinal mucus and from MNC culture supernatant were used to detect immunogenic antigens in A. suum L3 protein extracts (Fig 1). Several antigens were recognized by IgG and IgA antibodies isolated from both the challenge control pigs and the trickle infected pigs. One zone of approximately 12kDa in size was recognised by IgG and IgA antibodies from the immunized pigs (Group B). Although this zone was also detected by the challenge control animals (Group A), recognition was generally much more intense in the immune animals (Group B). This antigen, from here on out referred to as As12, was also detected in the water-insoluble protein fractions of L3 (S1 Fig). In addition, there was also increased recognition of a 21kDa antigen in the L3 PBS extract by mucosal IgA antibodies of immunized pigs.
An immunoblot of the PBS extracts of the different life stages of A. suum showed that the recognition of the As12 antigen was restricted to the early L3 stage and its E/S products (Fig 2A). The antigen was absent from the lung stage larvae onwards. In the L3 PBS extract from A. lumbricoides, a band of the same molecular weight as As12 was recognised (Fig 2B).
The As12 antigen could be visualised by a staining for carbohydrates or after immunoblotting but not by conventional protein staining methods like silver staining or coomassie staining (Fig 3A). The As12 antigen could be purified from the L3PBS protein extract by performing a Folch extraction for the isolation of total lipid. The antigen appeared to be the sole recognized antigen in both the upper methanol and lower chloroform phase after Folch extraction (Fig 3A).
To further explore the specific properties and composition of the As12 antigen, the antigen was subjected to different chemical and enzymatic treatments. The immunoblot reaction to the As12 antigen by purified mucosal antibodies following exposure of the L3 PBS extract to a range of chemical and enzymatic treatments is shown in Fig 3B. Overnight treatment of the L3PBS extract with pronase, lipase, trypsin or pepsin did not affect antibody recognition of the As12 antigen. Similarly, heating the L3PBS extract to 60°C or 90°C overnight did not reduce the recognition of As12, as well as treatment with 1M NaOH. Incubating the L3 PBS antigen with PNGaseF to cut off N-linked glycan structures or subjecting it to ß-N-acetylglucosaminidase treatment for the liberation of terminal ß-linked N-acetylglucosamine and N-acetylgalactosamine residues also did not diminish the recognition of the As12 antigen. The only chemical reactions that reduced or prevented recognition of As12 by porcine antibodies were treatment with 20mM periodic acid at 37°C or 60°C or 1M trifluoracetic acid and 1M HCl at 60°C. Finally, treatment of the As12 fraction with 48% HF for 48 hrs at 4°C for the removal of any PC groups on the antigen also reduced the immune recognition. The presence of a PC group on the As12 antigen was further confirmed by Western blot using specific anti PC antibodies (TEPC-15) (Fig 3C). Next to the As12 antigen, several other PC-bearing antigens were detected in the A. suum L3 extract.
Mass spectrometric analysis of glycans released from As12 after enzymatical treatments with PNGase-F and -A or by chemical β-elimination provided no indications for the occurrence of common N- or O-glycans on As12. Monosaccharide composition analysis indicated that the glycan portion of As12 consists mainly of GalNAc and GlcNAc in an approximate 1:2 ratio, with a minor amount of Glc (S2 Fig).
Freshly hatched A. suum L3 were incubated with antibodies purified from the intestinal mucus of immunized pigs (Group B). On Western blot these antibodies mainly recognized the As12 antigen (Fig 3C). Antibodies bound to the surface were detected with FITC labelled anti-pig total IgG. This showed complete staining of the outer surface of larvae that had shed their L2 sheath (Fig 4A). Other larvae with partially shed L2 sheaths showed only incomplete staining (Fig 4B). To further verify that the reactivity with mucosal antibodies was directed against the As12 antigen, L3 larvae were also incubated with monoclonal antibodies directed against PC. This labelling also resulted in staining of the L3 surface (Fig 4C). Live L3 appear to shed off the antigen-antibody complexes (Fig 4D). As a consequence, the number of L3 stained by mucosal antibodies diminished significantly over time when stained larvae were kept alive in culture. In contrast, when larvae were killed after staining, the number of stained L3 did not diminish over time (Fig 4E).
Results of the vaccination trial with the As12 antigen are shown in Fig 5A. Immunizing pigs with the total purified lipid fraction from A. suum L3, containing the As12 antigen, did not seem to induce protection to subsequent homologue challenge infection. There was no significant difference in number of liver white spots or number of L4 recovered from the intestine 14 days after a challenge infection with 1,000 infective A. suum eggs. Pigs of the vaccinated group did however show a significantly stronger IgG response against the As12 antigen compared to control pigs (Fig 5B).
To evaluate whether the As12 antigen could serve as a serodiagnostic antigen to measure exposure to Ascaris, the serum IgG antibody responses against the As12 antigen in the immunized pigs of group A was measured over time (Fig 6A). Anti-As12 antibodies were significantly elevated from 6 weeks after initial infection. Subsequently, sera from 91 experimentally infected pigs (31) were analysed using the As12 ELISA and showed that the reactivity to the As12 antigen increased significantly in pigs that were continuously infected with A. suum eggs (Fig 6B). After ROC analysis, using the week 0 sera as Ascaris-negative and both week 7 and week 14 sera as Ascaris positive, the cut-off for positive individuals was placed at an ODr of 0.50. Using this cutoff, the sensitivity of the ELISA was 98.4% (95% CI: 95.3–99.7), the specificity was 95.5% (95% CI: 88.9–98.8%) and the area under the curve was 0.99 (± 0.049). No relationship was found between anti-As12 antibody levels and the faecal egg count after 7 or 14 weeks of trickle infection or the number of adult worms at necropsy.
We subsequently tested whether the antibody reactivity to As12 was solely directed against the presence of PC on the antigen. For this, both specific anti-PC monoclonal antibodies (TEPC-15) and pooled purified mucosal antibodies from trickle infected pigs were pre-incubated with increasing concentrations of PC-Cl salt to block the PC-binding sites. After pre-incubation, antibody preparations were used to detect As12 by ELISA. Pre-incubation of TEPC-15 antibodies with increasing concentrations of PC-Cl salt reduced binding reactivity to As12 in a concentration dependent way (Fig 6C). Although pre-incubation with PC-Cl salt also affected the binding reactivity of pooled purified mucosal antibodies, the relative level of inhibition was not as pronounced as for the TEPC-15 antibodies (79% vs 27% and 91% vs 51% reduction of recognition at 10 and 25μg/ml PC-Cl salt respectively). This suggests that the recognition of As12 by Ascaris infected pigs was not solely directed against PC.
Finally, the use of As12 as a serodiagnostic antigen was also validated with human samples. Twenty five human plasma samples from Indonesia from patients with proven A. lumbricoides infection (EPG >50) were analysed using the As12 ELISA (Fig 6D). In addition, 20 non-endemic plasma samples (USA) were used as negative controls and 24 plasma samples from individuals with hookworm infection were tested to evaluate possible cross reactivity. The levels of anti-As12 IgG4 in the human samples was significantly elevated in humans infected with A. lumbricoides. ROC analysis placed the optimal cut-off for the human As12 ELISA at an OD of 0.26. Using this cut-off, the sensitivity of the ELISA was 92.0% (95% CI: 74.0–99.0%), the specificity was 90.0% (95% CI: 68.3–98.8%) and the area under the curve was 0.96 (± 0.026). A total of 24 out of 25 Ascaris infected individuals (96%), 2 out of 20 non-endemic plasma samples (10%) and 2 out of 24 hookworm infected individuals (8.3%) were positive for anti-As12 IgG4 antibodies when this cut-off was employed.
This study describes the identification and characterisation of As12, a phosphorylcholine-containing glycolipid-like structure present on the surface of infective Ascaris L3 larvae. This antigen is targeted by the intestinal antibody response of pigs that have been previously exposed to infection. Both IgG and IgA isotype antibodies against As12 were produced by local antibody secreting cells and both Ig isotypes were detected in the mucus of infected pigs. Recognition of As12 seems to increase over time when pigs are continuously exposed to infection. The As12 antigen was also found in A. lumbricoides L3 and is being recognized by IgG4 from A. lumbricoides infected humans. As a result, an As12 based ELISA test could identify 98% of the Ascaris exposed pigs and 92% of Ascaris infected humans, whereas samples from non-endemic humans or hookworm infected individuals were negative except for two samples in each case.
An amorphous envelope called the surface coat covers the L3 cuticle surface and it presents the greatest interface between the parasite and its host. Our results suggest that the As12 antigen is likely to be a component thereof. The surface coat or “fuzzy coat” is not derived from the cuticle but from specialized secretory glands. It is rich in carbohydrates, is dynamically responsive to changing host environments or immune attack and can be rapidly shed upon binding by antibodies and/or immune cells [33–35]. This process of shedding seems to be actively regulated in A. suum L3 since no decrease in staining was seen in dead L3 stained with immune pig antibodies whereas the percentage of live stained L3 decreased significantly over time. Similarly, antibody recognition of surface antigens on Toxocara canis L3 or Onchocerca cervicalis microfilaria was also lost in a temperature dependent manner or after incubation with antimetabolites [36, 37]. This process of sloughing off antibody-bound antigens is likely to be one of the mechanisms used by Ascaris L3 in order to evade host immune responses. The high antigenicity of the As12 antigen, presented on the L3 cuticle surface, might engage the host immune system while the active shedding of the antigen after antibody binding prevents actual protective mechanisms to damage the mobile larvae. The antigen might also be actively secreted to influence the responses of nearby immune cells.
The 35 kDa carbohydrate antigen (CarLA) detected on L3 of the sheep parasite Trichostrongylus colubriformis [38] shows characteristics similar to As12 as it is also very resistant to multiple enzymatic and chemical treatments, except for those that targeted carbohydrate structures. Although no glycans were detected following PNGase-F and -A digestion or chemical β-elimination, monosaccharide analysis indicated a simple composition suggesting a more polysaccharide nature with repeating units composed of GalNAc and GlcNAc.
Just like As12, CarLA is solely produced and excreted by the infective L3 stage and not by any other life stage [39]. However, unlike As12, the CarLA antigen does not appear to possess any PC group [38]. This might explain why the As12 antigen could not be recognized by immune sheep sera or monoclonal antibodies directed against CarLA. The presence of a PC hapten is very common on nematode carbohydrates attached to a protein or lipid backbone. Previous studies have suggested that PC containing glycolipids from adult Ascaris worms have immunomodulatory properties [40–42]. It is possible that this As12 antigen has similar properties, but further research is required to confirm this.
Despite the apparent antibody response against As12 in vaccinated pigs, there was no protection against subsequent challenge infection. A recent study by Masure et al., [12] has shown strong eosinophil presence in the intestinal tissue of pigs which have developed a ‘pre-hepatic barrier’. Possibly, the vaccination failed to induce an effective local intestinal cellular response which appears necessary to kill and thereby prevent the worms from migrating. Furthermore, systemic vaccination might not have induced the production of anti-As12 antibodies at the level of the intestine, where immune-mediated killing of the larvae is expected to occur [12]. We recognize the unfortunate fact that the presence of anti-As12 antibodies in intestinal mucus was not evaluated in this experiment. It is likely that the synergetic cooperation between both antigen specific antibodies and cellular responses at the site of infection is key in stopping the migrating larvae.
The strong antibody responses that are mounted by the host against As12 may be useful for diagnostic purposes. Nearly all trickle-infected pigs showed strong development of anti-As12 antibodies, independent of whether they were harbouring adult worms in their gut or not. The onset of detectable antibodies was however first apparent after 6 weeks, which is similar to purified Ascaris haemoglobin antigen [15]. Infected humans also mounted a significant antibody response to the As12 antigen, whereas non-endemic or hookworm infected humans did not seem to recognize the antigen. This result, together with the fact that pre-incubation with PC-Cl salt of mucosal antibodies from trickle infected pigs did not completely prevent binding of As12 on ELISA suggests that recognition of As12 not only depends on the PC-group but possibly involves other parts of the antigen. More tests with sera from individuals with other well-characterized helminth infections need to be performed to further determine the species specificity of As12 as a serodiagnostic antigen.
In conclusion, in this study we used locally secreted antibodies from the intestine of pigs with a pre-hepatic immunity to identify one major immunodominant antigen in the extracts and E/S material of infective Ascaris L3. This As12 antigen is stage specific, of a glycolipid nature and is being actively secreted. Experimentally infected pigs or naturally infected humans develop a measurable antibody response against the As12, advocating its possible use as a diagnostic antigen. The exact structure of this antigen and its biological role during infection is yet unknown and deserves further clarification.
|
10.1371/journal.ppat.1002924 | A Novel Rhabdovirus Associated with Acute Hemorrhagic Fever in Central Africa | Deep sequencing was used to discover a novel rhabdovirus (Bas-Congo virus, or BASV) associated with a 2009 outbreak of 3 human cases of acute hemorrhagic fever in Mangala village, Democratic Republic of Congo (DRC), Africa. The cases, presenting over a 3-week period, were characterized by abrupt disease onset, high fever, mucosal hemorrhage, and, in two patients, death within 3 days. BASV was detected in an acute serum sample from the lone survivor at a concentration of 1.09×106 RNA copies/mL, and 98.2% of the genome was subsequently de novo assembled from ∼140 million sequence reads. Phylogenetic analysis revealed that BASV is highly divergent and shares less than 34% amino acid identity with any other rhabdovirus. High convalescent neutralizing antibody titers of >1∶1000 were detected in the survivor and an asymptomatic nurse directly caring for him, both of whom were health care workers, suggesting the potential for human-to-human transmission of BASV. The natural animal reservoir host or arthropod vector and precise mode of transmission for the virus remain unclear. BASV is an emerging human pathogen associated with acute hemorrhagic fever in Africa.
| We used deep sequencing, a method for generating millions of DNA sequence reads from clinical samples, to discover a novel rhabdovirus (Bas-Congo virus, or BASV) associated with a 2009 outbreak of 3 human cases of acute hemorrhagic fever in Mangala village, Democratic Republic of Congo (DRC), Africa. The cases, presenting over a 3-week period, were characterized by abrupt disease onset, high fever, bloody vomiting and diarrhea, and, in two patients, death within 3 days. BASV was present in the blood of the lone survivor at a concentration of over a million copies per milliliter. The genome of BASV, assembled from over 140 million sequence reads, reveals that it is very different from any other rhabdovirus. The lone survivor and a nurse caring for him (with no symptoms), both health care workers, were found to have high levels of antibodies to BASV, indicating that they both had been infected by the virus. Although the source of the virus remains unclear, our study findings suggest that BASV may be spread by human-to-human contact and is an emerging pathogen associated with acute hemorrhagic fever in Africa.
| Viral hemorrhagic fever (VHF) encompasses a group of diseases characterized by fever, malaise, bleeding abnormalities, and circulatory shock [1], [2], [3]. Quality research on these infections is hindered by the fact that they are sporadic and often occur in geographically remote and politically unstable regions of the developing world. Most VHF diseases are associated with a short incubation period (2–21 days), abrupt onset, rapid clinical course, and high mortality, placing VHF agents amongst the most virulent human pathogens [4]. All known VHFs are zoonoses, and to date have been attributed to only four families of enveloped, single-stranded RNA viruses – Arenaviridae, Bunyaviridae, Filoviridae and Flaviviridae. Viruses from these families have caused major deadly outbreaks on the African continent (Fig. 1). Lassa fever virus (Arenaviridae) causes an estimated 500,000 cases each year in West Africa [5]. Crimean-Congo hemorrhagic fever (CCHF) and Rift Valley Fever viruses (Bunyaviridae) are associated with outbreaks in West, South and East Africa [6]. Ebola and Marburg viruses (Filoviridae) have caused several sporadic human outbreaks with high mortality (50–90%) in Central Africa, where they have also decimated local great ape populations [7]. Yellow fever and dengue viruses (Flaviviridae) are widely distributed throughout Sub-Saharan Africa where they cause both endemic and sporadic epidemic diseases in human populations [8].
Rhabdoviruses are members of the family Rhabdoviridae and order Mononegavirales and are enveloped viruses with single-stranded, negative-sense RNA genomes [9]. Their genomes encode at least five core proteins in the following order: 3′-nucleoprotein (N), phosphoprotein (P), matrix protein (M), glycoprotein (G) and large protein, or RNA-dependent RNA polymerase (L)-5′ (N-P-M-G-L). Rhabdoviruses are currently divided into six genera, with the two genera Ephemerovirus and Vesiculovirus, together with about 130 unclassified viruses, forming the dimarhabdovirus supergroup (“dipteran mammal-associated rhabdovirus”) [10]. Notably, although rhabdoviruses span all continents and exhibit a wide host range, infecting plants, invertebrates, vertebrate animals, and humans, relatively few are known to cause human infections. Rabies virus (RABV) and related viruses from the Lyssavirus genus and Chandipura virus (CHPV) from the Vesiculovirus genus are known to cause acute encephalitis syndromes [11], [12]. Other viruses from the genus Vesiculovirus cause vesicular stomatitis (mucosal ulcers in the mouth) and “flu-like” syndromes in both cattle and humans [13].
Unbiased next-generation or “deep” DNA sequencing is an emerging method for the surveillance and discovery of pathogens in clinical samples [14]. Unlike polymerase chain reaction (PCR), deep sequencing does not rely on the use of target-specific primers. Thus, the technique is particularly useful for the identification of novel pathogens with high sequence divergence that would elude detection by conventional PCR assays. Deep sequencing has been used previously to discover a new hemorrhagic fever-associated arenavirus from southern Africa, Lujo virus [15], as well as a new polyomavirus in human Merkel cell carcinoma [16]. With the depth of sequence data now routinely extending to >100 million reads, de novo genome assembly of novel viruses directly from primary clinical samples is feasible, as demonstrated by assembly of the 2009 pandemic influenza H1N1 virus genome from a single patient's nasal swab without the use of a reference sequence [17]. Here we report the critical role of deep sequencing in the discovery of a novel rhabdovirus associated with a small outbreak of fulminant hemorrhagic fever in the remote village of Mangala, Bas-Congo province, Democratic Republic of Congo (DRC), between May 25 and June 14, 2009.
A cluster of three human cases of typical acute hemorrhagic fever occurred between May 25 and June 13, 2009 in Mangala village, located in a remote tropical forest region in Central Africa. Cases were characterized by abrupt disease onset, high fever of >39°C when present, overt hemorrhagic symptoms with epistaxis, conjunctival injection, mouth and gastrointestinal bleeding, followed by death within 3 days of symptom onset in two patients (Table 1). The first patient, who died <48 hours after presentation, exhibited hemorrhagic symptoms without a documented fever, and only the third adult patient recovered from his illness. All three patients lived within a 2500-m2 area in the same neighborhood of Mangala, a remote village in Bas-Congo province of DRC (Fig. 1). The first two patients died rapidly in Mangala village, and no blood samples were collected. A blood sample was collected from the third surviving patient three days after symptom onset and sent to Centre International de Recherches Médicales de Franceville (CIRMF) for etiological diagnosis. The sample tested negative by TaqMan real-time PCR assays for all viruses known to cause acute hemorrhagic fever in Africa (data not shown).
To identify a potential causative pathogen in the third surviving patient with unknown hemorrhagic fever, RNA extracts from the serum sample were analyzed using unbiased deep sequencing (Fig. 2). The initial Roche 454 pyrosequencing library yielded a total of 4,537 sequence reads, of which only a single 220 bp read (0.022%) aligned with any annotated viral protein sequence in GenBank. The translation product showed similarity to a segment of the L protein, or RNA-dependent RNA polymerase, from Tibrogargan and Coastal Plains rhabdoviruses, with 41% identity to Coastal Plains virus (GenBank ADG86364; BLASTx E-score of 2×10−6). This finding suggested the presence of a novel, highly divergent rhabdovirus in the patient's serum. Attempts to extend the initial sequence by primer walking or PCR using rhabdovirus consensus primers failed due to limited sample availability; thus, we resorted to ultra-deep sequencing on an Illumina HiSeq 2000. Out of the 140,164,344 reads generated from Illumina sequencing, 4,063 reads (0.0029%) had nucleotide or protein homology to rhabdoviruses with an E-score of <10−5. These reads were used as “seeds” for iterative de novo assembly, resulting in construction of an estimated 98.2% of the genome of the novel rhabdovirus. We provisionally named this rhabdovirus BASV, or Bas-Congo virus, referring to the province from which the outbreak originated.
The coverage of BASV achieved by deep sequencing was at least 10-fold across nearly the entire genome and included 29,894 reads out of ∼140 million (0.021%) (Fig. 2). The viral load in the patient's serum was 1.09×106 RNA copies/mL by quantitative RT-PCR. The only moderately high titer is consistent with the fact that the sampled patient was a survivor of BASV infection and would thus be anticipated to have relatively lower viral titers in the blood, as also seen for survivors of Ebola virus infection [18].
Cultivation of the patient's serum in Vero, BHK, LLC-MK2 (rhesus monkey kidney), CCL-106 (rabbit kidney) and C6/36 (Aedes albopictus mosquito) cell cultures failed to show cytopathic effect, and serial quantitative BASV RT-PCR assays on primary and passaged cell culture supernatants turned negative. Subsequent electron microscopy of inoculated cell cultures was negative for viral particles. In addition, no illnesses or deaths occurred in suckling mice inoculated intracerebrally with the BASV-positive serum and observed over 14 days.
Phylogenetic trees reveal that BASV belongs to the dimarhabdoviridae supergroup and is distantly related to members of the Tibrogargan group and the Ephemerovirus genus, although it clusters separately from other rhabdoviruses in an independent deeply rooted branch (Figs. 3 and 4; Fig. S1). Comparative analysis of the concatenated BASV proteins with representative dimarhabdoviruses reveals very low overall amino acid pairwise identity of 25.0 to 33.7%, depending on the virus (Fig. 5). Notably, BASV diverges significantly from either of the two main recognized human pathogens among rhabdoviruses, rabies virus or Chandipura virus.
The sequence divergence of BASV relative to other rhabdoviruses is also correlated with differences in genome structure (Fig. 5). The prototype genome organization of rhabdoviruses, found in lyssaviruses, is N-P-M-G-L. However, molecular analysis of novel rhabdoviruses has often revealed more complex genomes, with up to 10 additional open reading frames (ORF) located within an existing gene or interposed between the five core genes [19], [20], [21]. Rhabdoviruses from the Tibrogargan group (TIBV and CPV) share a distinctive genome structure with three additional genes, two between M and G (U1 and U2) and one between G and L (U3) [22]. Interestingly, BASV also has these three additional genes (U1–U3), confirming the phylogenetic relationship and overall structural similarity to the Tibrogargan group viruses. Based on their size, the U3 proteins of TIBV, CPV, and presumably BASV are candidate viroporins [22]. BASV is more distant structurally and phylogenetically from the Ephemero and Hart Park Group rhabdoviruses (Figs. 3 and 4), which do not contain U1 or U2 genes, but rather an additional two or three genes between G and L (including a putative U3 viroporin in BEFV referred to as the alpha-1 protein) (Fig. 5, asterisk). Moussa virus (MOUV), another rhabdovirus recently discovered in Africa (Fig. 4), does not contain any accessory genes but instead, shares the prototype N-P-M-G-L rhabdovirus structure [23].
To confirm that BASV is infectious to humans, convalescent sera were collected in early 2012 from surviving Patient 3 as well as five additional health care workers from Mangala identified as close contacts and tested in a blinded fashion for the presence of neutralizing antibodies to BASV (Fig. 6). Two of the six sera tested strongly positive with 50% protective doses between 1∶1,000 and 1∶5,000 (Figs. 6A and 6F). Moreover, the observed neutralization was highly specific for BASV-G, since no neutralization was observed with pseudoviruses harboring the vesicular stomatitis virus glycoprotein (VSV-G). One of the neutralizing sera had been collected from surviving Patient 3 (Fig. 6A, “Patient 3”), whereas the other serum sample, containing even higher titers, corresponded to an asymptomatic nurse directly caring for Patient 3 during his period of acute hemorrhagic illness (Fig. 6F, “Contact 5”). Specifically, Contact 5 was the primary health care provider to Patient 3 at the health center and during his transfer to the general hospital at Boma. All 6 individuals, including Patient 3, tested negative for BASV viremia by specific RT-PCR (data not shown).
BASV was not detected by PCR in 43 serum samples from other unknown cases or outbreaks of hemorrhagic fever reported in the DRC from 2008–2010 (Fig. 7A, pink). Five of these 43 samples originated from the Bas-Congo outside of Mangala village and the Boma Bungu Health Zone. In total, the unknown hemorrhagic cases/outbreaks spanned 9 of the 11 provinces in the DRC, and all 43 samples also tested negative by PCR for the known hemorrhagic fever viruses circulating in Africa (data not shown). Fifty plasma samples collected from randomly selected blood donors in the Kasai-Oriental province of DRC (Fig. 7A, star; Table S2) were also screened and found to be negative for BASV-neutralizing antibodies (Fig. 7B).
Among more than 160 species of rhabdoviruses identified to date, fewer than 10 have been isolated from humans [24]. In addition, while human infection by rhabdoviruses has previously been associated with encephalitis, vesicular stomatitis, or “flu-like” illness, the discovery of BASV is the first time that a member of the Rhabdovirus family has been associated with hemorrhagic fever in humans with a fulminant disease course and high fatality rate. To our knowledge, this is also the first successful demonstration of de novo assembly of a novel, highly divergent viral genome in the absence of a reference sequence and directly from a primary clinical sample by unbiased deep sequencing.
Several lines of evidence implicate BASV in the hemorrhagic fever outbreak among the 3 patients in Mangala. First, this virus was the only credible viral pathogen detected in the blood of the lone survivor during his acute hemorrhagic illness by exhaustive deep sequencing of over 140 million reads. Analysis of the Illumina deep sequencing reads for the presence of other viral pathogens yielded only endogenous flora or confirmed laboratory contaminants (Table S1 and Fig. S2). Some enteric pathogens, such as E. coli O157:H7, Campylobacter, Shigella, and Salmonella, are diagnosed through fecal laboratory testing and not blood, and have been associated with hemorrhagic diarrhea [25]. However, these outbreaks are typically foodborne and associated with larger clusters and much greater numbers of clinical cases than reported here [26], [27], [28]. Furthermore, enteric diarrheal cases rarely present with systemic symptoms such as fever or generalized mucosal hemorrhage, with bleeding most often limited to the gastrointestinal tract, and overall mortality rates are generally low [26]. Thus, the clinical syndrome observed in 3 patients with hemorrhagic fever in the DRC, a region endemic for viral hemorrhagic fevers, is much more consistent with infection by a VHF disease agent. BASV is a plausible hemorrhagic fever candidate because it is a novel, highly divergent infectious virus, thus of unknown pathogenicity, and was detected at a titer of >1 million copies/mL in blood from an acutely ill individual. In addition, there is ample precedent for hemorrhagic disease from rhabdoviruses, as members of the genus Novirhabdovirus cause severe hemorrhagic septicemia in fresh and saltwater fish worldwide [29] (Fig. 4). The detection of BASV seropositivity in an asymptomatic close contact (Fig. 6) is not surprising given that up to 80% of patients infected with Lassa virus do not exhibit any hemorrhagic fever symptoms [30], [31].
Prior to the BASV outbreak, no hemorrhagic disease cases had been reported in Boma Bungu Health Zone. BASV was also not detected in 43 serum samples from unknown, filovirus-negative cases or outbreaks of hemorrhagic fever from 2008–2010 spanning 9 of the 11 provinces in the DRC (Fig. 7A). In addition, a serosurvey of 50 random blood donors from Kasai-Oriental province in central DRC was negative for prior exposure to BASV (Fig. 7B). Taken together, these data suggest that the virus may have emerged recently and locally from Boma Bungu in Bas-Congo, DRC.
We were unable to isolate BASV despite culturing the RNA-positive serum in a number of cell cultures and inoculation into suckling mice. One explanation for these negative findings may be that the virus inoculation titers of <50 µL were insufficient, although this is surprising given the concentration of >1 million copies per mL of BASV in blood from the lone survivor. A more likely explanation is viral inactivation resulting from the lack of adequate cold chain facilities in remote Boma Bungu. Viral RNA can often still be detected by RT-PCR in sera that is culture-negative [32]. In support of this premise, we have observed that the BASV-G/VSVΔG-GFP pseudotyped virus efficiently infects and replicates in a variety of insect and mammalian (including human) cell lines (Steffen, et al., manuscript in preparation). In the absence of a positive culture, a “reverse genetics” approach to produce recombinant BASV particles, if successful, would greatly facilitate further study of the virus, as established previously for other rhabdoviruses such as VSV [33].
Based on our findings, some speculations on the origin of and routes of transmission for BASV can be made. All 3 patients became ill with acute hemorrhagic fever over a 3-week period within the same 2500-m2 area of Mangala village, suggesting that all 3 cases were infected with the same pathogen. Waterborne or airborne transmission would be expected to result in more numerous cases than the 3 reported. There were no reports of animal die-offs that would suggest potential exposures to infected wild animals or livestock. Taken together, these observations suggest that an unknown arthropod vector could be a plausible source of infection by BASV. This hypothesis is consistent with the phylogenetic and structural relationship of BASV to rhabdoviruses in the Tibrogargan group and Ephemerovirus genus, which are transmitted to cattle and buffalo by Culicoides biting midges [9]. In addition, the recent discovery of Moussa virus (MOUV), isolated from Culex mosquitoes in Cote d'Ivoire, Africa [23], implies the presence of hitherto unknown arthropod vectors for rhabdoviruses on the continent. Nevertheless, at present, we cannot exclude the possibility of other zoonotic sources for the virus or even nosocomial bloodborne transmission (as Patients 1 and 2 have not clearly been established to be BASV cases by serology or direct detection), and the natural reservoir and precise mode of transmission for BASV remain unknown. A community-based serosurvey in Boma Bungu and an investigation to track down potential arthropod or mammalian (e.g. rodents and bats) sources for BASV are currently underway.
Although we cannot exclude the possibility of independent arthropod-borne transmission events, our epidemiologic and serologic data do suggest the potential for limited human-to-human transmission of BASV. Patient 3, a nurse, had directly taken care of Patients 1 and 2 at the health center, and another nurse (Contact 5), who had taken care of Patient 3 (but not Patients 1 or 2) had serologic evidence of asymptomatic BASV infection. We present a hypothetical model for BASV transmission during the hemorrhagic fever outbreak in which the initial infection of two children in Mangala (Patients 1 and 2) was followed by successive human-to-human transmission events involving two healthcare workers (Patient 3 and Contact 5) (Fig. 8). This pattern of transmission from the community to health care workers is also commonly seen in association with outbreaks of Ebola and Crimean-Congo hemorrhagic fever [6], [34].
While rhabdoviruses are distributed worldwide, some authors have suggested that the Rhabdoviridae family probably originated from tropical regions of the Old or New World [9]. The discovery of BASV in Central Africa suggests that additional rhabdoviruses of clinical and public health importance likely await identification, especially in these poorly investigated geographic regions. Active epidemiological investigation and disease surveillance will be needed to fully ascertain the clinical and public health significance of BASV infection in humans, as well as to prepare for potentially larger human outbreaks from this newly discovered pathogen.
Written informed consent for publication of their case reports was obtained from the sole survivor of the hemorrhagic fever outbreak and the parents of the two deceased children. Written informed consent was obtained from the surviving patient and 5 of his close contacts for analysis of the serum samples reported in this study. Samples were analyzed under protocols approved by the institutional review boards of University of California, San Francisco, the University of Texas Medical Branch, and the National Institute of Biomedical Research (INRB) and CIRMF in Gabon, and the Institutional Animal Care and Use Committee (IACUC) of the University of Texas Medical Branch.
No diagnostic samples were available from Patient 1 or Patient 2. Blood was collected in a red top serum tube from Patient 3 on June 16, during the acute phase, three days after hemorrhagic onset. The sample was transported at 4°C to the BSL-4 facility at CIRMF. Serum was obtained by centrifugation at 2300 rpm for 10 min. No other acute samples from Patient 3 were available. In January of 2012 (∼2.5 years after the outbreak), convalescent sera were collected from Patient 3 and close contacts (other workers at the health center) for BASV neutralization testing. Forty-three serum samples from other unknown hemorrhagic fever cases or outbreaks representing 9 of 11 provinces in the DRC were available for BASV PCR testing (Fig. 7A). Fifty available plasma samples from random blood donors (median age 27.5 years; age range 1–76 years) in Kasai Oriental province, DRC, were also tested for antibodies to BASV (Fig. 7A and B; Table S2).
RNA was extracted from 140 µl of serum using the QIAamp viral RNA mini kit (Qiagen). Taqman real-time reverse-transcription-PCR (RT-PCR) testing for known hemorrhagic fever viruses was performed using primers and probes specific for Marburg virus (MARV), all four species of Ebola virus (Zaire, ZEBOV; Sudan, SEBOV; Côte d'Ivoire, CIEBOV, and Bundibugyo, BEBOV), Crimean-Congo hemorrhagic fever virus (CCHFV), Yellow fever virus (YFV), Dengue virus (DENV), Rift Valley fever virus (RVFV) and Chikungunya virus (CHIKV) (available upon request).
200 µL of serum sample were inactivated in 1 mL of TRIzol (Invitrogen), and nucleic acid extraction and purification were performed according to the manufacturer's instructions. Roche 454 pyrosequencing using randomly amplified cDNA libraries was performed as described previously [35]. Viral sequences were identified using BLASTn or BLASTx by comparison to the GenBank nonredundant nucleotide or protein database, respectively (E-score cutoff = 10−5).
To recover additional BASV sequence, two sets of cDNA libraries were prepared from DNase-treated extracted RNA using a random PCR amplification method as described previously [36], or random hexamer priming according to the manufacturer's protocol (Illumina). The libraries were then pooled and sequenced on two lanes of an Illumina HiSeq 2000. Raw Illumina sequences consisting of 100 base pair (bp) paired-end reads were filtered to exclude low-complexity, homopolymeric, and low-quality sequences, and directly compared using BLASTn or BLASTx alignments to a library consisting of all rhabdovirus sequences in GenBank. The initial read obtained by 454 pyrosequencing as well as other reads aligning to rhabdoviruses were then inputted as “seeds” into the PRICE de novo assembler [37] (Fig. 2), with a criterion of at least 85% identity over 25-bp to merge two fragments. De novo assembly of the BASV genome was performed iteratively using PRICE and the Geneious software package (Biomatters) [38]. The near-complete whole genome sequence of the novel rhabdovirus (∼98.2% based on protein homology to other rhabdoviruses) was determined to at least 3× redundancy by de novo assembly as well as PCR and Sanger sequencing of low-coverage regions. Sanger sequencing was also performed to verify the accuracy of the assembly and confirm the genomic organization of BASV (Fig. 2).
Rapid classification of the ∼140 million 100-bp paired-end Illumina reads was performed using a modified cloud computing-based computational analysis pipeline [17] (Veeraraghavan, Sittler, and Chiu, manuscript in preparation). Briefly, reads corresponding to human sequences were taxonomically classified using SOAP and BLAT software [39], [40]. Other reads were then identified using BLASTn or BLASTx by comparison to GenBank-derived reference databases (E-score cutoff = 10−5).
To estimate the viral load in the patient's serum, we first designed a set of specific PCR primers for detection of BASV targeting the L protein, BASV-F (5′- CGCTGATGGTTTTTGACATGGAAGTCC-3′)/BASV-R (5′-TAAACTTCCTCTCTCCTCTAG-3′), for use in a SYBR-Green real-time quantitative RT-PCR assay. A standard curve for the assay was constructed as described previously [36]. The viral load in the patient's serum was determined by comparison to the standard curve.
Predicted open reading frames (ORFs) in the BASV genome were identified with Geneious [38]. Multiple sequence (Figs. 3 and 4; Fig. S1) and pairwise (Fig. 5) alignments of BASV proteins relative to corresponding proteins from other rhabdoviruses were calculated using MAFFT (v6.0) with the E-INS-i option and at default settings [41]. To generate the phylogeny trees, all rhabdoviruses in GenBank were included as well as representative members of other families within the order Mononegavirales. Bayesian tree topologies were assessed with MrBayes V.32 software (20,000 sampled trees; 5,000 trees discarded as burn-in) [42]. Convergence was confirmed by the PSRF statistic in MrBayes, as well as by visual inspection of individual traces using TRACER from the BEAST software package [43]. Trees were visualized after midpoint rooting with FigTree V1.31 [43].
Initial attempts were made to culture the virus using a total of 200 µL of BASV-positive serum inoculated onto confluent monolayers of Vero E6 and C6/36 (Aedes albopictus mosquito) cells in 6-well plastic tissue culture plates at 37°C and 28°C, respectively, in a 5% CO2 environment as previously described [44]. From 20–50 µL of serum were used to inoculate the cells, which were examined daily for cytopathic effect (CPE) at days 5, 7, and 14. Supernatants were harvested and two additional blind passages were performed, each passage followed by 14 days of observation for CPE. Cell culture supernatants were also monitored for evidence of viral replication by quantitative RT-PCR.
Using the remaining 100 uL of BASV-positive serum, further attempts were made to culture the virus in 5 cell lines and in suckling mice. The serum sample was split in half and diluted 1∶20 or 1∶10 in phosphate-buffered saline with 20% fetal bovine serum (FBS) to allow sufficient volume to inoculate cell cultures or mice, respectively. The first diluted sample was inoculated intracerebrally into a litter (n = 12) of 1 day old mice. Pups were observed daily for 14 days for lethality or signs of clinical illness. The second diluted sample was inoculated into 12.5 cm2 tissue culture flasks of Vero, BHK, LLC-MK2 (rhesus monkey kidney), CCL-106 (rabbit kidney) and C6/36 cells. Vertebrate cells were held at 37°C for 14 days and observed for evidence of CPE. Mosquito cells were maintained at 28°C for 10 days. Since no CPE was observed in any of the cultures, cells were subsequently fixed for transmission electron microscopy to see if viral particles could be visualized [45].
A pseudotype system based on a vesicular stomatitis virus (VSV) construct carrying a reporter gene for green fluorescent protein (VSVΔG-GFP) and bearing the predicted synthesized BASV glycoprotein (BASV-G) was used to generate a serum neutralization assay for BASV. Briefly, the predicted BASV glycoprotein (BASV-G) was synthesized (Genscript) and subcloned into the pCAGGS expression plasmid. Human embryonic kidney 293T cells were seeded (DMEM + 10% FBS + penicillin/streptomycin + Glutamax (Gibco) + non-essential amino acids (Gibco)) in 10 cm culture dishes 24 hours prior to transfection. Cells were transfected with 20 µg BASV-G, VSV-G, or empty pCAGGS DNA per dish following a calcium phosphate transfection protocol [46]. The culture medium was replaced 15 hours post-transfection and cells were stimulated with 6.2 mM valproic acid for 4 hours before the medium was replaced again. At 36 hours post-transfection the transfected cells were infected with VSVΔG-GFP/VSV-G pseudotypes at a multiplicity of 0.1–0.3. The inoculum was removed after 4 hours and replaced by fresh culture medium. At 24 hours post-infection, infectious supernatants were harvested, filtered through 0.45 µm filters, and concentrated 10-fold by centrifugation through a 100-kDA filter (Millipore). Concentrated viruses were aliquoted and stored at −80°C.
For serum neutralization testing, human hepatoma Huh-7 cells were seeded (DMEM +10% FBS + penicillin/streptomycin + Glutamax (Gibco) + non-essential amino acids (Gibco)) in 48-well plates 24 hours prior to infection. Per well 10 µl of pseudovirus harboring either BASV-G or VSV-G (adjusted to obtain 25–50% infection of target cells) was mixed with 10 µl of the respective serum dilution and incubated for 45 minutes at 37°C. Subsequently, the mix was added to the target cells (performed in triplicate) and cells were incubated for 24 hours at 37°C. The infected cells were detached with trypsin and washed with PBS before fixing with 2% paraformaldehyde for 1 hour at room temperature. GFP expression in infected cells was quantified by flow cytometry using a LSR II (BD Biosciences) and the collected data was analyzed with FlowJo software (TreeStar).
The annotated, nearly complete sequence of BASV has been submitted to GenBank (accession number JX297815). Deep sequencing reads have been submitted to the NCBI Sequence Read Archive (accession number SRA056894). Accession numbers used for the phylogenetic analyses in Figs 3, 4, and S1 are listed as follows, in alphabetical order: ABLV, Australian bat lyssavirus (NP_478343); ARAV, Aravan virus (ABV03822), BEFV, Bovine ephemeral fever virus (NP_065409); BYSMV, Barley yellow striate mosaic virus (BYSMV); CDV, Canine distemper virus (AAR32274); CHPV, Chandipura virus (ADO63669); CPV, Coastal Plains virus (ADG86364); COCV, Cocal virus (ACB47438); DURV, Durham virus (ADB88761); DUVV, Duvenhage virus (ABZ81216); EBLV1, European bat lyssavirus 1 (ABZ81166), EBLV2, European bat lyssavirus 2 (ABZ81191); EBOV, Ebola virus (AAG40171, AAA79970, BAB69010); EVEX, Eel virus European X virus (CBH20130); FDLV, Fer-de-lance virus (NP_899661); FLAV, Flanders virus (AAN73288); HeV, Hendra virus (NP_047113); HIRRV, Hirame rhabdovirus (ACO87999); HMPV, Human metapneumovirus (L_HMPVC); HPIV-1, Human parainfluenza virus type 1 (AAA69579); HPIV-2, Human parainfluenza virus type 2 (CAA40788); HPIV-3, Human parainfluenza virus type 3 (AAA46854); HPIV-4, Human parainfluenza virus type 4 (BAJ11747); INHV, Infectious hematopoietic necrosis virus (NP_042681); IRKV, Irkut virus (ABV03823); ISFV, Isfahan virus (CAH17548); KHUV, Khujand virus (ABV03824); LBV, Lagos bat virus (ABZ81171); LNYV, Lettuce necrotic yellows virus (YP_425092); MARAV, Maraba virus (AEI52253); MARV, Marburg virus (YP_001531159); MeV, Measles virus (AF266288); MMV, Maize mosaic virus (YP_052855); MOKV, Mokala virus (ABZ81206); MOUV, Moussa virus (ACZ81407); MUV, Mumps virus (AF201473); NCMV, Northern cereal mosaic virus (NP_597914); NDV, Newcastle disease virus (ADH10207); NGAV, Ngaingan virus (YP_003518294); NiV, Nipah virus (AAY43917); OVRV, Oak Vale rhabdovirus (AEJ07657); PFRV, Pike fry rhabdovirus (ACP28002); RABV, Rabies virus (NP_056797); RSV, Respiratory syncytial virus (NP_056866); RYSV, Rice yellow stunt rhabdovirus (NP_620502); SIGMAV, Sigma virus (ACU65444); SCRV, Siniperca chuatsi rhabdovirus (YP_802942); SHRV, Snakehead virus (AAD56771); SMRV, Scophthalmus maximus rhadovirus (ADU05406); SVCV, Spring viremia of carp virus (NP_116748); SYNV, Sonchus yellow net virus (NP_044286); TIBV, Tibrogargan virus (ADG86355); TUPV, Tupaia virus (AAX47602); TVCV, Tomato vein clearing virus (YP_224083); VHSV, Viral hemorrhagic septicemia virus (NP_049550); VSIV, Vesicular stomatitis virus, Indiana (NP_041716); VSNJV, Vesicular stomatitis virus, New Jersey (P16379); WCBV, West Caucasian bat virus (ABV03821); WONGV, Wongabel virus (YP_002333280).
|
10.1371/journal.pbio.2002623 | Transcriptome analysis of hypoxic cancer cells uncovers intron retention in EIF2B5 as a mechanism to inhibit translation | Cells adjust to hypoxic stress within the tumor microenvironment by downregulating energy-consuming processes including translation. To delineate mechanisms of cellular adaptation to hypoxia, we performed RNA-Seq of normoxic and hypoxic head and neck cancer cells. These data revealed a significant down regulation of genes known to regulate RNA processing and splicing. Exon-level analyses classified > 1,000 mRNAs as alternatively spliced under hypoxia and uncovered a unique retained intron (RI) in the master regulator of translation initiation, EIF2B5. Notably, this intron was expressed in solid tumors in a stage-dependent manner. We investigated the biological consequence of this RI and demonstrate that its inclusion creates a premature termination codon (PTC), that leads to a 65kDa truncated protein isoform that opposes full-length eIF2Bε to inhibit global translation. Furthermore, expression of 65kDa eIF2Bε led to increased survival of head and neck cancer cells under hypoxia, providing evidence that this isoform enables cells to adapt to conditions of low oxygen. Additional work to uncover -cis and -trans regulators of EIF2B5 splicing identified several factors that influence intron retention in EIF2B5: a weak splicing potential at the RI, hypoxia-induced expression and binding of the splicing factor SRSF3, and increased binding of total and phospho-Ser2 RNA polymerase II specifically at the intron retained under hypoxia. Altogether, these data reveal differential splicing as a previously uncharacterized mode of translational control under hypoxia and are supported by a model in which hypoxia-induced changes to cotranscriptional processing lead to selective retention of a PTC-containing intron in EIF2B5.
| Tumor hypoxia is a negative prognostic factor for many solid cancers. Cellular adaptation to hypoxia is largely mediated by widespread changes in gene expression and enables cancer cells to adjust and survive. Recently, alternative splicing has been implicated in this process. To identify biologically impactful hypoxia-responsive isoforms, we took an unbiased approach to deeply sequence RNA of normoxic and hypoxic head and neck cancer cells. This analysis identified >1,000 mRNAs as alternatively spliced under hypoxia, including a significant enrichment of alternatively spliced genes involved in adaptation to hypoxia. Most notably, we discovered a unique retained intron in the translation initiation factor EIF2B5 that creates a premature termination codon. We show that this retained intron leads to a 65kDa truncated isoform that opposes full-length eIF2Bε to inhibit global translation and enhances survival of head and neck cancer cells under hypoxia. Strikingly, this intron and several additional hypoxia-induced retained introns were overexpressed in solid tumors relative to normal tissues. Mechanistically, we propose that intron retention under hypoxia is influenced by changes to RNA polymerase II (RNAPII) activity at weak 3′ splice sites and carry out experimental validation for the retained intron in EIF2B5 to investigate intron retention under hypoxia.
| Reduced availability of oxygen, or hypoxia, is a major feature of solid tumors that contributes to metastasis and resistance to therapy [1]. Tumor hypoxia occurs due to several physiological factors, such as limited diffusion of oxygen and irregular vascular structure [2]. While oxygen levels can be measured directly in tumors, there is an immediate need to develop noninvasive clinical markers of hypoxic burden in tumors. Molecular imaging markers such as pimonidazole and fluorescence-based compounds [3–5] have been developed and refined to specifically label hypoxic tumors, which allows for specific interrogation of hypoxic gene expression programs within the tumor microenvironment [6]. Hypoxia-mediated changes in expression can be dynamic and robust, impacting pathways critical to tumor development and survival, such as angiogenesis, metabolism, and macromolecular synthesis [7,8]. Consequently, there is a nearly universal negative correlation between the level of hypoxia in tumors and overall survival in patients of many solid cancers, including head and neck squamous cell carcinoma (HNSC) [9]. As such, hypoxia “metagene” expression signatures have been successfully implemented as a surrogate method to classify tumor hypoxia for HNSC and other solid malignancies, including breast and prostate cancer [10,11].
Hypoxic stress influences processing and translation of mRNAs by regulating the levels and activity of diverse factors, including Hypoxia-Inducible transcription Factors (HIFs), small noncoding RNAs and miRNAs, and RNA binding proteins (RBPs) [12–14]. For example, RBPs such as HuR and PTB bind to and regulate the stability and localization of key regulators of hypoxic response such as HIF1α [15,16] and miRNA-199a [17]. Several kinases known to phosphorylate major RBPs and splicing factors are also hypoxia responsive [18]. Consequently, alternative splicing of select target genes of HIF1α has been reported in hypoxic cells [19]. Likewise, expression of noncoding mRNA isoforms are induced under hypoxia in part due to changes in splicing [20]. Several splicing factors, including SF3B1, are up-regulated in a HIF1α-dependent manner under physiological conditions of hypoxia in cardiac myocytes [21]; however, it remains unclear precisely how mRNA splicing is regulated during periods of oxygen deprivation in cancer cells and what the resulting biological implications are. Intriguingly, regulation of splicing is frequently altered in cancer and is affected by the same signaling pathways that are differentially regulated in hypoxic tumor microenvironments [22]. Moreover, many solid cancers affected by hypoxia display widespread alterations in splicing [23].
While splicing of specific genes has been shown to be dependent on the activity of HIF1α, differences in transcription elongation are known to impact regulation of cotranscriptional splicing [24–26]. Thus, we hypothesized that hypoxia-mediated changes to the RNA processing and transcription machinery could lead to extensive differences in mRNA splicing. Hypoxia was identified to induce phosphorylation of the C-terminal domain (CTD) of RNA polymerase II (RNAPII) and was observed to enhance binding of cofactors and increase control of transcriptional activation of HIF target genes [27]. Additional findings support the theory that changes in the activity and rate of transcription elongation play a key role in the maturation and processing of mRNAs under hypoxia. Under hypoxia, there are fewer changes in RNAPII binding near gene promoters but instead an increased accumulation of RNAPII observed along gene bodies [28]. Therefore, to better understand the link between hypoxia-mediated changes to mRNA regulation and to investigate the biological role of alternative splicing in response to hypoxia in cancer, we deeply sequenced mRNA of hypoxic and normoxic HNSC SQ20B cells.
The data led to the identification of more than 1,000 transcripts affected by alternative splicing under hypoxia and revealed 3 types of mRNA splicing as specifically enriched, including an increase in retained introns (RIs) in hypoxia compared to normoxia. Strikingly, for more than 90% of genes in this category, hypoxia increased the occurrence of RIs relative to normoxia, which is a phenomenon also observed in 16 cancer types, including head and neck, colon, breast, and lung cancers [23]. We found evidence of several hypoxia-induced RIs expressed in solid tumor data, indicating that tumor hypoxia contributes to this type of splicing. Most notably, a unique RI in the master regulator of translation initiation, EIF2B5, was significantly overexpressed in both head and neck and kidney renal clear cell tumors relative to normal tissues in a stage-dependent manner. Here, we present compelling evidence that hypoxia leads to retention of an intron in EIF2B5 that creates a phylogenetically conserved premature termination codon (PTC). This alternate transcript results in a truncated protein isoform of eIF2Bε predicted to lack enzymatic guanine exchange factor (GEF) activity. Remarkably, we demonstrate that the resulting truncated isoform of eIF2Bε is induced under hypoxia and provide data to show that this isoform acts in opposition to full-length eIF2Bε to inhibit protein synthesis. Cellular adaptation to hypoxia entails adjusting key metabolic processes, such as translation, to low energy due to reduced availability of oxygen [29]. Control of translation initiation specifically contributes to down-regulation of protein synthesis under hypoxia, which occurs through phosphorylation of eIF2α by hypoxia-mediated induction of the integrated stress response [30]. Here, we discover hypoxia-mediated induction of a dominant-negative isoform of eIF2Bε as a secondary method to inhibit translation and increase survival of head and neck cancer cells in periods of acute or prolonged hypoxia. We further investigated how splicing of EIF2B5 is controlled and uncovered several factors that contributed to the hypoxia-induced RI, including a weak splice site at the intron:exon junction, hypoxia-induced expression and binding of SRSF3, and an accumulation of RNAPII specifically at the RI. Some of these findings extend to the broader class of genes with RIs under hypoxia, supportive of a mechanism by which changes to RNAPII elongation contributes to retention of introns with weak 3′ splice sites under hypoxia.
The transcriptomes of normoxic and hypoxic SQ20B cells (maintained in 0.5% O2 for 16 h) were compared to identify expression differences at an individual mRNA transcript level. The RefSeq hg19 reference comprised of 46,017 transcript models was used for annotation. Of the 24,812 transcripts expressed in SQ20B cells, we detected 3,114 that significantly changed expression in hypoxia compared to normoxia (P < 0.05, false discovery rate [FDR] < 5%, Fragments Per Kilobase of transcript per Million mapped reads [FPKM] ≥ 0.5). In total, 1,519 transcripts representing 1,473 genes were induced and 1,595 transcripts expressed from 1,563 genes were repressed (FDR < 5%). Pathway-based analysis identified “cell adhesion,” “response to hypoxia,” and “metabolism” among the most enriched categories for hypoxia-induced transcripts (P < 0.01, DAVID GO [31]). Induction of select HIF1α target genes was validated by quantitative PCR (qPCR) (S1A Fig). Repressed genes were involved in regulating processing, stability, and translation of RNAs, with “ribosome biogenesis,” “nucleosome organization,” and “RNA splicing” as highly significant ontology groups (P < 0.01). This included core regulators of alternative splicing, such as the major splicing factor SF1, several serine–arginine splicing factors (SRSF1, SRSF3, and SRSF7), SF3 genes, and many translation initiation factors, including EIF2B family members, EIF5, and EIF6. Notably, a closer examination of genes involved in regulation of mRNA transcription, translation, and processing revealed that the clear majority of these genes were repressed under hypoxia (Fig 1A).
Variation in expression of individual transcripts of the same gene by hypoxia could be masked when analyzing expression changes at the total gene level; therefore, we used levels of individual isoforms to classify genes as hypoxia-responsive (defined as a gene with 1 or more individual transcripts detected as significantly induced or repressed, P < 0.05, FDR < 5%). This method uncovered 50 additional genes that may not yet be reported as hypoxia-regulated, including genes known to mediate protein localization (SEC24B, ARL17A, PLEKHA8, MLPH, STXBP5, MON2, and EXOC1), Mitogen-activated protein kinase kinase signaling pathway (MAP4K3, DUSP22, and MAP3K13), and key regulators of RNA metabolism (ZNF519, JDP2, CTBP2, ZNF248, RBM5, NFAT5, CREB3L2, SMARCA1 and ZFHX3). Consistent with the finding that transcript-level changes comprise an additional layer of hypoxia-regulated expression changes, only approximately 3% of genes that expressed multiple isoforms in SQ20B cells have been reported to be transcriptionally controlled by HIF binding at hypoxia-responsive elements (HRE) [32], suggesting an HRE-independent mode of regulation. This is not surprising, as regulation of transcription and mRNA splicing are observed to be distinct processes that impact different subsets of genes [33].
Next, we focused on those genes that expressed more than 1 transcript in SQ20B cells to assess changes in patterns of isoform expression. Altogether, of the 5,418 genes identified to express >1 transcript, 937 of these genes showed hypoxia-induced changes in expression of 1,015 transcripts (P < 0.05, FDR < 5%). Most these genes contained a single hypoxia-responsive isoform; only 8% of genes that expressed multiple transcripts showed more than 1 isoform that significantly changed expression in hypoxia. The isoform-level changes in expression were validated for select genes that displayed differential expression of isoforms predicted to have different biological functions (S1B Fig). Hypoxia selectively induced the MXI1 isoform, which codes for the shortest protein isoform that would lack 46 amino acids at the N-terminus compared to full-length isoforms. This difference was determined to alter the ability of the truncated protein to antagonize N-Myc activity and impact cell proliferation [34]. Likewise, the isoform of NDRG1 that exhibited the strongest induction arises from an alternative transcription start site, which leads to 66 fewer amino acids and exclusion of a proteolytic cleavage site that remains in the other two isoforms of NDRG1. Transcripts of 2 additional genes, NEK6 and FAM86C1, were validated as being selectively repressed under hypoxic conditions (S1C Fig).
We reasoned that hypoxia-responsive isoforms predicted to carry out different functions than isoforms expressed in normoxia would be the most biologically impactful changes that warranted further study. Therefore, we used the program MISO [35] to carry out an additional exon-level approach to identify changes in gene structure based on 8 annotated categories of alternative splicing. Hypoxia led to a change in expression for 1,103 alternatively spliced loci representing 819 unique genes (Fig 1B, ΔPsi > 10%, Bayes Factor > 20). Notably, there was a significant comparative enrichment for hypoxia-induced changes in 3 specific event types: expression of alternate last exons (ALEs), RIs, and tandem 3′ UTRs (TUTR) (Fig 1B). For the genes in these 3 splicing categories, gene ontology revealed processes central to hypoxic adaptation as significantly enriched, including “cellular protein metabolism,” “programmed cell death,” and “gene expression” (Fig 1C, DAVID GO, P < 0.05, S1 Data). Remarkably, nearly 90% of the genes in the RI category displayed increased retention of introns in hypoxic compared to normoxic cells (Fig 1D). Among these genes was ANKZF1, a gene implicated in mitochondrial and endoplasmic reticulum-associated protein degradation [36,37], the translation initiation factor EIF2B5, TGFB1, and the metionyl-TRNA synthetase, MARS. The RIs in these genes were validated by PCR using primers spanning the intron junction and cDNA prepared from oligo-dT–selected mRNA (Fig 2A–2D, S2 Fig).
To confirm expression of these hypoxia-induced introns in additional datasets and determine if they were expressed in patient samples, we interrogated The Cancer Genome Atlas (TCGA) and analyzed expression solid tumors known to be affected by hypoxic fractions [38–40]. This analysis validated increased expression of the hypoxia-induced RIs for ANKZF1, EIF2B5, MARS, and TGFB1 in HNSC tumors relative to matched normal tissues (Fig 2E–2H). The extended analysis of additional cancer types confirmed significantly increased expression of these hypoxia-induced RIs for 2 types of renal carcinoma, as well as lung, liver, and prostate cancers (S3 Fig).
Intron 12 of EIF2B5 showed a strong stage-dependent increase in expression for both head and neck and kidney renal clear cell carcinomas (KIRC) (Fig 3A and 3B). This trend was even more apparent with HNSC patients in late-stage disease, with some individuals exhibiting nearly 8-fold increased expression of intron 12 compared to controls (Fig 3A). These data suggest hypoxia-induced retention of EIF2B5 intron 12 may result in meaningful biological effects and encouraged us to examine the functional impact of this RI.
The RI in EIF2B5 stood out as the strongest candidate for functional studies for several additional reasons: (a) hypoxia led to a >40% increase in retention of intron 12 in a background of an overall 2-fold decrease in total expression of EIF2B5 (ΔѰ = 0.44, Bayes Factor > 20); (b) the retention occurred specifically at a single locus of EIF2B5; and (c) retention of intron 12 creates a PTC that remains in frame with the coding sequence (Fig 3C). The genomic locus around the PTC is highly conserved, even among lower organisms. Intriguingly, this stop codon is part of the 5′ splice site consensus sequence “GURAGU,” where URA can be either UAA or UGA (both of which would create a stop codon). This suggests a strong evolutionary pressure to preserve an early termination precisely at this location, where inclusion of the PTC may be influenced through regulation of splice site choice.
Increased expression of EIF2B5_intron12 was further confirmed by qPCR using intron-specific primers in a reaction with cDNA prepared from oligo-dT–selected mRNA (Fig 3D). Additionally, a deeper analysis of changes in splicing of EIF2B5 was carried out to closely examine inclusion of intron 12 and to validate the occurrence of this RI using another method; the software package MAJIQ [41] was used to assess local splicing variation in EIF2B5 for annotated and de novo splice events. A hypoxia-induced increase in expression of intron 12 was confirmed using this approach (S4 Fig), but additional sites of local splicing variation did not show significant (>20%) hypoxia-influenced changes.
Computational analyses approximate up to 20%–35% of alternatively spliced transcripts could contain PTCs and become targets of nonsense-mediated decay (NMD) [42,43]; however, in cases where NMD is inhibited, transcripts can be stabilized and subsequently translated into truncated proteins [44]. Moreover, NMD surveillance typically recognizes stop codons as premature if the stop occurs more than 50 nucleotides upstream of a splice junction [45]. Due to the unusual nature of this intron retention event and the role of hypoxia in suppressing NMD in an eIF2α phosphorylation-dependent manner [46], we predicted that this isoform would not be subject to NMD but would rather be translated into a truncated protein (Fig 4A). Consistent with this hypothesis, the MAJIQ splicing analysis of EIF2B5 revealed a 40%–50% decrease in expression of remaining exons following intron 12 (S4 Fig). These data support the notion that transcripts that retain intron 12 and the subsequent PTC would undergo read-through of intron 12 into intron 13, resulting in a reading frame for a truncated protein variant. Indeed, we observed induction of a 65kDa protein isoform of eIF2Bε under various conditions of hypoxia consistent with the predicted PTC inserted upon retention of intron 12 (Fig 4B). Induction of phospho-eIF2α was used as a marker for hypoxia in these experiments (Fig 4B). To further test whether this 65kDa protein detected in the immunoblot was indeed a truncated isoform of eIF2Bε, we used small interfering RNA (siRNA) to specifically target the entire EIF2B5 gene or intron 12 alone. Using this approach, we observed a substantial reduction in the levels of the 65kDa isoform under both conditions (Fig 4C, S5A Fig). We additionally observed hypoxia-induced expression of 65kDa eIF2Bε in the colorectal cancer cell line RKO (Fig 4D), demonstrating that this is not a cell-line specific event. Moreover, ultraviolet (UV) radiation (Fig 4E), but not thapsigargin-induced endoplasmic reticulum (ER) stress (S5B Fig), also led to induction of the truncated eIF2Bε, suggesting that expression of this isoform is induced by specific cell stresses. Furthermore, whole-cell lysates were isolated from hypoxic SQ20B cells, and proteins migrating at 80kDa and 65kDa were subjected to liquid chromatography tandem mass spectrometry (LC-MS/MS). Peptides corresponding to eIF2Bε were detected in both the 80kDa and 65kDa size analytes. The peptides corresponding to the 65kDa-sized isoform of eIF2Bε were located near the N-terminus or middle of the eIF2Bε sequence, consistent with a C-terminal truncation (Fig 4F). To additionally rule out that expression of the band migrating at 65kDa is a degradation product or may occur due to proteolytic cleavage, we used a plasmid to express a version of eIF2Bε with a C-terminal FLAG-tag in SQ20B cells and exposed these cells to UV radiation. If the tagged eIF2Bε were to undergo proteolytic cleavage, we predicted to observe a 15kDa band reactive to FLAG-tag antibody in addition to the 65kDa band reactive against eIF2Bε antibody in UV-treated cells. This experiment did not show evidence of the 15kDa tagged proteolytic cleavage product (S5C Fig). Finally, we carried out an experiment to test for the involvement of NMD in leading to the expression of truncated eIF2Bε. Although NMD is known to be inhibited in conditions of oxygen deprivation [46], NMD and alternative splicing are coupled processes, which can lead to alternative PTC-containing transcripts that ultimately become targets of NMD [47]. To rule out any major contribution of NMD in leading to expression of 65kDa eIF2Bε, we used siRNA to knock down expression of UPF1. UPF1 is a necessary component of the SURF complex, which is required for NMD [48,49]. This experiment failed to produce an increase in expression of the truncated protein (S5D Fig), further supporting the notion that this isoform occurs due to alternative splicing.
To identify potential regulators mediating retention of intron 12 in EIF2B5, we utilized the AVISPA tool [50] to carry out a splicing-relevant sequence feature analysis of the locus encompassing EIF2B5 exons 12–14. The nonmotif analysis revealed a relatively short distance to the nearest AG dinucleotide upstream of exon 13, as well as relatively unstructured RNA immediately downstream of exon 13 (Fig 5A). Both features indicate weakened splicing potential of exon 13 [51,52]. The motif analysis revealed sequence motifs, such as potential RBP binding motifs, which could be important for regulating splicing of the locus (Fig 5B). Notably, these sequence features were not identified in a search of additional loci within EIF2B5. As a control, we used AVISPA to predict the occurrence of alternative splicing in 4 additional exon triplets within EIF2B5 (exons 6–8, 7–9, 8–10, and 9–11). These loci were not predicted to be alternatively spliced. Altogether, these results suggest that alternative splicing is significantly more likely to occur at the intron 12 locus compared to the control loci tested.
The motif analysis also identified potential trans regulators with binding sites within this locus predicted to regulate splicing of this region, including NOVA, HNRNPC, and members of the CUGBP Elav-Like Family (CELF), as well as Ser/Arg-rich splicing factor 3 (SRP20, AKA SRSF3), HNRNPG, NPTB, HNRNPF, and Ser/Arg-rich splicing factor 2 (SC35, AKA SRSF2) (Fig 5B). Interestingly, several of the splicing factors predicted to have the greatest impact on regulation of this locus also changed expression under hypoxia, including SRSF2, SRSF3, HNRNPC, HNRNPF, and RBMX (AKA HNRNPG), which were repressed at the mRNA level, and CELF5 (a member of the CELF family), which was induced (Fig 1A, FC ≥ |1.2|, FDR < 5%). Upon closer examination of these splicing factors, we focused in on SRSF3 and CELF5 as prime candidates to test as regulators of EIF2B5 splicing, due to a large number of binding sites near the intron12:exon13 junction (S6A Fig) and the fact that these factors showed >2-fold changes in mRNA expression under hypoxia.
We next assayed for changes in protein expression of these splicing factors in hypoxic cells and observed a reproducible hypoxia-induced increase in SRSF3 protein (Fig 5C, S6B Fig) and a modest decrease in expression of CELF5. To test for their requirement in the splicing of EIF2B5 and production of the resulting 65kDa protein isoform, we used siRNA to knock-down expression of these factors and assayed for changes in expression of 65kDa eIF2Bε. Upon knockdown of SRSF3, we saw a concurrent disappearance in the expression of the 65kDa isoform of eIF2Bε in conditions of normoxia or hypoxia (Fig 5D), while siRNAs against other RBPs, such as CELF5, did not display the same effect (S6C and S6D Fig). RNA immunoprecipitation assays were next carried out to confirm a direct interaction between SRSF3 protein and EIF2B5 RNA. The data showed enrichment of SRSF3 at both exons flanking EIF2B5_intron12 (Fig 5E), validating the predicted binding sites observed in other systems (S6A Fig). Furthermore, binding of SRSF3 was increased in hypoxic cells relative to normoxic cells, providing additional support for the role of SRSF3 in regulating hypoxia-induced retention of intron 12.
In addition to the influence of sequence elements, mRNA splicing is a co-transcriptionally regulated process which is impacted by coordination of RNAPII. Phosphorylation of the CTD of RNAPII is known to impact the rate of transcriptional elongation, pausing, and the relative rate of splicing [53,54]. Therefore, we assayed for changes in phosphorylation of the CTD of the largest subunit of RNAPII in HNSC cells. Surprisingly, we saw a large (>10-fold) and reproducible increase in levels of phosphorylation at serine 2 residues and a concomitant decrease in phosphorylation of serine 5 and 7 residues of the CTD in hypoxic cells compared to normoxic controls (Fig 6A, S7 Fig). Interestingly, accumulation of the phospho-Ser2 form of RNAPII has been associated with RIs compared to constitutively spliced introns [55]. To assay for changes in binding of total and phospho-Ser2 RNAPII, we next used chromatin immunoprecipitation followed by qPCR. We observed a significant hypoxia-induced enrichment in both forms of RNAPII at intron 12 but did not detect the same enhanced binding at another nearby intron of EIF2B5 that did not undergo intron retention under hypoxia (Fig 6B).
In strong support of these data, cells expressing mutant forms of RNA polymerase with slower elongation rates led to extensive changes in mRNA splicing, including specific retention of EIF2B5 intron12 [57]. Moreover, 31 of 100 genes we identified to be affected by intron retention in hypoxia were classified as “rate-sensitive” and displayed altered expression in the RNA polymerase elongation mutants [57], suggesting that hypoxia-mediated changes in elongation of RNAPII likely influence splicing of additional loci. Those genes alternatively spliced in conditions of slow elongating RNAPII generally exhibited weaker 3′ splice sites compared to loci insensitive to changes in elongation rate [57], so we next carried out an analysis of splice site strength for genes affected by intron retention under hypoxia. This analysis detected significantly weaker 3′ splice sites for genes with changes in RIs under hypoxia (score = 7.6) compared to a control set of splice sites not affected by hypoxia (score = 8.9) (Fig 6C, P = 0.00289). This included intron 12 of EIF2B5, where the 3′ splice site strength score was 1.1 points lower compared to the 3′ splice site strength of the control set.
Next, we tested whether the expression of 65kDa eIF2Bε has an impact on the biological function of the endogenous, full-length eIF2Bε protein. The truncated protein isoform created from retention of intron 12 is predicted to lack 2 critical domains that occur in the C-terminus (Fig 4A): a GEF domain and a region required for interaction with eIF2α [58]. Thus, we hypothesized that the 65kDa isoform of eIF2Bε would inhibit translation and lead to reduced protein synthesis. To test this, we constructed a plasmid expressing the truncated 65kDa isoform using site-directed mutagenesis to insert a stop codon within 3 nucleotides of where retention of intron 12 results in a PTC (Fig 7A). Expression of this mutated version of eIF2Bε under normoxic conditions resulted in the appearance of a 65kDa protein isoform consistent with the size of the endogenous protein that is induced under hypoxia (Fig 7B).
To analyze the impact of the truncated isoform on protein synthesis, we used pulse-labeling of 35S methionine/cysteine in cells expressing 65kDa eIF2Bε, full-length eIF2Bε, or empty vector. Proteins isolated from the 35S methionine/cysteine-labeled cells were resolved on SDS-PAGE, after which the gels were dried and exposed to autoradiograph film to detect signal intensities as a measure of total protein synthesis. There was a pronounced and reproducible decrease (approximately 30%) in total protein synthesis in cells expressing 65kDa eIF2Bε compared to empty vector, while expression of full-length eIF2Bε did not have the same effect (Fig 7B and 7C).
Translation levels were assessed relative to cells expressing empty vector as a control in order to best isolate the effects of the truncated isoform from those of endogenous full-length eIF2Bε. Effects of inducing full-length eIF2Bε were not used as a control because full-length eIF2Bε remains stably expressed under hypoxic conditions and is not induced. In addition, expression of eIF2B has been shown to destabilize ternary complex formation under conditions where eIF2 is phosphorylated [59].
To further verify the effects observed in the 35S assay, we carried out polysome profiling as an additional measure of protein synthesis. The data confirmed a decrease in the total polysome profiles of cells expressing 65kDa eIF2Bε compared to control cells (Fig 7D, S8A Fig). Moreover, this decrease in translation was comparable to the decrease in translation observed in hypoxic conditions (Fig 7E). Consistent with these data, we observed an increase in adenosine triphosphate:adenosine monophosphate (ATP:AMP) ratio in cells expressing 65kDa eIF2Bε relative to cells expressing empty vector or the full-length isoform (S8B Fig). Finally, because down-regulation of translation is a known mechanism by which tumor cells overcome hypoxic stress [60], we predicted that expression of 65kDa eIF2Bε and the resulting repression of translation may promote survival of hypoxic cells. To test this, clonogenic assays were performed to measure survival and proliferation. The surviving fraction of cells expressing 65kDa eIF2Bε decreased under normoxic conditions but was significantly higher when cells were grown in 0.5% O2 conditions for either 16 h or 24 h (Fig 7F and 7G, S8C and S8D Fig, P < 0.01, Student t test). Collectively, these data establish a role for the 65kDa isoform of eIF2Bε in reducing protein synthesis in head and neck cancer cells to adapt to conditions of hypoxia (Fig 8). Mechanistically, we provide strong evidence that expression of this isoform is influenced by differential binding of SRSF3 at the locus of EIF2B5 intron12, a weak splice site coupled with an alternate splice site, and increased binding of total and P-Ser2 RNAPII (Fig 8).
This study has validated alternative splicing as a major, additional layer of complexity to the gene expression response to hypoxia. Previous work had described hypoxia-mediated changes in splicing, including in several known HIF-target genes [19,61] and further identified splicing as a means to induce expression of noncoding RNAs under hypoxia [20]. Here, we expand upon the current understanding of RNA processing under hypoxia and demonstrate that decreased oxygen in cancer cells leads to extensive changes in splicing, including a striking increase in retention of over 100 introns affecting a significant number of genes with key functional roles in cellular adaptation to hypoxia. As such, a major focus of our study became to investigate the biological significance of hypoxia-induced intron retention in the master regulator of translation initiation, EIF2B5.
Specific retention of intron 12 in EIF2B5 led to a previously undescribed 65kDa isoform of eIF2Bε that decreases overall protein synthesis in head and neck cancer cells. Full-length eIF2Bε is a necessary component of the eIF2B complex, containing both the active GEF domain and a region for association with eIF2α [62]; this complex binds eIF2α and exchanges guanosine diphosphate (GDP) for guanosine triphosphate (GTP) to initiate translation. During hypoxia, eIF2α is phosphorylated, and translation initiation is inhibited and was demonstrated to be critical for cell survival under extreme hypoxia [60]. However, this phosphorylation is not long-lasting and can be removed by GADD34 [29,63]. Moreover, in cells stably expressing an unphosphorylatable form of eIF2α with an S51A mutation, translation initiation resumed under hypoxia, but only to approximately 75% of the level observed in control cells [60]. These data suggested the existence of additional mechanisms to sustain reduced protein synthesis in conditions of low oxygen. Thus, we propose that, during periods of acute or prolonged hypoxia, intron retention in EIF2B5 leads to expression of a truncated dominant-negative isoform to further inhibit protein synthesis in cancer cells (Fig 8). Expression levels of 65kDa eIF2Bε are consistently up-regulated in hypoxia, with the greatest induction of this isoform observed at stringent (0.2% O2) and prolonged (≥ 16 h) hypoxic conditions (Fig 4B). Moreover, SQ20B cells expressing 65kDa eIF2Bε demonstrated an increase in ATP:AMP ratio and increased clonogenicity in conditions of low oxygen (S8 Fig). Altogether, these data establish a biological role for the 65kDa isoform of eIF2Bε in promoting adaptation to hypoxic stress in cancer cells.
The question remains as to whether induction of this isoform is a cause or effect of tumorigenesis; additional animal studies are necessary to evaluate its potential to influence tumor formation and will be a focus of future work. Intriguingly, altered regulation of eIF2B GEF activity and translational control has been observed in transformed cells relative to primary cells [64]. In conditions that activate the unfolded protein response (UPR), transformed cells displayed increased levels of GDP exchange from eIF2 relative to normal human cell lines, despite comparable levels of phospho-eIF2α [64]. These data provide an additional line of evidence that cancer cells in conditions of stress may require mechanisms outside of phospho-eIF2α to control eIF2B GEF activity and overall levels of protein synthesis.
Our findings led to the first reported case of splicing as a previously uncharacterized mode of translational control under conditions of hypoxia. The mechanism behind hypoxia-induced alternative splicing is influenced by several factors, including both -cis RNA sequence determinants and oxygen-sensitive -trans regulators. We used retention of intron 12 in EIF2B5 as a starting point to investigate regulation of intron retention in hypoxic cells. Several unique aspects of intron 12 may lead to its retention in EIF2B5, such as a weak 3′ splice site and a weakened splicing potential in this region of EIF2B5 influenced by a relatively short distance to the nearest AG dinucleotide upstream of exon 13, as well as the relatively unstructured RNA immediately downstream of exon 13. Intron length and GC content near the splice site can also influence splice site choice [65,66]. Alternatively, RIs often exhibit lower GC content compared to the immediate upstream and downstream exons [66]. Interestingly, the GC content of EIF2B5 intron 12 was 6% lower than that of the adjacent downstream exon 13. This was the intron:exon junction where our AVISPA analysis detected a relatively weaker splice site and a possible “GT” alternate splice site downstream in intron 13. The unique properties of this locus may promote specific retention of intron 12 compared to other introns under hypoxia.
The influence of hypoxia-responsive -trans regulators on splicing of EIF2B5 was also substantial. Our data support the theory that hypoxia-mediated changes in expression of splicing factors and regulation of transcription elongation contribute to splice site choice. We observed down-regulation of many genes that regulate splicing and RNA processing, including some that were predicted to bind at the EIF2B5 intron 12 locus. There is evidence that global down-regulation of splicing factors and RBPs can promote intron retention in a regulated manner during other physiological responses or processes, such as hematopoiesis [67]. This mechanism likely contributes to the RIs in some genes under hypoxia. However, we uncovered that the splicing factor predicted to contain the most binding sites within the EIF2B5 intron12 locus, SRSF3, was induced at the protein level under hypoxia and exhibited increased binding at the alternatively spliced locus under hypoxia (Fig 5C and 5E). Reducing levels of SRSF3 substantially decreased expression of the 65kDa protein isoform. The full-length eIF2Bε protein was reduced as well, suggesting a major role of SRSF3 in the processing and splicing of EIF2B5 transcripts. While there was not a significant enrichment for predicted SRSF3 binding sites for the class of 100 genes affected by RIs under hypoxia, we did identify a common feature of relatively weaker 3′ splice sites for this group of genes (Fig 6C). Previous work described that alternatively spliced loci with relatively weaker 3′ splice sites were more sensitive to changes in transcription elongation [57].
Thus, we propose that regulation of EIF2B5 intron12 and additional introns retained under hypoxia are likely influenced by hypoxia-mediated changes in activity of RNAPII. Under hypoxia, changes in phosphorylation of the CTD of RNAPII influence expression of HIF-1α target genes by affecting the binding of cofactors and the kinetics of transcriptional activation [27]. Our observation that UV radiation was another stress that also led to an induction of 65kDa eIF2Bε strongly supports this hypothesis. UV exposure is known to induce pyrimidine dimers and other blocks to transcription elongation, which alter RNAPII transcription kinetics and subsequently impact regulation of splicing [68,69]. We posit that other stresses that have an impact on transcriptional elongation will likely impact expression of 65kDa eIF2Bε as well.
Furthermore, hypoxia-mediated changes in phosphorylation of RNAPII may explain why we observed an enrichment of splicing changes at the 3′ end of genes (i.e., ALE, TUTR, and RI categories). Several RBPs, including polyadenylation factors and splicing regulators, interact specifically with phospho-Ser2 modifications of the CTD [70]. Intriguingly, increased occupancy of phospho-Ser2 RNAPII is associated with RIs compared with constitutively spliced introns [55]. Our data detected an enrichment of phosphor-Ser2 binding under hypoxia at intron 12 in EIF2B5, and future work by our group will determine whether this translates to other hypoxia-induced RIs with weak 3′ splice sites as well.
The physiological relevance of these findings is underscored by the fact that EIF2B5 intron 12 is overexpressed in tumor versus normal tissues of patients affected by 6 different cancers, including HNSC (Fig 2, S3 Fig). Interestingly, intron retention is relatively increased for nearly all solid tumors compared to normal tissue [23]. An interrogation of the TCGA database uncovered evidence of several hypoxia-induced RIs identified in this study as overexpressed in solid tumors relative to matched normal tissues (Fig 2, S3 Fig), supporting the notion that the hypoxic tumor microenvironment is a contributor to the intron retention observed in solid tumors.
There is a critical need to understand this important form of RNA processing and regulation in a cancer context. Many hypoxia-responsive isoforms, including those with RIs, have the potential to influence key biological pathways in cancer cells. For example, the hypoxia-induced RI in TGFB1 is predicted to create a different 5′ UTR and alternate transcription start site, which would code for a 416 amino acid peptide instead of the full-length 689 amino acids. This change in the C-terminus would affect part of a FAS1 domain involved in binding integrin to regulate cell adhesion, as well as an EMI (EMILIN protein family) domain that is thought to be a protein–protein interaction domain. Additional work is needed to investigate the functional consequences of additional hypoxia-induced introns, such as those in TGFB1 and MARS, predicted to create alternative protein isoforms. The identification of stress-responsive isoforms with biological functions that may differ from isoforms expressed under normal conditions will enable a deeper understanding of the link between stresses within the tumor microenvironment and regulation of RNA processing and splicing. These data will provide a new layer of information to refine prognostic hypoxia gene expression signatures and to investigate appropriate biological pathways to target hypoxic cancer cells.
SQ20B cells, derived from human head and neck squamous cell cancer, were obtained from American Type Culture Collection (Rockville, MD). The RKO cells were a generous gift from Dr. Cho (University of Chicago). Both cell lines were maintained in Dulbecco’s Modified Eagle Medium (DMEM) media supplemented with 4.5 g/L D-Glucose, 1X L-glutamine, 10% Fetal Bovine Serum, and 1X Penicillin/Streptomycin and cultured in a 37°C humidified 5% CO2 atmosphere. For oxygen deprivation experiments, cells were incubated in 37°C humidified 5% CO2 conditions with varying concentrations of O2 in an INVIVO2 400 chamber (Baker BioScience Solutions).
RNA was isolated from cells using the Trizol reagent (ThermoFisher) and purified according to the manufacturer’s protocol. All purified RNA was subsequently treated with DNaseI digestion to remove possible DNA contaminants (Qiagen). The quality of RNA used for cDNA library preparation was verified using the RNA nano 6000 analysis chip on a BioAnalyzer 2000 series instrument (Agilent Technologies) to ensure an RNA integrity value greater than or equal to 9.
The cDNA libraries for sequencing were prepared from poly(A)+-selected mRNA, according to Illumina’s TruSeq Stranded mRNA sequencing preparation kit. Briefly, 1 μg RNA was purified for mRNA. Then mRNA material was fragmented and denatured, in preparation for first- and second-strand cDNA synthesis steps. Finally, the 3′ ends were adenylated to ligate strand-specific adapter sequences to cDNA material and amplified using PCR. Purity and size of cDNA library products were confirmed using a BioAnalyzer instrument. Library concentrations were determined via RT-qPCR using the Library Quantification Kit (KapaBiosystems). The samples were then prepared and sequenced on an Illumina HiSeq Series instrument, with 1 sample per sequencing lane to achieve > 2 x 108 reads per sample.
The sequencing data were aligned using a RefSeq hg19 reference with STAR software (version 2.3.0.1), resulting in an average of 1.7 x 108 uniquely aligned reads per sample. The Cufflinks software suite (version 2.1.1) was used for differential expression analysis, with standard parameters and RefSeq hg19 reference annotations. Gene ontology analyses were carried out using DAVID software [31].
The mixture of isoforms software, MISO, version 0.5.1 (February 23, 2014 release) was used for the exoncentric isoform quantification analysis. Each of the 4 hypoxia and normoxia replicates were merged into 1 file for each treatment for MISO analysis. For the exoncentric analysis, the hg19 GFF3 annotation files for each of the splicing event categories (A3SS, A5SS, AFE, ALE, MXE, RI, SE and Tandem UTR) were downloaded from http://genes.mit.edu/burgelab/miso/ as human genome (hg19) alternative events v1.0. Standard analysis parameters were used, with a filter option applied to require a minimum of 20 reads to support an event identification.
To identify regulatory elements that may affect retention of EIF2B5 intron 12, we used AVISPA [50]. This method, based on computationally derived splicing codes, has been used previously to detect and experimentally verify novel regulators of exon splicing in a variety of experimental conditions [71–73]. Since AVISPA was built for analyzing differential splicing determinants around cassette exons, we extracted hg19 coordinates for Ensembl-defined exons 11–13 and 12–14, then analyzed the genomic regions of these 2 triplet exons using AVISPA. The loci containing triplets of exons 6–8, 7–9, 8–10, and 9–11 were also analyzed to serve as negative controls, as these exons are known to be constitutive and not exhibit intron retention. The splicing-related top motifs and regulatory features were defined by their normalized feature effect (NFE) and their relative enrichment compared to alternative and constitutive exons. Briefly, the NFE value represents the effect on splicing prediction outcome if a motif is removed in silico, normalized by the total effects observed from removing each of the top features in this way.
PCR reactions were carried out with cDNA prepared from RNA treated with DNaseI (Qiagen) to minimize contamination of DNA and in a reverse transcription reaction using oligo-dT primers to enrich for mature mRNA. Reverse transcription was carried out according to manufacturer’s protocol (Taqman RT reagents). PCR reactions were carried out in a PTC-100 Thermocycler (MJ Research, Inc.) for 40 cycles, using annealing temperatures optimal for each primer set. (Primer sequences available in Supporting Material). qPCR reactions were prepared using Power SYBR Green PCR Master Mix (Applied Biosystems) and carried out on a QuantStudio 6 Flex Real-Time PCR Instrument (Thermo Fisher Scientific). Primer sequences made available in S1 Table.
SQ20B cells were plated 24–48 h before transfection and grown to approximately 60%–70% confluency. RNAi transfections were carried out using a mixture of lipofectamine RNAi Max (ThermoFisher) diluted in OptiMEM media with siRNA to 10–50 nM, which was added to cell culture plates in complete DMEM. Cells were placed in incubator for 24 h, at which point the media was replaced. After an additional 24 h, cells were either harvested or used for subsequent experiments. For expression experiments, plasmids were purchased from Origene. Site-directed mutagenesis was used to alter the original plasmid and introduce a TAG stop codon, and confirmed by Sanger sequencing (see Fig 7A). Expression plasmids were transfected in cells plated to the same confluency as described above using a mixture of lipofectamine 2000 reagent (ThermoFisher) diluted in OptiMEM media with varying concentrations of pCMV expression plasmid added to cells in complete DMEM. The plates of cells were incubated for 4–12 h, at which point the media was washed off and replaced. Cells were harvested or used for downstream experiments 24–48 h post-transfection.
Whole cell lysates were collected using a lysis buffer of 2% Triton-X, 1X Complete Mini protease inhibitor cocktail (Roche), and 1X phosphatase inhibitor cocktail 2 (Sigma) in PBS. The nuclear/cytosol fractionation reagents (BioVision) supplemented with 1X phosphatase inhibitor cocktail 2(Sigma) were used to extract cytoplasmic and nuclear extracts from the same sample. Lysates isolated from frozen tumor and normal mouse tissue were lysed in a buffer containing 1% Triton X-100, 50 mM HEPES, ph 7.4, 150 mM NaCl, 1.5 mM MgCl2, 1 mM EGTA, 100 mM NaF, 10 mM Na pyrophosphate, 1 mM Na3VO4, 10% glycerol, protease inhibitors (Roche #04693116001), and phosphatase inhibitors (Roche #04906845001). All protein concentrations were determined using DC protein assay (BioRad). Equal amounts of protein were resolved on 10% or 12% sodium dodecyl sulfate polyacrylamide gels and transferred to polyvinylidene fluoride membranes. Membranes were blocked with 5% nonfat dried milk in TBS-T (20 mM Tris, 137 mM NaCl, 0.1% Tween-20; pH 7.6) and then incubated in a 1:1,000 dilution of primary antibody in 5% milk/TBS-T followed by 1:5,000 dilution of secondary antibody in 5% milk/TBS-T. After washing, membranes were treated with ECL chemicals and exposed to autoradiograph film.
The LC-MS/MS was carried out at the Wistar core facility. Complete protocol available online: https://www.wistar.org/our-science/shared-facilities/proteomics-facility/helpful-information.
Following transfection experiment, SQ20B cells were incubated with methionine-/cysteine-free media for 30 m. For labeling, 1 set of experimental cell plates were incubated for 30 m with media supplemented with 0.075 mCi/ml [35S]-methionine/cysteine, while a second set of plates were incubated with “cold” media supplemented with nonradioactive 1X methionine/cysteine. Protein was harvested from the “cold” samples, and a protein assay was performed to obtain concentrations. These proteins were analyzed using standard immunoblot assay conditions. Proteins collected from the [35S]-methionine/cysteine labeled cells were resolved on SDS-PAGE. The resulting gel was then fixed using a solution of 20% methanol and 10% acetic acid for 30 m, washed with deionized water, and then incubated on a rocker at room temperature with enlightening solution (PerkinElmer) for 15–30 m. The gel was then dried for 16 h using a Bio-Rad gel-drying apparatus and then exposed to autoradiography film at -80°C for 2 h before processing.
Polysome profiling was carried out per previously described conditions [74]. The HPLC ATP:AMP protocol was performed using previously established conditions determined by our group [75].
For ChIP experiments, SQ20B cells were plated 24 h before placing into hypoxia chamber or standard incubator. Six 10-cm plates at 70% confluency were grown in normoxic or hypoxic (0.5% O2) conditions for 16 h, at which time cells were crosslinked for 10 m at room temperature using 1% formaldehyde in minimum essential medium. Crosslinking was stopped and cells were washed and lysed according to manufacturer’s protocol (Active Motif ChIP kit #53035). Shearing conditions were carried out for six 20-s pulses at 25% power, with 30 s rest in between pulses on ice. Immunoprecipitations were carried out according to protocol, with 40 ug chromatin and 3 ug antibody for RNAPII (Active Motif #39097, mouse IgG #53010) and 60 ug chromatin with 3 ug (rabbit IgG Santa Cruz #sc-2027X) or 30 ug P-Ser2 RNAPII (Abcam #5095). For the resulting qPCR reactions, primers specific to EIF2B5 intron 10 and intron 12 were used (S1 Table), and control primers were purchased from Active Motif (GAPDH-2 for RNAPII control #71006, GAPDH-1 for P-Ser2 RNAPII control #71004, and Negative-1 for a 78-bp intergenic region of chromosome 12 as a negative control #71001).
SQ20B cells were grown in 2 x150-cm tissue culture plates for each condition (normoxia and a 0.5% hypoxia time point). Antibody against SRSF3 (SRp20) was purchased from MBL, Inc. (#RN080PW). Rabbit IgG was provided by the RIP kit for use as a control (RIP Assay kit, MBL, Inc. # RN1001). Protein A Sepharose CL-4B (GE Healthcare) were prepared fresh for the beginning of the experiment in a slurry of 75% beads in 25% 50 mM Tris pH 7.5. Beads were prewashed in PBS before adding 15 μg of antibody and then incubated at 4 degrees Celsius with rotation for 4 h. At this time, beads were prewashed before lysing cells and adding the lysates to the Protein A beads to preclear for 1 h at 4 degrees with rotation. Input samples were collected at this time before adding the precleared lysate to the antibody-bound beads. Samples were incubated for 3 h at 4 degrees with rotation. After wash steps, samples were aliquoted for quality control analysis of the protein. RNA was isolated according to the manufacturer’s protocol and analyzed using the Nanodrop1000 UV spectrometer. Equal amounts of RNA were then processed into cDNA and used for subsequent qPCR analysis. Input RNA was used as a standard to calculate the quantity for each primer. Expression data are reported as the relative enrichment of SRSF3 immunoprecipitation over the control Rabbit IgG immunoprecipitation.
All statistical analyses and drawings were done in R (version 3.2.5) (http://www.r-project.org/), and the statistical significance was defined as a P value <0.05.
For gene expression levels of the regions of interest, we downloaded RPKM data from TANRIC (http://ibl.mdanderson.org/tanric/_design/basic/index.html) [76]. The expression data is log2 transformed.
We restricted ourselves to the 8 types of cancers with available data for at least 30 normal samples. These are breast invasive carcinoma (BRCA), HNSC, KIRC, kidney renal papillary cell carcinoma (KIRP), liver hepatocellular carcinoma (LIHC), lung adenocarcinoma (LUAD), prostate adenocarcinoma (PRAD), and thyroid carcinoma (THCA). We downloaded patient clinical information for the patients of these cohorts from cbioPortal (http://www.cbioportal.org/).
To be able to determine the expression difference for the regions of interest between normal and tumor tissue, respectively, among normal and tumor tissue separated according to the cancer stage, we first employed a Shapiro-Wilk test to verify if the data follows a normal distribution. Accordingly, t test, respectively ANOVA test (depending on the number of groups considered), or the nonparametric test Mann-Whitney-Wilcoxon, respectively Kruskal-Wallis test, was applied to assess the relationship between mRNA expression and tissue type.
A box-and-whisker plot (Box plot represents first [lower bound] and third [upper bound] quartiles, whiskers represent 1.5 times the interquartile range) was used to visualize the data.
|
10.1371/journal.pgen.1004828 | Epigenome-Guided Analysis of the Transcriptome of Plaque Macrophages during Atherosclerosis Regression Reveals Activation of the Wnt Signaling Pathway | We report the first systems biology investigation of regulators controlling arterial plaque macrophage transcriptional changes in response to lipid lowering in vivo in two distinct mouse models of atherosclerosis regression. Transcriptome measurements from plaque macrophages from the Reversa mouse were integrated with measurements from an aortic transplant-based mouse model of plaque regression. Functional relevance of the genes detected as differentially expressed in plaque macrophages in response to lipid lowering in vivo was assessed through analysis of gene functional annotations, overlap with in vitro foam cell studies, and overlap of associated eQTLs with human atherosclerosis/CAD risk SNPs. To identify transcription factors that control plaque macrophage responses to lipid lowering in vivo, we used an integrative strategy – leveraging macrophage epigenomic measurements – to detect enrichment of transcription factor binding sites upstream of genes that are differentially expressed in plaque macrophages during regression. The integrated analysis uncovered eight transcription factor binding site elements that were statistically overrepresented within the 5′ regulatory regions of genes that were upregulated in plaque macrophages in the Reversa model under maximal regression conditions and within the 5′ regulatory regions of genes that were upregulated in the aortic transplant model during regression. Of these, the TCF/LEF binding site was present in promoters of upregulated genes related to cell motility, suggesting that the canonical Wnt signaling pathway may be activated in plaque macrophages during regression. We validated this network-based prediction by demonstrating that β-catenin expression is higher in regressing (vs. control group) plaques in both regression models, and we further demonstrated that stimulation of canonical Wnt signaling increases macrophage migration in vitro. These results suggest involvement of canonical Wnt signaling in macrophage emigration from the plaque during lipid lowering-induced regression, and they illustrate the discovery potential of an epigenome-guided, systems approach to understanding atherosclerosis regression.
| Atherosclerosis, a progressive accumulation of lipid-rich plaque within arteries, is an inflammatory disease in which the response of macrophages (a key cell type of the innate immune system) to plasma lipoproteins plays a central role. In humans, the goal of significantly reducing already-established plaque through drug treatments, including statins, remains elusive. In mice, atherosclerosis can be reversed by experimental manipulations that lower circulating lipid levels. A common feature of many regression models is that macrophages transition to a less inflammatory state and emigrate from the plaque. While the molecular regulators that control these responses are largely unknown, we hypothesized that by integrating global measurements of macrophage gene expression in regressing plaques with measurements of the macrophage chromatin landscape, we could identify key molecules that control macrophage responses to the lowering of circulating lipid levels. Our systems biology analysis of plaque macrophages yielded a network in which the Wnt signaling pathway emerged as a candidate upstream regulator. Wnt signaling is known to affect both inflammation and the ability of macrophages to migrate from one location to another, and our targeted validation studies provide evidence that Wnt signaling is increased in plaque macrophages during regression. Our findings both demonstrate the power of a systems approach to uncover candidate regulators of regression and to identify a potential new therapeutic target.
| Underlying many cardiovascular diseases is atherosclerosis, an arterial accumulation of lipid-laden macrophages (foam cells) with attendant chronic inflammation [1]. Atherosclerosis is endemic in adults [2] and can progress over decades before presenting with overt symptoms [3]. Atherosclerosis is correlated with risk factors such as hyperlipidemia, especially high plasma levels of low-density lipoprotein (LDL) cholesterol [4], [5]. There is significant interest in finding new therapies to promote the regression of atherosclerotic plaques [6]–[8], with particular emphasis on the plaque macrophage as a potential target because of its roles in lipoprotein uptake and vascular inflammation [9]–[12], plaque destabilization [13], [14], and plaque remodeling in response to dietary changes [15].
In mice, plaque regression has been achieved through expression [16]–[18] or inactivation [19], [20] of lipid metabolism-modifying genes; aortic tissue grafting [21]–[23]; and treatment with small-molecule therapeutics [24]–[26] or biologics [27], [28], frequently in conjunction with a diet alteration. We have been studying the role of macrophages during plaque regression in two mouse models of regression, the Reversa mouse [19], [20], [29], [30] and an aortic arch transplant model [21], [22], [31]. Reversa mice are Ldlr−/− Apob100/100 [32], [33] and they possess a loxP-flanked allele of the gene encoding microsomal triglyceride transfer protein (Mttp) and a drug-inducible Cre recombinase (Cre) transgene [19], [34]. In the Reversa mouse, hyperlipidemia can be reversed by the transient induction of Cre expression in the liver, which causes recombinant knockout of Mttp, and thus significantly lowers the plasma levels of very low density lipoprotein (VLDL) and LDL [19]. In prior work we have shown that inactivation of Mttp in the Reversa mouse, when combined with a switch from Western diet to chow, leads to plaque regression at the aortic root [20]. In the aortic transplant model, an atherosclerotic aortic arch from an Apoe−/− mouse [35] is grafted in place of a segment of the abdominal aorta [31] of either a wild-type (WT) or an Apoe−/− mouse. Grafts within WT recipients exhibit plaque regression, while plaques within Apoe−/ recipients continue to progress [22]. A common feature of the Reversa and transplant regression models is the substantial reduction in plaque macrophages during regression [20], [23], [29], [30], [36]. It is at present unknown whether there are molecular pathways that may initiate the depletion of plaque macrophages that are shared across regression models. Identifying such core pathways could yield new therapeutic approaches to stimulate regression and/or beneficial remodeling of plaque.
Recently, we studied transcriptome differences between macrophages in regressing versus progressing plaques in the aortic transplant model [36]. We observed patterns of differential expression that suggest that macrophages in regressing plaques up-regulate genes associated with cell motility, down-regulate genes associated with cell adhesion, and up-regulate genes associated with an anti-inflammatory, “M2 macrophage” [37] phenotype. Although our analysis pointed to several candidate molecular regulators, it is unknown whether there are transcription factors that are common to the transplant and Reversa regression models that act as master controllers for the responses of plaque macrophages to lipid lowering.
In the current study we carried out a systems investigation of the Reversa model, with three goals: (i) to identify pathways and gene functions that are associated with plaque macrophage responses to lipid lowering in vivo; (ii) to understand the connections between gene sets identified in our study and gene sets from previous studies of macrophage foam cell formation, aortic wall changes during plaque progression and regression, and atherosclerosis risk alleles in humans; and (iii) to identify transcriptional regulators that are associated with macrophage responses to lipid lowering in vivo. We therefore carried out microarray transcriptome profiling of plaque macrophages isolated from the aortic root of Reversa mice undergoing lipid lowering in vivo and analyzed differentially expressed genes using multiple bioinformatics approaches. Gene functional enrichment analysis (i) detected over-representation of cytoskeletal-binding and Rho GTPase genes among genes that are upregulated in response to lipid lowering, pointing to cytoskeletal reorganization during regression. Comparative analysis (ii) of gene sets from our study with those from previous studies of in vitro models of macrophage foam cell formation revealed a statistically significant overlap. Additionally, we found a significant overlap between human monocyte expression quantitative trait loci (eQTLs) for human orthologs of genes that are differentially expressed during plaque regression, and genetic loci that are associated with risk of atherosclerosis or coronary artery disease. Promoters of genes identified in our study were analyzed for transcription factor binding site over-representation (iii) using a novel chromatin-guided method, REMINISCE (Figure S1 and Methods). REMINISCE leverages macrophage epigenomic and chromatin measurements, such as histone acetylation [38], [39] and DNase I hypersensitive sites [40], in the analysis of the 5′ regulatory regions of differentially expressed genes, including enhancers as well as promoters. In both the Reversa and aortic transplant regression models, we found that the consensus binding site sequence for the T-cell specific, HMG-box factors (TCF) and lymphoid enhancer factors (LEF), together known as the TCF/LEF family of transcription factors, was overrepresented within the 5′ regulatory regions of genes that are upregulated in plaque macrophages during regression.
TCF/LEF transcription factors are activated by nuclear β-catenin (CTNNB1) that accumulates in response to activation of the canonical Wnt signaling pathway [41], [42]. The Wnt pathway controls multiple functions in development, tissue organization, cell proliferation, cell migration [42], and inflammation [43]. In macrophages, canonical Wnt signaling is thought to promote cell motility through a β-catenin-dependent mechanism [44]. In this study we show, for the first time, that activation of canonical Wnt signaling, defined by up-regulation of β-catenin, occurs within macrophage-rich plaque in two mechanistically distinct lipid-lowering models of plaque regression, the Reversa and aortic transplant models.
Reversa mice were maintained on Western diet for 16 weeks, which allowed for the development of hypercholesterolemia (903±71 mg/dL total cholesterol [mean ± standard error, SE]) and substantial, macrophage-rich plaques at the aortic root, as determined by immunostaining with the macrophage marker Cluster of Differentiation 68 (CD68). At 16 weeks, a group of animals were sacrificed to measure the baseline plasma total cholesterol levels and CD68+ areas within plaque within aortic root sections. The remaining animals were switched to chow diet and divided into two groups, an experimental group that received four injections of polyinosinic:polycytidylic acid (poly I:C) to induce Mx1:Cre in order to inactivate Mttpfl/fl, and a control group that was injected with vehicle (saline) on the same schedule. On day seven or 14 after the final injection, animals were sacrificed and plasma cholesterol and plaque CD68+ areas were measured. By day 14, cholesterol levels had decreased by 88% in the Mttp-inactivated animals vs. baseline (Figure 1A and Table 1, P<0.001) and plaque CD68+ area had decreased by 36% in the Mttp-inactivated animals vs. baseline (Figure 1B and Table 2, P<0.05). Additionally, plaque CD68+ area was reduced in Mttp-inactivated vs. vehicle-treated animals at seven and 14 days (P<0.05); at day 14, CD68+ area was 26% smaller in the Mttp-inactivated animals to the vehicle-treated animals (Figure 1C). These results indicated that there is plaque regression that is attributable to inactivation of Mttp. In contrast, comparisons of the saline-treated sample groups to baseline showed no statistically significant differences in CD68+ area that could be attributed to diet switch alone.
In order to focus on the molecular signaling events that initiate the reduction in plaque macrophages, we selected an early time point (day seven) when there was a statistically significant difference in the CD68+ area in the treatment group vs. vehicle (Figure 1, B and C). CD68+ plaque macrophages were isolated from aortic root sections from Mttp-inactivated Reversa, vehicle-treated Reversa, and poly I:C-injected Ldlr−/− animals (Table S1) by laser capture microdissection (LCM) [45], [46], a procedure that we have previously demonstrated specifically enriches for the mRNAs of macrophage-related genes by more than 30-fold over levels seen in aortic homogenates [45]. Global transcriptome profiles were then obtained from the LCM-derived samples using high-density oligonucleotide microarray hybridization. A total of 213 genes were detected as differentially expressed in plaque macrophages from Mttp-inactivated vs. vehicle-treated animals (93 had higher expression levels in Mttp-inactivated than in control, and 120 had lower levels; Table S2). In order to control for any effect of the poly I:C injections on the macrophage transcriptional profile we compared the transcriptomes of plaque CD68+ cells from saline-treated Reversa mice with those from poly I:C-treated Ldlr−/− mice. We found that only one of the 213 lipid-responsive genes, Trnt1, was detected as differentially expressed in the Reversa saline vs. Ldlr−/− poly I:C comparison, and this gene was excluded from further bioinformatic analysis. To further substantiate that the 213 genes in plaque CD68+ cells are responding to lipid lowering rather than directly to poly I:C, the list of 213 genes was compared with a list of genes that are differentially expressed (vs. vehicle) in bone marrow-derived macrophages stimulated in vitro with poly I:C (Table S3) to determine whether the degree of overlap in the two gene lists is greater than would be expected by chance. The analysis revealed that there is not more overlap with in vitro poly I:C responses than would be expected by chance (P>0.1, Fisher's Exact Test).
To gain insight into the biological processes that occur during lipid lowering in vivo, we performed gene functional annotation enrichment analysis. The most significantly overrepresented processes associated with the upregulated genes are related to cell motility, cytoskeletal binding, and Rho GTPases (Figure 2A and Table S4). For the genes that were downregulated in macrophages from lipid-lowered vs. control animals, the gene annotation term that was the most significantly overrepresented was mitochondrion (Figure 2B and Table S4). In order to determine whether lipid lowering in vivo might have increased plaque macrophage apoptosis, multiple gene functional enrichment analysis tools were applied, and in each case, no finding of a statistical enrichment for apoptosis-related gene annotations was detected (see Methods).
We hypothesized that among the plaque macrophage genes that are responsive to lipid lowering in vivo, genes that had been previously reported to be differentially expressed in response to lipid loading in vitro would be statistically overrepresented. We therefore combined differentially expressed gene lists from 21 datasets from 15 studies in the atherosclerosis and macrophage foam cell literature (Table S5), and scored genes based on the number of studies in which they were detected as differentially expressed during foam cell formation or disease progression (similar to the approach that was used in a recent meta-analysis to identify obesity-related genes [47]). We found that 19 out of the 213 differentially expressed genes in our study were also detected among the high-scoring genes in the in vitro lipid-loaded macrophage meta-analysis gene set, corresponding to a 2.9-fold enrichment (P<0.0001) vs. for all mouse genes.
Because in vivo lipid lowering is associated with plaque regression in the Reversa and aortic transplant models (Figure 1 and [19], [20], [23], [36]), we conjectured that the gene expression changes in plaque CD68+ cells in response to lipid lowering would be associated with atheroprotective effects in humans. Because such effects might be subtle for any individual gene, and because “master regulators” of the transcriptional response are often not transcriptionally regulated themselves, we tested our hypothesis by computationally searching for connections between human atherosclerosis/CAD risk-associated single nucleotide polymorphisms (SNPs) and monocyte eQTLs for orthologs of genes that are differentially expressed in plaque CD68+ cells in response to lipid lowering (as had been previously applied to a transcriptome study of whole aortic arches [30]). We assumed that if there are no atheroprotective effects associated with the observed gene expression changes in CD68+ cells, then the fraction of atherosclerosis/CAD risk SNPs within monocyte eQTLs for the 213 lipid lowering-responsive genes should be the same as the fraction of atherosclerosis/CAD SNPs within all monocyte eQTLs. Using a list of human eQTLs from two population genetic studies of peripheral blood monocyte gene expression [48], [49], we identified SNPs associated with mRNA levels (eSNPs) of genes whose mouse orthologs are differentially expressed in CD68+ cells during regression in the two mouse models. We expanded the regression-gene-set-associated eSNPs to include all validated SNPs in linkage disequilibrium (using haplotype information from the 1,000 Genomes Project [50]), obtaining a set of regression-associated monocyte eQTLs. We then assessed the frequency of atherosclerosis/CAD risk SNPs (starting with a list of 761 risk SNPs obtained from the NCBI PheGenI database [51]) within these regression-associated monocyte eQTLs and compared these frequencies to the frequency of atherosclerosis/CAD risk SNPs within monocyte eQTLs in general. We found that the regression-associated eQTLs are substantially more enriched for atherosclerosis/CAD risk SNPs than monocyte eQTLs in general (Figure 2C and Table S6), in both regression models (P<0.05 for each model); overall, a total of 17 atherosclerosis/CAD risk-associated SNPs were found within regression-associated eQTLs (Table S7). Further investigation revealed that one of the Reversa regression-associated eQTLs encompasses rs599839, a SNP in the 1p13.3 locus that has been found to be associated with LDL cholesterol (LDL-C) levels and with risk of CAD and MI [52]–[55], as well as with hepatic mRNA expression of the intracellular lipoprotein sorting receptor Sortilin 1 (SORT1) [56], [57]. In human monocytes [48] and in liver cells [56], variant rs599839 is also an eSNP for mRNA expression of the gene Proline/serine-rich coiled coil 1 (PSRC1), which is one of most strongly upregulated genes in CD68+ cells in Mttp-inactivated vs. vehicle-treated Reversa mice (Table S2).
We applied our novel promoter scanning analysis method, REMINISCE (Figure S1), to identify candidate transcription factors that are associated with the transcriptional responses of plaque macrophages to lipid lowering in vivo. REMINISCE combines 17 different types of macrophage chromatin and epigenomic measurements (see Table S8) that were obtained from NCBI GEO and from the Mouse ENCODE project [58] in a machine-learning algorithm (using genome regions that coincide with transcription factor binding sites that are obtained from ChIP-seq as a training dataset of regulatory regions) to map macrophage cis-regulatory regions, and then scans genomic sequences within these regions to identify statistically overrepresented transcription factor binding site sequence patterns (see Methods). In a direct comparison of REMINISCE with non-epigenome-guided motif-scanning of all noncoding sequence within ±1 kbp or ±5 kbp of the transcription start site for gene sets derived from macrophage transcriptional responses to the atherogenic lipoprotein oxidized LDL, REMINISCE detected substantially more binding site motifs above the significance threshold (Figure S2). Applying REMINISCE to the 213 lipid lowering-responsive genes from the Reversa model, we identified 15 transcription factor binding site sequence motifs that are statistically overrepresented within the 5′ regulatory regions of upregulated genes (Figure S3A), and 14 motifs that were associated with the downregulated genes (Figure S3B). Because any experimental model is subject to model-specific artifacts, we also analyzed our previously published data on the transcriptional response in CD68+ cells isolated from regressing plaques in the aortic transplant model using the REMINISCE method. We found that of the 15 motifs that were overrepresented among upregulated genes in the Reversa model, eight were also overrepresented among upregulated genes in the aortic transplant model (Figure 3A), including the motifs for PPARγ (nuclear receptor direct repeat 1, or NR DR1), the TCF/LEF family of transcription factors, and activating protein 1 (AP-1). The transcription factors represented by the 15 motifs that were associated with upregulated genes in the Reversa model are highly interconnected in the human protein-protein interactome [59] (clustering coefficient of 0.31, vs. 0.14 for the whole interactome [60]) (Figure 3B). Similarly to the case with transcription factors for upregulated genes, of the 14 motifs that were overrepresented among downregulated genes in the Reversa model, six were also detected in the aortic transplant model (Figure 3C and Table S9).
Under the hypothesis that the transcription factors that control macrophage responses to lipid lowering in vivo would have binding site sequences within the 5′ regulatory regions of a high proportion of differentially expressed genes, we ranked the motifs from Figure 3 by the fraction of the differentially expressed genes that possess a match for each motif (Table S10). Notably, for the upregulated genes, the motif whose matches were present within the largest percentage of genes, 33%, was the CTTTGA motif that is recognized by TCF/LEF and SOX transcription factor families. The finding of enrichment of TCF/LEF binding site sequences within 5′ regulatory regions of genes that are upregulated in the Reversa mouse was confirmed by additional computational analysis using high-precision motifs based on three-dimensional structural modeling of protein-DNA interactions [61] (Table S9). Because the reduced macrophage content of the plaque in the transplant regression model has been associated with macrophage emigration from the plaque, we also ranked transcription factor binding site motifs based on the proportion of cytoskeletal-binding and RhoGAP domain-containing genes (Table S4) whose 5′ regulatory regions contain matches for each motif. We found that the motif with the highest percentage, 50%, was the CTTTGA motif (corresponding to TCF/LEF and SOX) (Table S10). Next, we used the global mammalian protein interaction network to examine which signaling pathways could connect between lipoproteins or cholesterol and these transcription factors. We found that the Wnt signaling pathway [62], [63], through β-catenin and through TCF/LEF and SOX family members [64], satisfies these criteria (Figure 4).
Gene regulation likely occurs at multiple levels in plaque macrophages. Therefore, we scanned the 3′ untranslated regions (UTRs) of the two sets of differentially expressed genes in plaque CD68+ cells in the Reversa model for microRNA target sequences. We identified eighteen microRNAs whose target sequences were overrepresented within the 3′ UTRs of the upregulated and downregulated genes (six for upregulated genes, and twelve for downregulated) (Figure S4, Table S11). Under the hypothesis that some of these microRNAs might work in concert in response to lipid lowering in vivo, we analyzed molecular pathways for enrichment of the number of pathway-associated genes that possess target sequences for any of the microRNAs in Figure S4A, using a microRNA pathway analysis tool [65]. For the upregulated genes, the two pathways with the strongest enrichment were axon guidance (P<10−4) and focal adhesion (P<0.001); also enriched was the Wnt signaling pathway (P<0.05) (Table S12).
β-catenin-dependent activation of TCF/LEF is the final step in the canonical Wnt signaling pathway. To investigate the possibility of canonical Wnt pathway activation during plaque regression, we measured the relative mRNA level of β-catenin (Ctnnb1) in CD68+ cells isolated from aortic grafts from WT and Apoe−/− recipient mice in the aortic transplant regression model. By quantitative PCR (qPCR), we found that the relative level of Ctnnb1 mRNA is approximately 2.2-fold higher in CD68+ cells in regressing plaques than in progressing plaques (Figure 5A) (P<0.05). Additionally, we measured mRNA levels of two known Wnt target genes, Lrp6 (for which loss-of-function has been associated with premature CAD [66]) and Gja1 [67], both of which were indicated by microarray profiling to be upregulated in regressing vs. progressing plaque CD68+ cells in the aortic transplant model [36]. By qPCR, both Lrp6 and Gja1 were found to be upregulated in CD68+ cells in regressing vs. progressing plaques (Figure S5). Given a previous report that β-catenin deficiency impairs macrophage migration [44] and in light of evidence of plaque macrophage emigration during regression in both the Reversa [20] and aortic transplant [22] models, we investigated whether canonical Wnt pathway activation stimulates macrophage migration in vitro. In a scratch-wound cell migration assay [68], [69], primary murine macrophages treated with the canonical Wnt pathway agonist Wnt3a scored more than two-fold higher than vehicle-treated macrophages (Figure S6).
To confirm the prediction that Wnt signaling is activated in plaque macrophages during regression in vivo, we analyzed the expression of β-catenin in plaque within aortic root sections from Reversa mice at day seven following Mttp-inactivation or vehicle treatment, by fluorescence immunohistochemistry. Consistent with the systems biological analyses, the level of β-catenin was significantly greater (approximately 2.6-fold, P<0.001) in the plaques isolated from the Mttp-inactivated vs. vehicle-treated animals (Figure 5B, 5C). Because the TCF/LEF binding site was detected as being overrepresented among genes that are upregulated during plaque regression in the transplant regression model as well, we also analyzed β-catenin expression in plaques in the aortic arch graft of WT and Apoe−/− recipient animals. As with the Reversa model, we found that β-catenin is expressed at an approximately 2.5-fold higher level (P<0.0001) in regressing vs. progressing plaques (Figure 5D, Figure 5E, Figure S7), with increased abundance in the nuclear area (Figure S8).
It is clear that atherosclerosis regression involves the interactions of many cellular pathways and multiple cell types and plaque components. To address this complexity, a systems biology approach employing large-scale molecular profiling is needed to identify core, model-independent mechanisms [8], [10], [70], [71]. To our knowledge, the present work, along with complementary measurements obtained from the aortic transplant regression model [36], represent the first reports of transcriptome measurements specifically of CD68+ macrophages within regressing plaques. In a previous transcriptome study [29] of changes in vessel wall tissue after lipid lowering in vivo in the Reversa mouse, a modest number of genes were detected as differentially expressed in the aortic wall (37 genes in [29], and 42 genes for arteries containing early lesions in [30]). Notably, none of these was detected as differentially expressed in our study of responses of plaque macrophages to lipid lowering, pointing to the need to separately measure the gene expression changes of constituent cell types of the plaque [71], as we have done for macrophages in the present work. While the 213 genes we identified as differentially expressed in plaque CD68+ cells in the Reversa mouse in response to lipid lowering are likely only a subset of the transcriptional changes in plaque macrophages over various time points, the enrichment of atherosclerosis/CAD risk alleles within human monocyte expression QTLs for orthologs of these genes (Figure 2C) provides intriguing support for their collective function in macrophages in the context of atherosclerosis.
Given the central role of macrophages in vascular inflammation [72] and in plaque destabilization in humans [73]–[75], our finding that Mttp-inactivation in the Reversa mouse reduces the macrophage content of the plaque (Figure 1 and [20]) points to the potential for beneficial remodeling. However, it also raises the question of how much of this reduction is driven by three processes (all of which have been reported in different models of regression): emigration from the plaque [22], deceased infiltration of monocytes [17], and increased macrophage cell death [25]. To the extent that accumulation of apoptotic macrophages within shoulder regions of advanced plaque is considered a marker of vulnerable plaques [76], the balance of these mechanisms is potentially clinically significant.
In the aortic transplant regression model, plaque macrophages in WT recipients rapidly emigrate from the graft to regional lymph nodes [22], while remaining macrophages do not exhibit abnormal levels of apoptosis [22]. Measurements of macrophage transcriptional changes during plaque regression versus progression in the aortic transplant model are consistent with an induced emigration hypothesis; macrophages in regressing (vs. progressing) plaques had lower expression levels of genes associated with adhesion and higher expression levels of genes associated with cell motility [36]. In regards to the emigration hypothesis in the Reversa model (which is supported by macrophage trafficking studies in vivo [20]), the plaque macrophage transcriptome measurements in this work provide several key insights. First, among genes that are upregulated in macrophages from Mttp-inactivated versus control animals, genes encoding cytoskeletal-binding proteins and Rho GTPases are overrepresented, consistent with increased motility. Second, among genes that are differentially expressed, no overrepresentation of apoptosis-related genes was found at the time point that we examined. A third insight, involving the Wnt signaling pathway, comes from the transcription factor analysis as described below.
In this study, we used an integrative computational approach, REMINISCE, to identify transcription factors that are likely to mediate macrophage responses during plaque regression. A novel aspect of the approach is that it incorporates macrophage epigenome measurements to guide the search for transcription factor binding sites (resulting in improved sensitivity; Figure S2) in the context of an enrichment analysis. In the analysis, we made use of transcriptome measurements from both the aortic transplant and Reversa regression models as well as macrophage-specific chromatin measurements to detect overrepresentation of known transcription factor binding site sequences upstream of genes that are differentially expressed during plaque regression. To focus on transcription factors that may have a model-independent role in macrophage responses to lipid lowering, we selected factors whose binding site sequences were overrepresented in both regression models. The significant overlap between the sets of transcription factors that were identified from the two models was unexpected due to the distinct differences in genetic backgrounds, plaque locations, and regression time-scales of the models. Of these factors, the association of PPARγ with genes that are upregulated during plaque regression is consistent with one of the principal findings from a previous transcriptome profiling study of whole aortic arches in the Reversa mouse with early lesions [30]. While our previous work investigating the effect of treatment of Reversa mice with a PPARγ agonist did not observe a significant effect of the agonist on the abundance of macrophages in the plaque [20], PPARγ is certainly worthy of future studies to clarify its precise function in plaque regression (especially given conflicting findings in different models and disease stages [20], [77]–[79]).
Among the other transcription factors whose binding site motifs were found to be associated with genes that are upregulated in CD68+ cells in regressing plaques in both regression models, the TCF/LEF and SOX transcription factor families (whose binding site sequences were associated with genes that are upregulated in response to lipid lowering) were particularly notable because of their connection to Wnt signaling, and thus, cell motility. Additionally, pathway enrichment analysis based on genes with target sequences for the microRNAs that were associated with upregulation in plaque CD68+ cells identified axon guidance and Wnt signaling as enriched pathways. Based on the transcription factor binding site analysis, we hypothesized that canonical Wnt signaling may be increased in plaque macrophages during regression. This hypothesis was confirmed by multiple observations including Ctnnb1 mRNA measurements in CD68+ cells, mRNA measurements of Wnt pathway target genes, and immunohistochemical analysis demonstrating increased β-catenin in regressing plaques in the Reversa and aortic transplant mouse models. Although the molecular mechanisms that cause increased Wnt signaling in plaque in response to lipid lowering are not known, and, of course, in spite of our associative findings, it is possible that plaque regression is mediated by Wnt-independent pathways, nonetheless, the observation of increased Wnt signaling in response to lipid lowering is consistent with previous reports that, in cultured cells, depletion of cholesterol increases canonical Wnt signaling [63].
In both the Reversa and aortic transplant models of regression, lipid lowering in vivo leads to decreased macrophage content in plaque (Figure 1B, [20], [23], [28]–[30], [36]), and trafficking studies in both of these models suggest that lipid lowering triggers macrophage emigration from the plaque [20], [22]. Our own functional studies of macrophage responses to Wnt3a stimulation (Figure S6) as well as previous studies of the role of β-catenin in macrophage migration [44], [80], are consistent with the model that increased canonical Wnt signaling – resulting in increased β-catenin levels – stimulates macrophage motility. A transitioning of macrophages from a sessile to a motile state would be expected to facilitate macrophage egress from the plaque [81].
This work is the first to our knowledge to propose that canonical Wnt signaling is involved in atherosclerosis regression in response to lipid lowering in vivo, and adds to several lines of evidence that support that alterations in Wnt signaling play a role in the progression of atherosclerosis. LRP6 genetic variants that impair Wnt/β-catenin signaling have been reported that are associated with increased risk of carotid atherosclerosis [82] and early CAD [66] in humans, suggesting that decreased canonical Wnt/β-catenin signaling may be pro-atherogenic; conversely, by this reasoning, increased Wnt signaling would be expected to have protective, pro-regressive effects. Consistent with this model, the canonical Wnt signaling molecule Wnt3a has been demonstrated to mediate anti-inflammatory effects via β-catenin-dependent suppression of the expression of proinflammatory cytokines [83]. Canonical Wnt signaling is also thought to play an important role in maintaining vascular cellular homeostasis, for example, in orchestrating lineage commitment within pericytes in the vessel wall [84]. Furthermore, Wnt signaling has been shown to potently inhibit lipid accumulation in cultured cells and to regulate plasma cholesterol levels in vivo [85]–[87], as well as (noted before) to promote macrophage migration ([44] and Figure S6).
While the present work provides evidence implicating β-catenin-interacting transcription factors in atherosclerosis regression, it is highly probable that other transcription factors are involved, given the complexity of macrophage transcriptional regulation in response to changing lipoprotein levels [70]. Of particular interest for future studies are the transcription factors that have overrepresented binding site sequences within the upregulated gene group, and that are connected to the TCF/LEF family members in the protein interaction network (Figure 3B). It should also be noted that a potential limitation of the analytical approach used in this work is the partial reliance on proximity to key factors in the global protein interaction network, in prioritizing transcriptional regulators for targeted validation; the use of such a proximity criterion could introduce bias against transcription factors that are less well-studied.
While this work used curated chromatin measurements from primary macrophages to guide the bioinformatic search for transcription factor binding sites, the numbers of cultured cells that are required for a ChIP-seq assay is rapidly being reduced; continued assay improvements may make possible genome-wide location analysis of chromatin factors within specific cellular constituents of plaque, which would likely enable more precise mapping of regulatory elements that mediate transcriptional changes during plaque regression. Equally important, comprehensive datasets of genetic variants and their associations with cardiovascular phenotypes (which may achieve significance only in a meta-analysis), such as those provided by the CARDIoGRAMplusC4D and dbGAP databases, will undoubtedly be beneficial for contextualizing the functions of these regulatory elements in atherosclerosis. Another enhancement in future studies would be to use lineage-tracing techniques in vivo to definitively characterize the extent of cellular phenotypic conversion in the Reversa model at the aortic root. Our previous work, however, has demonstrated that CD68-guided LCM of aortic root plaque yields a highly macrophage-enriched cell population [45]. Furthermore, no transcriptional changes were observed during plaque regression that would suggest a change in the proportion of macrophages in the cells captured by CD68-guided LCM in regressing vs. progressing plaques.
Beyond the enrichment of TCF/LEF transcription factor binding sites, the finding of enrichment of binding sites for the oxidative stress-responsive transcription factor nuclear respiratory factor 2 (NRF-2) within the regulatory regions of the downregulated gene group (Figure S3) is interesting in light of the observed down-regulation of mitochondrial transcription factor A (Tfam, a regulator of mitochondrial biogenesis and a known NRF target gene [88]). Downregulation of NRF-2 activity might reflect decreased oxidative stress in plaque macrophages due to the substantially reduced plasma lipoprotein levels [89].
Another intriguing finding from our work is the higher level of expression of Psrc1 in plaque macrophages in the Reversa regression model. Psrc1 mRNA has been previously been reported to be elevated in macrophage-rich regions of human atheroma [90]. The function of PSRC1 in macrophages is not known, but in other cell types it stabilizes microtubules [91], [92] and functions in cell migration [92]. Given our findings of elevated β-catenin in regressing plaques, it is interesting that a study of PSRC1 function found that PSRC1 stimulates the β-catenin pathway [93]. While strong evidence indicates that the effect of SNP rs599839 on LDL-C is mediated by hepatic expression of Sortilin 1 (SORT1) [56], the link between rs599839 and monocyte expression of PSRC1 seems worthy of further study in the context of plaque regression.
The multi-model analyses of plaque regression we have carried out is likely to ultimately shed light on the process of plaque stabilization in humans, given that atherosclerosis in genetically modified mice recapitulates many features of clinical disease. Large clinical trials have shown that in humans treated with lipid-lowering drugs, plaque area, as assessed by intravascular ultrasound (IVUS), changes very little [94], [95], while atherothrombotic event rates are dramatically decreased [96], [97], presumably due to stabilization resulting from changes in plaque macrophages not detected by IVUS. Indeed, even in mouse models, we have found that plaque size measurements do not always reflect changes in plaque macrophage content because of increased matrix proteins (e.g., [98]). Demonstrating similar phenomena in human plaques awaits improvements in non-invasive imaging. Overall, in human plaques, a drastic reduction in the abundance of macrophages within the plaque would be presumed to be associated with protective remodeling of the atheroma, and given the similarities in the disease process between mice and people, some of the factors and pathways we have identified are likely to be clinically relevant (e.g., Wnt signaling [66]).
In conclusion, this work demonstrates the value of a genomic, systems biology approach for elucidating the cellular pathways and transcriptional regulatory mechanisms that are dysregulated or altered by perturbations in vivo. Although the present study is focused on macrophage responses to lipid lowering in vivo, our approach of integrating LCM-derived transcriptome measurements with epigenomic and chromatin structural information from primary cell models could be applied to other constituents of the plaque (e.g., smooth muscle cells and endothelial cells) and then ultimately combined in a comprehensive regulatory network. More broadly, with today's panoply of publicly available epigenomic and chromatin structural measurements from a wide array of mouse and human cell types and tissues (e.g., through the ENCODE project), we anticipate that integrative approaches such as the one used in this work could be useful in the regulatory analysis of transcriptome measurements from unique, highly specific cell populations in other in vivo contexts.
This study was carried out in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals (National Institutes of Health), and in accordance with animal use protocols that were approved by the NYU Langone Medical Center Institutional Animal Care and Use Committee (PHS Animal Welfare Assurance Number A3435-01) or the Institute for Systems Biology Institutional Animal Care and Use Committee (PHS Animal Welfare Assurance Number A4355-1). All tissue harvest (except mouse peritoneal macrophage harvest) and all survival surgeries were carried out under ketamine/xylazine-induced anesthesia, and for survival surgeries, animals were carefully monitored post-surgery for signs of pain or distress. For peritoneal macrophage harvest, mice were euthanized by CO2 asphyxiation following standard operating procedures established by the Institute for Systems Biology IACUC.
The development of the Reversa mouse (JAX Stock Number 004192, Ldlr−/− Apob100/100 Mttpfl/fl Mx1:Cre+/+ animals on mixed background, in which injection of poly I:C causes silencing of hepatic Mttp expression) has been previously described [19]. In this work, for experimental studies of the Reversa model of plaque regression, four-week-old Reversa mice were weaned onto a Western diet (Catalog no. 100244, 40% kcal from milk fat; Dyets Inc., Bethlehem, PA) for 16 weeks (baseline time point) to establish aortic plaque and then switched to a chow diet and injected intraperitoneally every other day, for a total of four injections, with either 500 µg poly I:C (Sigma) or vehicle (saline). Blood samples were obtained from the retro-orbital plexus at baseline, seven days after the last poly I:C injection, and 14 days post-injection. Animals were sacrificed for arterial tissue collection at either baseline, seven days after the last poly I:C injection, or 14 days post-injection, as indicated in the Results section.
The aortic transplant model of plaque regression has been previously described [21], [23], [31]. In this work, for studies of plaque regression in the aortic transplant model, C57BL/6–Apoe−/− mice [35] were maintained on Western diet for 16 weeks and then the aortic arches were harvested and grafted into the abdominal aortas of Apoe−/− and WT mice. The grafts were harvested at three or five days post-surgery, as indicated in the Results section.
For all experimental data reported in this work, animal cohort sizes are as reported in the corresponding figure captions, table captions, or main article text, except for the microarray transcriptome profiling of plaque macrophages in the Reversa model, for which the cohort sizes are described in Table S1.
All cell culture was performed at 37°C and in the presence of 5% CO2.
Bone marrow-derived macrophages (BMDM) were prepared as follows: L929-conditioned medium was prepared by growing L929 cells to confluency in Dulbecco's Modified Eagle Medium (DMEM)+10% (v/v) fetal bovine serum (FBS)+100 I.U./mL penicillin and 100 µg/mL streptomycin (hereafter, 1% penicillin/streptomycin). The media was then replaced with DMEM containing 2% FBS+1% penicillin/streptomycin. Conditioned media was collected every three days, filtered through a 0.22 µm filter and fresh media containing 2% FBS was added. Bone marrow cells were isolated from the femur and tibia of 8–12 week old mice. After red blood cell lysis (RLB, Sigma), cells were plated in Petri dishes in L929-conditioned medium and incubated for seven days. Cells were then serum-starved in 0.2% FBS overnight and treated with 400 ng/mL Wnt3a (R&D Systems, Minneapolis, MN) for 24 h. RNA was isolated from cells using TRIzol (Life Technologies). cDNA synthesis was performed using Verso cDNA kit (Thermo Scientific) and qPCR analysis was performed as described below (see section qPCR).
Peritoneal macrophages were prepared as follows: In each of three separate replicate experiments, resident peritoneal macrophages were isolated from 2–4 female C57BL/6J mice (ages 8–12 weeks) using intraperitoneal lavage as described previously [70]. Peritoneal exudate cells were pooled, resuspended in peritoneal macrophage medium (Roswell Park Memorial Institute 1640 medium plus L-glutamine plus 10% FBS plus 1% penicillin/streptomycin) and then plated into two tissue culture wells at approximately 106 cells/well. After two hours, wells were triple-washed with warm PBS to select for adherent cells. Cells were incubated for 24 h in the presence of 25 µg/mL (by protein content) CuSO4-oxidized human LDL (oxLDL, Intracel, Frederick, MD) or vehicle.
BMDM (prepared from N = 3 mice) were differentiated in six well plates and grown to confluency. Cells were serum-starved in 0.2% FBS overnight and a single scratch wound was made with a plastic p200 tip. Cells were washed with starvation media to remove non-adherent cells and incubated with or without Wnt3a (400 ng/mL) for 24 h. The cells were then fixed in 3.7% formaldehyde, and wound area visualized. The percentages of area covered in 24 h were quantified using NIH ImageJ software. All conditions were done in quadruplicates.
Plasma cholesterol levels were determined by a colorimetric enzymatic assay (Wako Diagnostics, Richmond, VA). Mice were sacrificed at the time points indicated below (see Results) and hearts (for aortic root analysis) or aortic arch grafts were harvested and frozen in Optimum Cutting Temperature (OCT) compound and serial-sectioned at a thickness of 6 µm onto positively charged glass slides (Superfrost Plus, Fisher Scientific, Pittsburgh, PA).
For CD68 immunostaining and laser capture microdissection, slides were hematoxylin-stained, cleared with xylenes, air-dried, and foam cells were identified by light microscopy. Every seventh section was immunostained with a CD68-specific antibody (AbD Serotec, MCA 1957) as previously described [23] and used as a guide slide for laser-capture microdissection for the next six serial sections. Sections were imaged at 40× on a Leica DM4000B microscope. In addition, slides that were used for morphometric analysis were counterstained with eosin and the CD68-immunostained areas were quantified by computer-aided morphometric analysis of digitized images (Image-Pro Plus software v5, Media Cybernetics, Silver Spring, MD). For isolation of CD68+ cells from the plaques, laser-capture microdissection (LCM) was performed with the PixCell II instrument (Arcturus Bioscience) as previously described [45], [46]. All LCM procedures were performed under RNase-free conditions. Student's t-test was used for comparisons of CD68+ pixel areas and cholesterol levels for each sample group vs. baseline. A two-way ANOVA was used to test for differences in CD68+ pixel areas and cholesterol levels, between Mttp-inactivated and vehicle-treated sample groups at the two time points (seven and 14 days post-injection). In the linear model for the ANOVA, the post-treatment time point (day seven or day 14) was taken as the first factor, and the treatment (poly I:C or saline) was taken as the second factor.
β-catenin immunohistochemistry (IHC) was carried out using frozen tissue sections from aortic grafts from eight animals (N = 4 animals per genotype group). Sections were fixed in acetone for 10 min, blocked with DAKO blocking buffer (DAKO, Cat. # X0909) and incubated with β-catenin rabbit polyclonal antibody (Santa Cruz, Cat, # sc7199, 1∶200) overnight at 4°C. As a negative control, sections from Apoe−/− recipient animals were assayed without the primary antibody. Sections were then incubated with FITC-conjugated secondary antibody for one hour, washed and mounted onto slides using VECTASHIELD mounting medium with diamidino-2-phenylindole (DAPI) (Vector Labs, Burlingame, CA). Images at 63× magnification were obtained using a Leica SP5 confocal microscope. Quantification of the mean fluorescent pixel intensity was performed on immunofluorescent images using Image-Pro Plus software v5. Six sections per animal were analyzed. An unpaired, two-tailed Student's t-test was used to test for differences in the numbers of pixels that were positive for β-catenin fluorescence, between the sample groups.
For LCM-derived samples, RNA was isolated from LCM caps using the PicoPure RNA Isolation Kit (Life Technologies) following the manufacturer's instructions. Total RNA was quantified using the NanoDrop spectrophotometer and assayed for quality using the Agilent Bioanalyzer 2100. Single-stranded cDNA was produced in microgram quantities from the RNA template using the Ovation Pico WTA System v2 (NuGEN), with the QIAquick PCR Purification Kit (QIAGEN) used for cDNA purification. The cDNA was analyzed on the Agilent 2100 Bioanalyzer and then fragmented and biotin-labeled using the Encore Biotin Module (NuGEN).
For peritoneal macrophage-derived samples, total RNA was isolated using TRIzol and RNA quality was analyzed with an Agilent 2100 Bioanalyzer. The three independent experiments thus yielded six RNA samples. Sample mRNA was amplified and labeled with the Affymetrix One-Cycle Eukaryotic Target Labeling Assay protocol and reagents. Labeled cDNA was hybridized to the Affymetrix GeneChip Mouse Exon 1.0 ST array using standard protocols and reagents from Affymetrix.
Fragmented, biotin-labeled cDNA was hybridized to the Affymetrix Mouse Exon Array 1.0 ST GeneChip and the GeneChip was washed and stained using the protocol and reagents in the Affymetrix GeneChip Hybridization, Wash, and Stain Kit. The stained GeneChips were imaged using the Affymetrix GeneChip Scanner 3000 and probe intensities were quantify from the scanned images using the Gene Chip Operating Software from Affymetrix. The transcriptome profiling data for this study are available online in the NCBI Gene Expression Omnibus (GEO) database, under accession numbers GSE52482 and GSE58913.
Affymetrix microarray data from the Cho et al. study [100] and the Hägg et al. study [101] (see Table S5 for references) were obtained from NCBI GEO and processed using the Robust Multichip Average method [102] in the software package Bioconductor, to obtain background-adjusted, quantile-normalized, and probeset-summarized log2 intensities. Human gene probesets were mapped to mouse orthologs using the Bioconductor tools for querying NCBI HomoloGene, and differential expression testing was carried out using a two-way ANOVA with a Benjamini-Hochberg [103] false discovery rate cutoff of 0.15. From all other studies, lists of differentially expressed genes were obtained from data tables from the publications referenced in Table S5, and mapped to mouse orthologs using HomoloGene. For each mouse gene in the Entrez Gene database, the number of independent analyses in which the gene was reported as differentially expressed during foam cell formation or atherosclerosis progression (among the studies in Table S5) was tabulated and used as the gene's meta-analysis score. Genes with a score greater than or equal to two (see Table S2, column L) were counted as high-scoring in the meta-analysis, for statistical enrichment test (see Results). Statistical enrichment of the set of genes with meta-analysis scores of two or greater, within the set of 213 genes that are differentially expressed in plaque CD68+ cells, was tested using Fisher's Exact Test, with the larger sample set consisting of the set of 21,594 mouse genes for which a meta-analysis score was tabulated.
For microarray transcriptome profiling of plaque macrophages in the Reversa model, GeneChip scan images were processed into probe intensities using the Affymetrix GeneChip Operating Software system and then background-adjusted, quantile-normalized, and probeset-summarized using the Robust Multichip Average feature of Affymetrix Power Tools, with transcript-level probeset definitions from the Version 15 of the CustomCDF Project [104] (Ensembl Transcript-derived probeset definitions). Probe intensities were separately processed into exon-level normalized probeset intensities using the Robust Multichip Average Feature of Affymetrix Power Tools (version 1.10.2), and (separately) using exon-level probeset definitions from the CustomCDF project (Ensembl Exon) and from the Affymetrix Mouse Exon Array 1.0 ST NetAffx Annotations version 29. Exon-level detect-above-background calls were made using the two separate sets of (CustomCDF and Affymetrix) exon-level probesets using the Detection Above Background algorithm in Affymetrix Power Tools. For each exon-level probeset, detect-above-background P values were combined for all samples within a sample group using the geometric mean, and then across sample groups by taking the minimum P value, to obtain a single “present” P value for each exon-level probeset. For each Ensembl Transcript corresponding to one of the transcript-level probesets for which at least 75% of the constituent exon-level probesets was interrogated on the array, the transcript was considered to be above background in the sample if at least half of its constituent exons (that were interrogated with probes on the array) were detected above background with a “present” P value of less than or equal to 0.01 [105], [106]. For all probesets that passed the filter, log2 probeset intensities across all 24 Reversa samples (see Table S1) were analyzed for differential expression between the Mttp-inactivated and control sample groups, as described below.
For comparative transcriptome profiling of plaque macrophages in the transplant regression model, probe intensity values were obtained from a two-color 3′ microarray dataset of CD68+ cells from aortic arch grafts harvested from Apoe−/− and WT host animals three days post-surgery (N = 14 for Apoe−/−, N = 18 for WT). This previously published dataset was obtained from the lab of one of the authors (EAF) [36], and was based on the Rosetta/Merck Mouse 23.6K 3.0 A1 microarray (custom-synthesized by Agilent; design files are available at GEO accession number GPL9733). Probe intensities in each of the two color channels (corresponding to sample-derived RNA and a universal mouse RNA reagent (Stratagene) that was used as an internal control) were combined to quantify absolute expression (geometric mean) and normalized expression (ratio of the macrophage sample-derived intensity to the internal control-derived intensity) for each of the samples. Probes were filtered for minimum absolute log2 intensity (greater than or equal to −5) to remove probable below-background probe intensities. For all probes that passed the filter, log2 expression ratios across all 32 samples were analyzed for differential expression between the WT and Apoe−/− sample groups, as described below.
Testing for differential expression was performed using a two-factor model (sex and treatment for the Reversa study, sex and genotype for the transplant study) with empirical Bayes variance estimates, using the Limma package in Bioconductor [107], with a proportion factor of 0.03. Transcripts were selected as differentially expressed between the sample groups if the treatment factor P value was less than 0.01. For each unique gene among the set of differentially expressed transcripts, a representative transcript was selected by selecting (from among the differentially expressed transcripts that are associated with the gene in the Ensembl database) the transcript with the largest number of exons represented on the microarray. Enrichment tests for gene annotation terms were carried out separately for the sets of upregulated and downregulated genes in the Reversa transcriptome dataset, using the gene annotation analysis tool DAVID [108] with a P value cutoff of 0.05. Gene sets were analyzed with three different functional annotation enrichment tools to determine whether gene annotations related to apoptosis or cell death are overrepresented: DAVID, Gene Set Enrichment Analysis [109], and Ingenuity Pathway Analysis (Ingenuity Systems, Redwood City, CA).
Microarray analysis of poly I:C-treated macrophages was carried out using the raw microarray intensity data from a previous study [110] in which primary murine bone marrow-derived macrophages (BMDM) were treated in vitro for two hours with 6 µg/mL poly I:C and then profiled using the Affymetrix Mouse Exon Array 1.0 ST GeneChip. Six CEL files (three from each of the two sample groups, untreated and poly I:C-treated) were obtained from NCBI GEO (accession number GSE7052) and processed as described above (save for the proportion factor used for the empirical Bayes step in LIMMA being 0.1). The statistical cutoff used for differential expression tests was P<0.01. The statistical test for overlap between the poly I:C-responsive genes in BMDM and the genes that are transcriptionally responsive to lipid lowering in vivo was carried out using Fisher's Exact Test (two-sided).
Microarray transcriptome profiling probe-level measurements from mouse peritoneal macrophages were quantile-normalized without background adjustment and then summarized for probeset intensities using the Affymetrix PLIER algorithm [111] and for detection above background using the Affymetrix DABG algorithm [112] using probe-to-probeset mappings from the CustomCDF project [104] version 15, based on the mouse transcripts and mouse exons lists from the Ensembl database release 65. Transcripts were marked as “present” only if at least 2/3 of the constituent exons had a DABG P value of less than 0.01 [105]. Differential expression testing for each Ensembl transcript was carried out using a t-test with empirical Bayes variance estimates, using the Limma package [107] in Bioconductor with 0.1 as the prior expected proportion of differentially expressed genes. P values for differential expression between the sample groups (vehicle and oxLDL) were converted to q-values using the Bioconductor package qvalue [113]. Transcripts were filtered with a false discovery rate cutoff of 0.05 and a minimum absolute fold-change cutoff of 2.0, and filtered to include only transcripts that could be mapped to the RefSeq NM identifier. Transcripts were mapped start site coordinates on the mm9 genome assembly, and noncoding sequences within ±5 kbp of the transcription start sites were retrieved, using the BioMart tool in Ensembl release 65.
In this section, we describe the computational method REMINISCE (Regulatory Element Machine-learning to Infer Networks Instructing Specific Cellular Expression; see Figure S1) that was used to identify transcription factor binding site motifs whose matches are overrepresented within 5′ regulatory regions of differentially expressed genes. Using the BioMart interface to the Ensembl database (release 65), transcription start site locations (on the NCBI M37 mouse genome assembly) were obtained for the sets of differentially expressed transcripts that were detected by microarray in CD68+ cells in plaques in the Reversa or transplant regression models. Noncoding DNA sequences (defined by excluding locations marked as protein coding within at least one exon within the Ensembl database) within ±5 kbp of a transcription start site were analyzed using a Random Forest classifier [114] to identify probable cis-regulatory regions, with a score assigned at 100 bp intervals based on the voting fraction of trees in the classifier. The features used for the classification (Table S8) include genome-wide measurements (binarized peak locations) of histone acetylation in bone marrow-derived macrophages obtained by ChIP-seq [39], “valley scores” in histone acetylation ChIP-seq signal [39], DNase I hypersensitive sites in bone marrow-derived macrophages [39] detected by short-read sequencing, vertebrate PhastCons interspecies sequence conservation scores from the University of California Santa Cruz Genome Bioinformatics database [115], GC content, computational predictions of CpG island locations [116], and basepair distances to the nearest binarized feature locations of each type. The Random Forest classifier for predicting cis-regulatory regions was trained using compendium of transcription factor binding site locations obtained by combining ChIP-seq datasets targeting 19 transcriptional regulators in murine macrophages, that were obtained from the literature (Table S8). Noncoding sequence within regions predicted to have a cis-regulatory function by the Random Forest model (voting fraction greater than 0.5) was scanned for transcription factor binding site motif matches using a combined set of vertebrate binding site position weight matrices from the TRANSFAC Professional 2010 [117] and JASPAR 2010 [118] databases. The scanning was performed using the software tool Clover [119] (Feb. 19, 2010 release) on noncoding sequences within cis-regulatory regions (after masking low-complexity repeats using NCBI DustMasker) proximal to transcription start sites of differentially expressed transcripts as well as proximal to transcription start sites of a set of 1,000 randomly selected genes that were detected above background in plaque macrophages by microarray hybridization (see Microarray Analysis). For the Clover scanning, a pseudocount value of 0.3 [120] was used. Matches for binding site sequence patterns within the promoter sequences for upregulated or downregulated genes were assessed for statistical overrepresentation using the frequencies compiled on the promoter regions for the 1,000 randomly selected macrophage-expressed genes, using the binomial test (P value threshold of 0.01). Subsequent to the motif analysis using TRANSFAC 2010, 5′ cis-regulatory regions for the upregulated Reversa gene set were re-analyzed using a substantially larger set (numbering 56) of TCF/LEF transcription factor family motifs from an updated motif database (TRANSFAC Professional 2013.4), using the same methods.
Official gene symbols for genes encoding protein components of transcription factors that were implicated in the transcription factor binding site enrichment analysis (see Figure 3A) were uploaded to the GeneMANIA network analysis web application [59]. The interaction network for the transcription factors was assembled using protein-protein interactions (direct physical interactions only), and visualized using Cytoscape [121].
MicroRNA enrichment analysis was carried out using the CORNA software algorithm [122] using conserved microRNA target predictions from miRBase (v5) and using Ensembl BioMart [123] (release 70) for gene identifier translation. Statistical enrichment was assessed using Fisher's Exact Test (P value threshold of 0.01). Pathway enrichment analysis of genes with target sequences for the microRNAs shown in Figure S4 was carried out using the analysis program DIANA-miRPath [65], with target sequences predictions from the DIANA-microT-CDS v.5.0 algorithm [124].
A set of 239,165 eSNP-to-gene associations (that collectively connect 157,668 eSNPs to monocyte expression levels of 13,989 genes) was compiled by merging (eSNP,gene) pair lists from two human population genetic studies of monocyte gene expression regulation [48], [49] comprised of samples from 1,490 and 283 individuals, respectively. From these 239,165 pairs, eSNPs that were associated with human genes whose mouse orthologs (obtained via successive mapping using the HomoloGene version 68 and the INPARANOID [125] version 8.0 databases) were detected as differentially expressed in plaque CD68+ cells in the Reversa model of plaque regression (for Reversa, the 213 genes in Table S2) or in the aortic transplant model of plaque regression (from Table S2 of [36], the 549 genes with absolute geometric-mean expression ratio >1.25 between the WT and Apoe−/− recipient sample groups) were then obtained, resulting in lists of 2,380 (Reversa) and 5,024 (transplant) plaque regression-associated eSNPs, respectively. The three sets of eSNPs – Reversa eSNPs, transplant eSNPs, and the full 157,668 “monocyte eSNPs” – were expanded to include proxy SNPs that are both in linkage disequilibrium (LD) with an eSNP (R>0.9) and within 250 kbp of the eSNP, yielding LD-expanded sets of 6,214, 13,248, and 363,283 SNPs, respectively. The expansion to proxy SNPs was carried out using LD maps from the 1,000 Genomes Project (CEU population group) and using the SNP analysis software SNAP [126]. A set of 761 SNPs that are associated with risk of atherosclerosis or CAD was obtained by searching the NCBI Phenotype-Genotype Integrator database [51] for the MeSH terms “atherosclerosis”, “coronary artery disease”, and “carotid stenosis”. Overlaps were computed for the set of 761 atherosclerosis/CAD SNPs with four different sets of SNPs: the 6,214 LD-expanded eSNPs from the Reversa model (overlap = 6), the 13,248 LD-expanded eSNPs from the aortic transplant model (overlap = 11), the 363,283 LD-expanded monocyte eSNPs (overlap = 134), and the set of all 44,278,189 validated SNPs in dbSNP build 138 (overlap = 761). From these overlaps, two 2×2 contingency tables were constructed to test for association between atherosclerosis/CAD SNPs and regression gene-associated eSNPs in the Reversa and aortic transplant models, respectively, using the LD-expanded set of monocyte eSNPs as a background set. Additionally, two 2×2 contingency tables were constructed to test for association between atherosclerosis/CAD SNPs and regression gene-associated eSNPs in the Reversa and aortic transplant models, respectively, using the set of all validated SNPs as a background set. P values and odds ratios (Figure 2C and Table S6) were obtained for these four contingency tables using Fisher's Exact Test. Confidence intervals (shown in Figure 2C) on the expected overlap of atherosclerosis/CAD SNPs and regression-associated eSNPs were obtained using the quantile function of the hypergeometric distribution.
Computational analyses were carried using scripts in the R statistical computing environment, also using the Bioconductor set of packages for R), and in the Perl programming language. Bar graphs were prepared using Prism (GraphPad Software, La Jolla, CA). All custom software scripts used in the computational analysis are available from the authors upon request. Statistical tests were carried out using R and GraphPad Prism.
|
10.1371/journal.pbio.0050132 | Distinct Functions for Different scl Isoforms in Zebrafish Primitive and Definitive Hematopoiesis | The stem-cell leukemia (SCL, also known as TAL1) gene encodes a basic helix-loop-helix transcription factor that is essential for the initiation of primitive and definitive hematopoiesis, erythrocyte and megakarocyte differentiation, angiogenesis, and astrocyte development. Here we report that the zebrafish produces, through an alternative promoter site, a novel truncated scl (tal1) isoform, scl-β, which manifests a temporal and spatial expression distinct from the previously described full-length scl-α. Functional analysis reveals that while scl-α and -β are redundant for the initiation of primitive hematopoiesis, these two isoforms exert distinct functions in the regulation of primitive erythroid differentiation and definitive hematopoietic stem cell specification. We further demonstrate that differences in the protein expression levels of scl-α and -β, by regulating their protein stability, are likely to give rise to their distinct functions. Our findings suggest that hematopoietic cells at different levels of hierarchy are likely governed by a gradient of the Scl protein established through temporal and spatial patterns of expression of the different isoforms.
| Hematopoiesis is the process that generates all the body's blood cells. In vertebrates, hematopoietic development occurs in two phases: a transitory embryonic, or primitive, wave produces only erythrocytes (red blood cells) and myeloid cells; an adult, or definitive, wave gives rise to at least three blood cell lineages, including erythrocytes and two types of immune cells—myeloid cells and lymphocytes. Previous studies have shown that the stem-cell leukemia (SCL) gene is essential for hematopoietic stem cell specification and erythrocyte maturation. Yet how SCL regulates these distinct processes is not fully understood. This study demonstrates that zebrafish produce a smaller isoform of scl, scl-β, that plays an overlapping role with the full-length scl-α in the initiation of primitive hematopoiesis, but possesses distinct functions in regulating primitive erythrocyte maturation and definitive hematopoietic stem cell development. We further show that the distinct functions of scl-α and -β are likely due to differences in their protein expression level through the regulation of protein stability. We postulate that hematopoietic cell development at different levels of hierarchy is governed by a gradient of the Scl protein created by temporal and spatial expression of different scl isoforms.
| On the basis of anatomic locations of development, time of initiation, and cell type produced, vertebrate hematopoiesis can be divided into primitive and definitive programs [1–3]. In mouse, the primitive, or first, wave of hematopoiesis initiates in the yolk sac at about embryonic day 7.5 and produces primarily nucleated embryonic erythrocytes and macrophages [4,5]. The definitive, or second, wave of hematopoiesis is believed to originate from the intra-embryonic aorta–gonads–mesonephros at approximately embryonic day 8.5 and give rise to all the mature blood cell types [6–8].
Similar to that of mammals, zebrafish hematopoiesis also consists of primitive and definitive programs, and produces differentiated cells analogous to most of the mature blood lineages found in mammals [9–11]. Zebrafish primitive erythropoiesis originates from the posterior lateral mesoderm (PLM) as a pair of bilateral stripes at approximately the five-somite stage [9,10,12]. These bilateral stripes extend anteriorly and posteriorly, and converge in the midline at the 20-somite stage to form the main structure of the intermediate cell mass (ICM), where the erythroid progenitors further develop. On the other hand, primitive myelopoiesis is believed to arise from the rostral blood island of the anterior lateral mesoderm (ALM) region at around the ten-somite stage, and produces mainly macrophages [10,13]. Compared to the onset of primitive hematopoiesis, the onset of zebrafish definitive hematopoiesis is less well defined. Preliminary studies indicate that the earliest definitive hematopoietic stem and progenitor cells arise from the ventral wall of dorsal aorta (DA) at around 26 to 30 h postfertilization (hpf) and subsequently migrate to the kidney, the adult hematopoietic organ in zebrafish, by 5 d postfertilization (dpf) [10,14,15].
Stem-cell leukemia (SCL, also known as TAL1) was originally identified as a proto-oncogene through the study of T cell acute lymphoblastic leukemia patients with a chromosomal translocation at the breakpoint of t(1;14) (p32;q11) [16–18]. The importance of SCL in normal hematopoiesis and angiogenesis was revealed by gene targeting analysis in mouse embryonic stem cells. Mice lacking SCL function failed to form vitelline vessels in the yolk sac and died at embryonic day 8.5 of development because of the complete absence of primitive hematopoiesis [19–21]. SCL-null embryonic stem cells, when injected into blastocysts, failed to contribute to any hematopoietic lineage in mouse chimeras [22,23]. These results demonstrate that SCL is essential for the generation of primitive and definitive hematopoietic cells as well as for the formation of yolk sac vessels. In addition to its pivotal role in early hematopoiesis, SCL also exerts important biological functions in subsequent hematopoietic lineage specification. Enforced SCL expression in hematopoietic cell lines favors erythroid differentiation [24,25], and ablation of SCL in adult mice impairs erythropoiesis and megakarypoiesis [26,27]. Despite its important functions, the molecular mechanisms of how SCL mediates these multiple functions remain obscure. Previous in vitro studies in human and mouse malignant hematopoietic cell lines have described several SCL isoforms involved in T cell leukemia development and differentiation of erythrocytes and megakaryocytes [28–33]. However, the existence and biological functions of these SCL isoforms in vivo have not been demonstrated.
In this study, we report that the zebrafish produces, through an alternative promoter site within exon 2, a novel scl (tal1) isoform, scl-β, that encodes a truncated protein lacking the N-terminal 118 amino acids (aa). Whole-mount in situ hybridization (WISH) demonstrated that scl-β exhibits temporal and spatial expression that overlaps but is clearly distinctive from that of the full-length scl-α: scl-β emerges first and expresses in the entire ALM and PLM and the ventral wall of DA, where the first definitive hematopoietic progenitors arise; scl-α expresses later in the posterior of the ALM and the PLM, possibly in a manner overlapping with scl-β, and is subsequently restricted to the ICM region. Loss-of-function analysis using an antisense morpholino oligonucleotide (MO) knockdown approach revealed that while scl-α and -β are redundant in the initiation of primitive hematopoiesis, these two isoforms appear to function at different stages of primitive erythroid cell differentiation. Analysis of definitive hematopoiesis of scl-α and -β MO-injected embryos (morphants) showed that the knockdown of Scl-β, but not Scl-α, protein expression resulted in the loss of c-myb (cmyb) and runx1 expression in the ventral wall of DA as well as rag1 expression in the thymus, demonstrating that scl-β but not scl-α is essential for definitive hematopoietic stem cell development. Interestingly, we found that scl-β and -α exhibited a remarkable difference in their protein expression levels both in vitro and in vivo, indicating that differences in the protein expression level of scl-α and -β isoforms likely confer their distinct functions in the regulation of hematopoietic cell development.
To investigate whether different scl isoforms exist in zebrafish, RNA samples were prepared from 18-somite-stage embryos and kidney, the adult hematopoietic organ in zebrafish [10], and subjected to Northern blot analysis. The result showed that two transcripts, one 2.6 kilobases (kb) and the other 2.2 kb, were specifically hybridized to the probes corresponding to the coding sequence and the 3′ untranslated region (UTR) of the zebrafish scl cDNA (data not shown), suggesting that the 2.6-kb and 2.2-kb transcripts may represent two different scl isoforms. To characterize the nature of these two transcripts, we carried out a rapid amplification of cDNA ends (RACE) experiment and obtained one 3′ RACE and two 5′ RACE products (data not shown). DNA sequencing revealed that the larger 5′ RACE product was identical to the published full-length scl sequence [34,35], whereas the smaller fragment was also identical except that it lacked the first 438 base pairs at the 5′ end of the full-length scl, indicating that the 2.6-kb transcript is the full-length scl and the 2.2-kb transcript represents a novel scl isoform. This was confirmed by Northern blot analysis, which showed that while both transcripts were hybridized to the scl 3′ UTR probe (3′-probe), only the larger 2.6-kb species was recognized by the scl 5′ 414-bp probe (5′-probe) (Figure 1A). We hereafter designate the 2.6-kb full-length and the 2.2-kb truncated forms as scl-α and scl-β, respectively. Time course analysis by Northern blot revealed that the expression of scl-β began at the one- to two-somite stage, peaking at around the 11- to 18-somite stage, and maintained a lower level in the adult kidney (Figure 1A). On the other hand, the scl-α transcript was not detected until the four-somite stage, after which it rapidly increased its level to that of scl-β, subsequently becoming the dominant form in the adult kidney (Figure 1A). Considering the fact that the scl-β transcript initiates within the exon 2, where a potential TATA-box sequence can be found at position −31 of the predicted transcription initiation site (Figure 1B), and expresses earlier than scl-α (Figure 1A), we conclude that scl-β is generated from an alternative promoter site within exon 2 encoding a truncated Scl protein lacking the N-terminal 118 aa (Figure S1).
WISH was performed to examine the temporal and spatial expression of scl-α and -β. As shown in Figure 2, the 3′-probe, which recognized both scl-α and -β, exhibited a pattern identical to that of scl expression described previously [34,35]. It first emerged at around the two-somite stage as one pair of stripes in the PLM followed by the appearance of a second pair of stripes in the ALM at the four-somite stage (Figure 2A and 2B). These two pairs of stripes, which represented the combination of both scl-α and -β transcripts (referred to as scl-α/β stripes hereafter), extended anteriorly and posteriorly from the four-somite stage onwards (Figure 2B and 2C). And by the 18-somite stage, the scl-α/β stripes were mainly localized in three regions: the ALM, the anterior of the PLM (APLM), and ICM (Figure 2D). In contrast to the 3′-probe-positive signals, the 5′-probe-positive signals, which represented only the full-length scl-α, appeared at the four- and six-somite stage as two pairs of stripes (scl-α stripes) in the PLM and ALM, respectively (Figure 2H, and data not shown). By the 18-somite stage, the scl-α expression was restricted to the ICM (Figure 2J), where primitive erythropoiesis actively occurs at this stage [10], suggesting that scl-α is predominantly expressed in the erythroid lineage from the 18-somite stage onwards. This possibility was further supported by the lack of scl-α expression in the erythrocyte-deficient mon (trim33) mutant embryos [36] (Figure 2F and 2L). Notably by 26 hpf, only a weak signal of the 3′- but not the 5′-probe was detected in the ventral wall of DA (Figure 2E and 2K), where the first definitive hematopoietic stem cells presumably emerge [10,14,15], suggesting that scl-β is the main isoform expressed in the definitive hematopoietic stem cells. To test this possiblity, we carried out double staining analysis in which 26 hpf embryos were stained with anti-Scl-α (Ab-Scl-N) or anti-Scl-α/β (Ab-Scl-C) antibodies together with c-myb WISH. The transverse section through the trunk region of these embryos showed that Scl-β but not Scl-α protein was present in the ventral wall of DA in a manner partially overlapping with c-myb-positive cells (Figure 2M–2R), indicating that the definitive hematopoietic stem cells predominantly express scl-β. Based on these observations, we conclude (as illustrated in Figure 2S) that scl-β appears first and expresses in the entire ALM and PLM regions; scl-α emerges later in the ALM and ICM, possibly in a manner overlapping with scl-β, and is subsequently restricted to the ICM by the 18-somite stage. Notably, only scl-β expresses in the ventral wall of DA where the first definitive hematopoietic stem cells arise.
To determine the biological functions of scl-α and -β in hematopoiesis, two MOs, scl-α MO and scl-β MO (Figure 1B), that specifically inhibited the protein syntheses of scl-α and -β, respectively, were injected into wild-type zebrafish embryos. Immunohistochemistry staining showed that scl-α MO and -β MO specifically abolished the Scl-α and -β protein expression as indicated by lack of anti-Scl-N staining in the ICM of scl-α morphants (Figure 3N, n = 35/35), and the selective loss of anti-Scl-C staining in the APLM (where only scl-β was transcribed [Figure 2D]) of scl-β morphants (Figure 3S, n = 32/32). Microscopic analysis of the 24-hpf embryos injected with either MO revealed no obvious abnormality in general morphology (data not shown). Examination of the expression of gata1 (Figure 3A, n = 46/46; 3B, n = 43/43; and 3C, n = 41/41), βe1-globin (hbbe1) (Figure 3E, n = 49/49; 3F, n = 51/51; and 3G, n = 49/49), and pu.1 (spi1) (Figure 3I, n = 41/41; 3J, n = 41/43; and 3K, n = 45/45) by WISH confirmed that the initiation of primitive erythropoiesis and myelopoiesis were intact in the scl-α and -β morphants. We reasoned that the lack of phenotypes in both morphants was likely due to the functional redundancy of scl-α and -β. To test this possibility, we co-injected scl-α MO and scl-β MO to block both protein syntheses (Figure 3P, n = 45/45; and 3T, n = 50/50), and found that the expression of gata1, βe1-globin, and pu.1 were either absent or drastically reduced in the co-injected morphants (Figure 3D, n = 42/42; 3H, 46/46; and 3L, n = 47/47). This phenotype is very similar to that found in the scl-sp morphants, in which both Scl-α and -β protein expression were eliminated by scl-sp MO (Figure 1B) that interfered with the splicing between exon 2 and 3 of the scl gene [37,38]. Furthermore, injection of in vitro synthesized scl-α or -β mRNA was sufficient to rescue the expression of gata1, βe1-globin, and pu.1 in the scl-sp morphants (data not shown). Taken together, we conclude from these results that scl-α and -β are functionally redundant in the initiation of primitive hematopoiesis.
To address their functions in the late developmental stages of primitive erythropoiesis, we examined the scl-α and -β morphants beyond 30 hpf. Although primitive erythropoiesis initiated normally, o-dianisidine staining revealed that the red blood cells (RBCs) in the scl-β morphants were significantly reduced by 2 dpf (data not shown) and finally not detectable by 3 dpf (Figure 4C). On the other hand, RBCs in the scl-α morphants were normal before 3 dpf (Figure 4B) but began to decrease by 4 dpf (data not shown) and were severely reduced by 5 dpf (Figure 4E). These data indicate that the loss of either Scl-α or -β protein renders abnormal RBC differentiation at different developmental stages, eventually resulting in anemia. To provide an additional test of this possibility, circulating RBCs were collected from the scl-α and -β morphants, stained with May-Grunwald Giemsa, and compared to those from wild-type embryos. Based on the size of cell, shape of nucleus, and staining of cytoplasm, normal primitive RBCs from 30 hpf to 5 dpf can be classified into four main stages: stage I, basophilic erythroblast; stage II, polychromatophilic erythroblast; stage III, orthochromatophilic erythroblast; and stage IV, mature erythrocyte (Figure 4G). In the 2-dpf wild-type embryos and 2-dpf scl-α morphants, more than 98% of the circulating RBCs were at stage II (Figure 4H and 4I). However, 90% of the circulating RBCs arrested at stage I in the 2-dpf scl-β morphants (Figure 4J). As expected, although the circulating RBCs developed normally before 3 dpf, 80% of them were blocked at stage II in the 4-dpf scl-α morphants (Figure 4L), compared to the number of RBCs at stage III in the 4-dpf wild-type embryos (Figure 4K). These phenotypes concurred with the virtual Northern blot result showing that circulating RBCs from 30-hpf wild-type embryos contain both scl-α and -β, whereas those from 2-dpf wild-type embryos express predominantly scl-α (Figure 5). Collectively, we conclude that scl-β plays a critical role in the differentiation of basophilic erythroblasts to polychromatophilic erythroblasts, whereas scl-α is pivotal for the transition from polychromatophilic erythroblasts to orthochromatophilic erythroblasts.
We next explored the roles of the scl-β and -α isoforms in definitive hematopoiesis. In zebrafish, definitive hematopoietic stem cells originate from the ventral wall of DA at around 26 hpf to 30 hpf, and these cells are enriched in c-myb and runx1 expression [10,14,15]. We therefore first examined c-myb and runx1 expression in the 30-hpf scl-α and -β morphants. WISH revealed that expression of c-myb and runx1 in the ventral wall of DA was abolished in the scl-β morphants (Figure 6C, n = 43/45; and 6F, n = 40/43) but not in the control embryos (Figure 6A, n = 40/40; and 6D, n = 42/42) or scl-α morphants (Figure 6B, n = 45/46; and 6E, n = 43/45). As artery endothelial cells appeared to be retained in both scl-α and -β morphants, as indicated by two artery-specific markers deltaC (dlc) and grl (hey2) (Figure S2), these data indicate that scl-β, but not -α, is essential for definitive hematopoietic stem cell development. This result is consistent with the finding that only scl-β is expressed in the ventral wall of DA (Figure 2). To further test whether scl-β is indeed required for the development of definitive hematopoietic stem cells, we investigated T cell development in both morphants by examining rag1 expression at 5 dpf. As expected, rag1 was detected in the thymus of control embryos and scl-α morphants (Figure 6G, n = 30/30; and 6H, n = 42/45) but not scl-β morphants (Figure 6I, n = 46/50). Taken together, these data demonstrate that scl-β is essential for the development of definitive hematopoietic stem cells while scl-α is dispensable.
To gain insight into the molecular basis underlying the distinct functions of scl-α and -β in hematopoietic cell development, we performed rescue experiments—by co-injecting in vitro synthesized scl-α or scl-β mRNA with scl-sp MO into wild-type embryos—to test whether the lack of the N-terminal 118 aa could lead to the differences in their biological functions. Examination of the expression of βe1-globin and c-myb by WISH showed that either isoform was sufficient to rescue both primitive and definitive hematopoietic defects in the scl-sp morphants (Figure 7A, n = 35/35; 7B, n = 41/41; 7C, n = 44/44; 7D, n = 47/47; 7E, n = 36/36; 7F, n = 40/45; 7G, n = 41/45; and 7H, n = 42/46). The data indicate that the N-terminal 118-aa segment is not essential for determination of the functions of these two Scl protein isoforms, consistent with previous findings showing that the basic helix-loop-helix domain of the murine SCL protein is sufficient for hematopoiesis [39]. As recent studies have suggested that different SCL protein expression levels are required at different levels of hematopoietic hierarchy [40,41], these findings raise the possibility that the distinct functions of scl-α and -β are due to differences in their protein expression levels, resulting in establishment of a gradient of Scl protein at different stages of hematopoietic hierarchy. To test this hypothesis, we performed immunoblotting of protein extracts from 18-somite-stage wild-type embryos, in which the transcription levels of the scl-α and -β isoforms are similar (the scl-β RNA is slightly higher) (Figure 1A). We found that the protein expression level of Scl-β was much lower than that of Scl-α in these embryos (the Scl-β protein was hardly detectable at all by Western blot; Figure 7I). When COS7 cells were transiently transfected with construct expressing either the full-length scl-α or scl-β, the protein expression level of Scl-β was also much lower than that of Scl-α, while their RNA levels were comparable (Figure 7J). These data strongly indicate that a post-transcriptional mechanism is involved in the regulation of Scl-β protein expression level. To test this, equal amounts of in vitro synthesized scl-α and -β mRNA were injected into one-cell-stage wild-type embryos, and protein levels were examined at different time points post-injection (Figure 7K). As anticipated, real-time reverse transcriptase PCR analysis revealed that mRNAs of both isoforms behaved similarly in these injected embryos (data not shown). Immunoblotting of whole embryo protein extracts showed that protein expression levels were comparable at 3 h post-injection, indicating that both isoforms are effectively translated. However, by 4 h post-injection, Scl-β protein level was greatly reduced, while Scl-α level increased (Figure 7K), showing that the low protein expression level of Scl-β was likely due to the rapid degradation of its protein. Taken together, these data strongly indicate that differences in the protein expression levels of Scl-α and -β isoforms, via regulation of their protein stabilities, likely confer their distinct functions in the regulation of hematopoietic cell development.
It is believed that the complexity in morphology and behavior of higher organism is achieved not only by higher gene numbers, but also by multiple protein isoforms being encoded by a single gene locus and by the complexity of protein–protein interactions. The most well studied phenomenon that results in the generation of multiple protein isoforms from a single gene is alternative splicing of pre-mRNA [42]. However, other mechanisms such as use of an alternative promoter—a phenomenon that is as equally widespread in higher organisms as alternative pre-mRNA splicing [43,44]—are less appreciated. In this article, we described how zebrafish produce, through alternative promoter sites, two scl isoforms, the full-length scl-α and a novel truncated scl-β (Figure 1). We further showed that these two scl isoforms manifest distinct temporal and spatial expression (Figure 2) and exert distinct functions in the regulation of primitive and definitive hematopoiesis (Figures 3–6). The identification of the alternative-promoter-generated scl-β isoform in zebrafish has not only revealed new insight into the roles of scl in the regulation of hematopoietic cell development, but also provided another example to highlight the importance of alternative promoter usage in generating protein and regulatory diversity.
Previous studies have revealed that mammals contain several SCL isoforms generated by either alternative splicing, alternative promoters, or alternative translation initiation sites [28–33]. These mammalian SCL isoforms arise from alternative promoters in exon 1a and exon 1b, and encode identical proteins, the full-length SCL. In contrast, scl-β in zebrafish is generated through an alternative promoter site within the scl exon 2, the equivalent of the mammalian SCL exon 4, and encodes an N-terminal truncated protein (Figures 1 and S1). Thus, scl-β is clearly distinct from the previously described mammalian SCL isoforms. One intriguing question raised is whether the mechanism of generating different scl isoforms found in zebrafish is evolutionarily conserved in higher vertebrates, especially in mammals. Considering the facts that (1) the SCL locus is highly conserved in vertebrates [45], (2) there are observations showing that the murine SCL exon 4 has promoter activity in the context of SCL 3′ stem cell enhancer [46], and (3) a truncated SCL transcript initiated from exon 4 can be detected in some of the human T cell leukemia cell lines [28,29], we speculate that a scl-β equivalent may exist in higher vertebrate species.
Another interesting issue raised by this study is the regulation of scl-α and -β transcripts during hematopoiesis. Notably, scl-β first appears in hematopoietic stem and progenitor cells and soon diminishes in the differentiated primitive RBCs (Figures 2 and 5). In contrast, scl-α emerges later and is predominantly restricted to RBCs (Figures 2 and 5). Thus, it appears that an on–off switch, from scl-β to -α expression, must occur during primitive RBC development. Considering the facts that (1) the earliest definitive hematopoietic stem/progenitor cells located in the ventral wall of DA express only scl-β and (2) scl-α becomes the predominant isoform expressed in the adult kidney marrow, where definitive hematopoiesis takes place presumably from 5 dpf onwards in zebrafish development, it is conceivable to speculate that this on–off switch may also exist during definitive erythroid cell development. However, it is unclear at this moment whether this on–off switch takes place at the transcriptional level or the post-transcriptional level, or perhaps a combination of both. Nevertheless, we believe that the on–off switch of these scl isoforms must play a crucial role in normal hematopoietic cell development, at least for RBCs, and that the underlying molecular basis of this regulation warrants further studies.
Our study has provided evidence indicating that differences in the protein expression levels of the scl-α and -β isoforms are likely to confer their distinct functions in regulating hematopoietic cell development (Figure 7). Although we cannot rule out the possibility that translational control may contribute to the regulation of their protein expression levels, the fact that the Scl-β protein was initially expressed at a level comparable to that of Scl-α but soon reduced dramatically upon injection of equal amounts of in vitro synthesized scl-α and -β mRNA (Figure 7K) strongly indicates that differences in the protein expression levels of the scl-α and -β isoforms are largely due to the rapid turnover of the Scl-β protein. However, the triggers causing the onset of rapid degradation of Scl-β protein in vivo are unclear. Given the differences in their N-terminal residues, one could speculate that, perhaps, the short half-life of the Scl-β protein is mediated through the N-end rule degradation, a common proteolytic pathway that is present in prokaryotes, fungi, plants, and animals [47]. Further biochemical analyses are required to clarify this issue.
Finally, our data strongly suggest that the establishment of an appropriate Scl protein gradient at different levels of hematopoietic hierarchy—a low level in hematopoietic stem and progenitor cells and a high level in differentiated RBCs—is essential for hematopoietic cell development. The phenomenon of a lower Scl protein level in hematopoietic stem and progenitor cells correlating to the importance of hematopoietic cell development is intriguing. One possibility is that the specification of definitive stem and progenitor cells requires the low concentration of Scl protein, which occurs by preferential expression of scl-β during early hematopoiesis. However, the fact that injection of either scl-α or -β mRNA is sufficient to rescue the c-myb expression at 30 hpf in the ventral wall of DA in scl-sp morphants (Figure 7) suggests that this may not be the case. A high Scl protein level, which is known to be required for the maturation of RBCs [40,41], has the tendency to promote hematopoietic stem and progenitor cell differentiation into erythroid lineage, so a more likely explanation is that the low concentration of the Scl-β protein ensures the proper expansion of these cell pools by promoting their proliferation rather than their differentiation. In addition, the low concentration of Scl-β may also be crucial for maintaining an unbiased differentiation potential of hematopoietic stem and progenitor cells during ontogeny. Uncovering the molecular basis of scl-α- and -β-mediated actions will provide further insight into our understanding of the specification, proliferation, and differentiation of hematopoietic lineages.
Zebrafish were maintained at 27 to 28 °C as described in [48]. The clos5 and montg234 mutants were kindly provided by Didier Y. R. Stainier (University of California San Francisco, United States) and Artemis Pharmaceuticals (Germany), respectively.
The full-length scl-α and -β DNA were amplified by RT-PCR and cloned into pCS2+ vector. The scl-5′ and scl-3′ constructs contained the first 414 bp of the 5′ UTR of scl-α and the last 539 bp of the 3′ UTR of scl-α/β, respectively. They were amplified by PCR using two sets of specific primers (scl-5′: 5′-acttcagtgcatctaaaacctcag-3′/5′-ttttatatccgcgctccctcctc-3′; scl-3′: 5′-tggaaattcaagcgggtaatgac-3′/5′-gggcttttcatataaaatttgtgag-3′) and subcloned into pGEM-T Easy and pGEM-T vector (Promega, http://www.promega.com).
Total RNA was extracted from wild-type embryos and subjected to 5′ and 3′ RACE using the SMART RACE cDNA Amplification Kit (Clontech, http://www.clontech.com) according to the manufacturer's instructions. For 5′ and 3′ RACE, two sets of primers (scl-5′-P1: 5′-aagttgatgtacttcatggccag-3′; scl-5′-P2: 5′-atacatcccatactgttccgcatctccagc-3′; scl-3′-P1: 5′-gcggaacagtatgggatgtatcct-3′; scl-3′-P2: 5′-ctagtgcgggacgacctc-3′) were used. The RACE products were cloned into pGEM-T Easy vector (Promega) and subsequently sequenced.
The T-Coffee method [49] was used for protein sequence alignment.
Total RNA from different stages of embryos, adult kidney, and RBCs were prepared using the RNeasy Kit (Qiagen, http://www.qiagen.com) according to the manufacturer's instructions. mRNA was purified using the NucleoTrap mRNA mini kit (Macherey-Nagel, http://www.macherey-nagel.com) according to the manufacturer's instructions. Three micrograms of each embryonic mRNA and 30 μg of total kidney RNA were used for Northern blot analysis (Figure 1A). For virtual Northern blot, total RNA from 30-hpf and 2-dpf embryos and RBCs were reverse transcribed into cDNA and then amplified by the SMART PCR cDNA synthesis kit (Clontech). The amplified cDNAs were used as targets for hybridization with the DIG-labeled scl-5′ and scl-3′ probes. DIG labeling was carried out by PCR amplification with the DIG Probe Synthesis Kit (Roche Applied Science, http://www.roche-applied-science.com) according to the manufacturer's instructions. Northern blot and virtual Northern blot analyses were performed as previously described [50,51].
Fish embryos were stained for 15 min in the dark in o-dianisidine staining solution as previously described [12].
Generation of the DIG-labeled anti-sense RNA probes and whole-mount in situ hybridization were performed as described in [48]. For transverse cryosection, embryos were embedded in OCT solution (Sakura, http://www.sakura.com) after WISH and immunochemistry staining. Embedded embryos were sectioned at a thickness of 10 μm using a Leica (http://www.leica.com) microtome. The sections were mounted on glass slides and imaged using a Zeiss (http://www.zeiss.com) AxioPhot 2 imaging system.
Anti-zebrafish Scl-α N-terminal (aa 17 to 62, referred to as Ab-Scl-N) and Scl-α/β C-terminal (aa 255 to 325, referred to as Ab-Scl-C) antisera were generated by immunizing rabbits with the GST-Scl-N and GST-Scl-C fusion proteins using standard protocol. Antibody purification and immunohistochemistry staining were carried out as previously described [52].
Anti-sense MOs (Gene Tools) were designed as follows: scl-α MO, 5′-gctcggatttcagtttttccatcat-3′; scl-β MO, 5′-gcggactcaactgcaccattcgagt-3′; scl-sp MO, 5′-agatttaaaatgctcttaccatcgt-3′. Three nanograms of scl-α MO, 8 ng of scl-β MO, and 8 ng of scl-sp MO mixed with phenol red were separately injected into wild-type embryos at the one-cell stage. Similarly, 3 ng of scl-α MO and 8 ng of scl-β MO were used for the co-injection experiment. Wild-type embryos injected with a mixture of sterile water and phenol red were used as control.
Fish embryos were anesthetized in calcium- and magnesium-free PBS (pH 7.4) containing 0.02% tricaine (Sigma-Aldrich, http://www.sigmaaldrich.com) and 1% BSA (Sigma-Aldrich). After tail clipping using surgical scissors, blood cells were collected by pipetting and cytospun onto slides by centrifugation at 450 rpm for 3 min using a Cytospin 4 (Thermo Scientific, http://www.thermo.com). The slides were then air-dried and subjected to May-Grunwald Giemsa staining according to the standard protocol.
In vitro transcription was carried out using the mMESSAGE mMACHINE sp6 kit (Ambion, http://www.ambion.com) according to the manufacturer's instructions. For rescue experiments, 200 pg of scl-α or -β RNA was co-injected with 8 ng of scl-sp MO into wild-type embryos at the one-cell stage.
For protein stability analysis in fish embryos, 500 pg of in vitro synthesized scl-α or -β mRNA was injected into embryos at the one-cell stage, and protein extracts were prepared as described in [48]. Total RNA from both scl-α- and -β-injected embryos at 6 h post-injection were extracted and reverse transcribed using random hexamer as primer. The cDNAs were examined with real-time PCR using scl-specific primers (scl-961F, 5′-ctagtgcgggacgacctc-3′; scl-1528R, 5′-ggaactaaactgtgccga-3′), which can amplify both injected scl-α and -β mRNA. COS7 cells were maintained in DMEM (Gibco, http://www.invitrogen.com/content.cfm?pageid=11040) supplemented with 10% bovine calf serum (Hyclone, http://www.hyclone.com). Transient transfection was carried out by SuperFect Reagent (Qiagen) according to the manufacturer's protocol. Real-time RT-PCR was performed to ensure that the transfection efficiency was similar using the same RT-PCR protocol for protein stability analysis in fish embryos. Cell extract preparation and Western blot were carried out as described previously [50].
|
10.1371/journal.pgen.1006176 | Multiple Rad52-Mediated Homology-Directed Repair Mechanisms Are Required to Prevent Telomere Attrition-Induced Senescence in Saccharomyces cerevisiae | Most human somatic cells express insufficient levels of telomerase, which can result in telomere shortening and eventually senescence, both of which are hallmarks of ageing. Homology-directed repair (HDR) is important for maintaining proper telomere function in yeast and mammals. In Saccharomyces cerevisiae, Rad52 is required for almost all HDR mechanisms, and telomerase-null cells senesce faster in the absence of Rad52. However, its role in preventing accelerated senescence has been unclear. In this study, we make use of rad52 separation-of-function mutants to find that multiple Rad52-mediated HDR mechanisms are required to delay senescence, including break-induced replication and sister chromatid recombination. In addition, we show that misregulation of histone 3 lysine 56 acetylation, which is known to be defective in sister chromatid recombination, also causes accelerated senescence. We propose a model where Rad52 is needed to repair telomere attrition-induced replication stress.
| Telomeres are essential structures located at the ends of chromosomes. The canonical DNA replication machinery is unable to fully replicate DNA at chromosome ends, causing telomeres to shorten with every round of cell division. This shortening can be counteracted by an enzyme called telomerase, but in most human somatic cells, there is insufficient expression of telomerase to prevent telomere shortening. Cells with critically short telomeres can enter an arrested state known as senescence. Telomere attrition has been identified as a hallmark of human ageing. Homologous recombination proteins are important for proper telomere function in yeast and mammals. Yeast lacking both telomerase and Rad52, required for almost all recombination, exhibits accelerated senescence, yet no apparent increase in the rate of telomere shortening. In this study, we explore the role of Rad52 during senescence by taking advantage of rad52 separation-of-function mutants. We find that Rad52 acts in multiple ways to overcome DNA replication problems at telomeres. Impediments to telomere replication can be dealt with by post-replication repair mechanisms, which use a newly synthesized sister chromatid as a template to replicate past the impediment, while telomere truncations, likely caused by the collapse of replication forks, can be extended by break-induced replication.
| Telomeres, nucleoprotein structures located at the ends of linear chromosomes, prevent natural chromosome ends from being recognized as DNA double-strand breaks (DSBs) [1]. Due to incomplete DNA replication and nucleolytic degradation, telomeres shorten with each round of replication, which can eventually lead to a growth arrest, known as replicative senescence, or to apoptosis. Telomere shortening can be counteracted by a specialized reverse transcriptase called telomerase, which is composed of a protein catalytic subunit and an RNA subunit [2, 3]. Telomerase extends telomeres by iterative reverse transcription of a short sequence to the 3′ ends of telomeres, using the RNA subunit as a template [2, 4, 5].
Most human somatic cells do not express sufficient telomerase to prevent telomere shortening, which may be a contributing factor towards human ageing. This absence of telomere maintenance may have evolved as a barrier to tumorigenesis (reviewed in [6]). Indeed, cancer cells need to activate a telomere maintenance mechanism, and in approximately 85–90% of cancers, this occurs through the upregulation of telomerase [7]. The remaining 10–15% of cancers employ telomerase-independent, recombination-based mechanisms, collectively termed alternative lengthening of telomeres (ALT) [8].
ALT mechanisms were first described in the budding yeast Saccharomyces cerevisiae, where cells using ALT are called “survivors” [9]. There are two main types of survivors: type I and type II. Both types of survivors require the major recombination protein Rad52 and the DNA polymerase δ subunit Pol32 [9, 10]. Pol32 is essential for break-induced replication (BIR) [10], while Rad52 is important for almost all recombination-related activities, including BIR (reviewed in [11]). Type I survivors also require Rad51, Rad54, and Rad57, and maintain telomeres by amplification of subtelomeric Y′ elements [9, 12]. Formation of type II survivors, which exhibit amplification of the C1–3A/TG1–3 telomeric repeats, is Rad51-independent, but requires the MRX complex (consisting of Mre11, Rad50 and Xrs2), Rad59, and Sgs1 [12–15]. BIR can be Rad51-dependent or Rad51-independent, suggesting that type I and type II survivors maintain telomeres through Rad51-dependent BIR and Rad51-independent BIR, respectively [16, 17].
While recombination is clearly important for the maintenance of telomeres in survivors, recombination proteins are also important in pre-senescent cells [18]. Rad52 can be detected at telomeres well before the appearance of survivors [19]. Furthermore, telomerase-negative cells lacking Rad51, Rad52, Rad54, Rad57, Rad59, Pol32, or Sgs1 senesce very rapidly [9, 14, 15, 20–22]. With the exception of Sgs1, the enhanced senescence does not appear to cause a change in bulk telomere shortening [9, 12, 20, 23], although rare telomere loss events may be occurring. tlc1Δ sgs1Δ strains fail to resolve recombination intermediates at telomeres in pre-senescent cells, which may explain their accelerated senescence [24].
Rad52 mediates the exchange of RPA for Rad51 on single-stranded DNA to promote Rad51-catalyzed strand invasion [25, 26]. While this Rad51 pathway, which also requires Rad54, Rad55, and Rad57, is important for the majority of homology-directed repair (HDR), Rad52 also has Rad51-independent functions. These functions involve its DNA annealing activity, which is augmented by Rad59 [27–29]. The Rad51-mediator and the DNA annealing functions of Rad52 are separable. An alanine scan mutation study identified a class of rad52 mutants (class C mutants) that can still promote recruitment of Rad51 but is deficient in DNA annealing [30, 31]. These mutants are defective in repairing DSBs and in sister chromatid recombination (SCR) but perform BIR with only slightly reduced efficiency [30, 32, 33].
The mechanism by which HDR prevents accelerated senescence has been poorly characterized. This is in part due to the multiple Rad52 subpathways within HDR. Rad51-dependent BIR and Rad51-independent BIR have been previously implicated in delaying senescence [21–23]. In this study, we make use of rad52 class C mutants to show that SCR is also important during senescence. We also demonstrate that proper regulation of the acetylation of lysine 56 of histone 3 (H3K56) is important during replicative senescence, and we propose a model where Rad52 is repairing damage at telomeres in the absence of telomerase.
Previous studies have used telomere sequencing to detect recombination events in senescing S. cerevisiae cells [24, 34–42]. This assay takes advantage of the fact that yeast telomerase adds imperfect, degenerate repeats [43]. Sequencing multiple copies of the same telomere derived from a clonal population of cells reveals a centromere-proximal region of stable sequence and a distal region with differing degenerate repeats [44, 45]. The variation in the sequence of the distal region is largely abolished in the absence of telomerase [45], but rare sequence divergence events can be detected and have been presumed to be caused by recombination [34]. More precisely, since equal SCR generates repair products without changes in DNA sequence, the assay detects sequence divergence due to unequal SCR, intertelomere recombination, or BIR that does not result from perfect alignment with a sister telomere. These recombination events may be directly important in delaying senescence, or they may be a byproduct of other recombination-mediated activities that delay senescence. To determine the nature of these events, we sequenced telomere VI-R from est2Δ strains—EST2 encodes the protein catalytic subunit of telomerase [3]—that are also deleted for either RAD52, POL32, or RAD59. All three of these genes are required for recombination of telomeric repeats in type II survivors [9, 10, 12]. In est2Δ cells, 8.6% of the telomeres exhibit sequence divergence, similar to what has previously been reported [37]. Surprisingly, even though rad52Δ, pol32Δ, and rad59Δ telomerase-null strains senesce rapidly [9, 21, 22], we find that the divergence events do not decrease in the absence of Rad52, Pol32, or Rad59 (Fig 1), indicating that these events are not involved in the recombination-mediated delay of senescence. In fact, divergence events increase in the absence of Pol32.
pol32Δ rad52Δ double mutants are synthetic lethal [46, 47]. One interpretation of this genetic interaction is that in the absence of Pol32, DNA replication is compromised, resulting in damage that is repaired by Rad52-dependent HDR. Indeed, we see elevated levels of Rad52 focus formation in pol32Δ cells (S1 Fig). The increased divergence seen in est2Δ pol32Δ telomere sequences could be due to an increase in Rad52-dependent recombination at telomeres. Consistent with this hypothesis, we observe a further increase in Rad52 focus formation in est2Δ pol32Δ cells.
To determine the source of the Rad52-independent divergence events, we performed two controls. First, we took two plasmids with cloned and sequenced telomeres (one of 166 bp and the other 213 bp in length), amplified the telomeres by PCR, re-cloned them into the same vector, and sequenced multiple clones. We found that 3.8% of the clones exhibited sequence divergence (S2 Fig). The divergence events can be due to sequence alterations caused during PCR amplification, propagation in bacteria, and/or DNA sequencing. Second, we integrated two telomeres, 166 bp and 230 bp in length, into the URA3 locus in wild-type and rad52Δ strains, which were then clonally propagated for ~30 population doublings. We amplified these internal telomeres by PCR, cloned the PCR products, and sequenced multiple clones. We found that 4.2% and 7.3% of the clones from wild type and rad52Δ, respectively, exhibit sequence divergence (S2 Fig). The higher percentage in strains lacking Rad52 likely reflects a role for Rad52 in suppressing the accumulation of mutations [48]. While an internally-integrated telomere is not equivalent to a natural telomere, our data suggest that a significant fraction of sequence divergence events at natural telomeres in telomerase-null cells occur because of technical reasons related to amplification, cloning, and/or sequencing of telomeres. Our findings have implications with regard to using this assay to study recombination at telomeres (see Discussion), and show that the function of Rad52 in delaying senescence is unrelated to the sequence divergence events observed in senescing telomerase-negative cells. In addition, our data indicate that any Rad52-mediated HDR events during senescence most likely involves perfectly aligned sister telomeres, which would not alter the sequences of recombining telomeres, and would therefore not be detected using this assay.
Although absence of Rad52 does not alter the rate of bulk telomere shortening, truncation of a small number of telomeres may be occurring that results in accelerated senescence. It was previously determined that telomeres less than 125 bp in length are highly unlikely to arise due to the standard end-replication problem [49, 50], which shorten telomeres by 3–4 bp per generation in yeast [51], so such telomeres would mostly likely have undergone a truncation event. To determine whether Rad52 prevents such truncation events, we sequenced telomeres from two wild-type and two rad52Δ telomerase-positive strains, which were derived from the meiosis of a single rad52Δ/RAD52 diploid cell, after ~35 generations of clonal expansion. In telomerase-positive strains, most sequence divergence events are due to telomerase-mediated telomere extension, and not the telomerase- and Rad52-independent divergence events discussed above. The length of the undivergent region of each telomere indicates how short the telomere became before being extended by telomerase. It has previously been shown that telomeres with undivergent regions less than 125 bp in length do occur even in wild-type cells [49]. We confirm this observation and also find no change in the frequency of these truncation events in the absence of Rad52 (Fig 2). This suggests that Rad52 does not prevent telomere truncation events, although it may have a role in repairing such truncations.
Telomerase-negative strains lacking Pol32 have previously been shown to exhibit an accelerated rate of senescence [22], indicating the importance of BIR during senescence. Thus, the function of Rad52 in preventing accelerated senescence may be to promote repair of truncated telomeres through Pol32-mediated BIR. If the accelerated senescence of an est2Δ rad52Δ mutant is due to the role of Rad52 in BIR, then est2Δ rad52Δ and est2Δ pol32Δ mutants, derived from the same parental diploid, should have similar rates of senescence. Interestingly, we find that est2Δ rad52Δ mutants senesce faster than est2Δ pol32Δ mutants (Fig 3), indicating that although BIR is important to prevent accelerated senescence, other Rad52-mediated activities are also required.
To further dissect the function of Rad52 at telomeres in the absence of telomerase, we used a specific class of rad52 mutants (class C mutants, specifically rad52-Y66A and rad52-R70A) that are proficient for mitotic recombination but defective in DNA strand annealing and the repair of DSBs [30–32]. The efficiency of BIR is reduced only 2.7-fold in class C mutants [32], whereas rad51Δ mutants exhibit a ~140-fold reduction using the same assay [16], suggesting that class C mutants can perform Rad51-dependent BIR. We find that est2Δ rad52-R70A and est2Δ rad52-Y66A double mutants senesce faster than est2Δ single mutants (Fig 4A, est2Δ vs. est2Δ rad52-R70A, p < 10−6; Fig 4B, est2Δ vs. est2Δ rad52-Y66A, p = 0.003; Fig 4C, est2Δ vs. est2Δ rad52-Y66A, p = 0.001). Interestingly, survivors generated from est2Δ rad52-R70A and est2Δ rad52-Y66A strains are all type I (Fig 4D and Fig 5D). Since type II survivors grow better than type I survivors, survivors generated from a liquid culture senescence assay, as done here, should all be type II [52] unless the strain in question has a defect in forming type II survivors. In our experiments, 9 out of 9 survivors generated from est2Δ mutants, examined 6 population doublings (PDs) after they had recovered from the point of maximum senescence (for each est2Δ mutant, this means the first time point after the point of maximum senescence in the liquid culture senescence assays), were type II. Deleting RAD51 in telomerase-null cells blocks the formation of type I survivors [12]. We find that est2Δ rad51Δ rad52-Y66A triple mutants cannot form any survivors (Fig 4B), supporting that the strand annealing activity of Rad52 is needed to perform Rad51-independent BIR and to form type II survivors.
The rad52 class C mutants behave similarly to rad59Δ, which is also defective for Rad51-independent BIR and causes telomerase-negative strains to senesce fast and to be unable to form type II survivors [12, 17, 21]. We find that est2Δ rad52-Y66A senesces faster than est2Δ rad59Δ (p = 0.02), and deletion of RAD59 does not enhance the senescence of est2Δ rad52-Y66A (Fig 4C), indicating that rad52-Y66A is epistatic to rad59Δ during senescence. However, rad52-Y66A has a greater effect on senescence than rad59Δ, suggesting that the accelerated senescence of telomerase-null rad52 class C mutants is not solely due to a loss of Rad51-independent BIR. Interestingly, est2Δ rad52-Y66A rad59Δ triple mutants show a defect in the formation of survivors (Fig 4C and S3 Fig). Of the four est2Δ rad52-Y66A rad59Δ followed, three showed a prolonged delay before survivors arose and one never formed survivors at all during the duration of the experiment. This observation implies that, while Rad52 and Rad59 function in the same pathway during senescence, they have nonoverlapping functions with regard to survivor formation, and is consistent with other reports suggesting that Rad59 has Rad52-independent functions [53–55].
Having established that Rad52 has non-BIR-related functions in preventing accelerated senescence, we asked whether Rad52 participates in error-free post-replication repair (PRR) at telomeres during senescence. Error-free PRR is thought to utilize the newly synthesized sister chromatid as a template for DNA synthesis to bypass DNA lesions (reviewed in [56]). It was shown that error-free PRR utilizes recombination proteins in the repair of MMS- and UV-induced DNA damage [57]. Rad5 is a key component of the error-free PRR pathway and absence of Rad5 accelerates senescence in a telomerase-negative strain [22]. We analyzed the rate of senescence of est2Δ, est2Δ rad5Δ, est2Δ rad52-Y66A, and est2Δ rad5Δ rad52-Y66A strains (S4 Fig). We find that est2Δ rad5Δ exhibits accelerated senescence (est2Δ vs. est2Δ rad5Δ, p = 0.001), as previously reported, and est2Δ rad52-Y66A senesces faster than est2Δ rad5Δ (p < 10−6). The est2Δ rad5Δ rad52-Y66A triple mutant senesces the fastest (est2Δ rad52-Y66A vs. est2Δ rad5Δ rad52-Y66A, p = 0.001), appearing to have combined the effects of rad5Δ and rad52-Y66A in an additive manner, suggesting that these mutations affect separate pathways. Consistent with our data, it has been previously reported that rad5Δ rad52Δ mutants exhibit a strong synthetic growth defect that is exacerbated in the absence of telomerase [22], and that Rad5 and Rad52 have independent functions during the bypass of thymine dimers [58]. These results indicate that Rad52-mediated HDR and Rad5-mediated error-free PRR act in at least partially non-overlapping pathways to prevent accelerated senescence.
There are several other possible non-BIR mechanisms through which Rad52 may prevent accelerated senescence, including recombination involving sister chromatids. rad52 class C mutants have been previously reported to be defective in SCR [33]. This study also found that defective regulation of H3K56 acetylation also impairs SCR. H3K56 is acetylated by the histone acetyltransferase Rtt109 [59–61] and deacetylated by the histone deacetylases Hst3 and Hst4 [62, 63]. Both hyper-acetylation (e.g. hst3Δ hst4Δ and H3K56Q mutants) and hypo-acetylation of H3K56 (e.g. rtt109Δ and H3K56R mutants) decrease SCR [33]. We hypothesized that defective SCR may explain the senescence and survivor phenotype of est2Δ rad52 class C mutants. If so, then mutants affecting H3K56 acetylation should behave similarly with respect to senescence and survivor formation. Consistent with this idea, both est2Δ hst3Δ hst4Δ and est2Δ rtt109Δ mutants exhibit accelerated senescence (est2Δ vs. est2Δ hst3Δ hst4Δ, p < 10−6; est2Δ vs. est2Δ rtt109Δ, p = 0.02) and are defective in type II survivor formation (Fig 5). However, unlike est2Δ rad52 class C mutants, which do not form any type II survivors, est2Δ hst3Δ hst4Δ and est2Δ rtt109Δ mutants are defective, but still able to form type II survivors. As mentioned above, survivors generated from a liquid culture senescence assay are typically all type II. Two of seven est2Δ hst3Δ hst4Δ survivors (Fig 5C), and five of ten est2Δ rtt109Δ survivors (Fig 5D), were type II.
hst3Δ hst4Δ rad52-Y66A strains are synthetic lethal (S5A Fig), similar to hst3Δ hst4Δ rad52Δ [64]. In addition, rtt109Δ rad52-Y66A double mutants are synthetic sick (S5A Fig), similar to rtt109Δ rad52Δ double mutants [65]. These results indicate that while Rad52-dependent strand annealing and H3K56 acetylation are both important for SCR, to delay senescence, and for type II survivor formation, they function in different pathways. It is also possible that acetylation of H3K56 is important to delay senescence and promote type II survivor formation independently of its role in SCR.
The importance of Rad52 in delaying senescence and for telomerase-independent telomere maintenance in post-senescence survivors was first described over twenty years ago [9]. While much is now known about the role of HDR in telomerase-independent telomere maintenance in yeast as well as other organisms, including humans, the function of HDR during senescence is less-well understood. Rad52-mediated BIR has previously been implicated in preventing accelerated senescence, and it is thought that both Rad51-dependent and Rad51-independent BIR are involved [21–23]. In this study, we show that non-BIR functions of Rad52, involving recombination between sister chromatids, are also required to delay senescence. We also find that proper regulation of H3K56 acetylation is important in preventing accelerated senescence. We present a model where Rad52-mediated HDR mechanisms act at telomeres during telomere attrition-induced senescence.
We first tried to study the role of Rad52 during senescence with a telomere sequencing assay that has been used to detect telomere recombination events, or more specifically, intertelomeric recombination events and unequal sister telomere recombination events. Using this assay, it was estimated that such recombination events occur at a rate of 0.3% per telomere per generation [34]. Surprisingly, we find that the frequency of these events is not reduced by the deletion of RAD52—in fact, the frequency is even slightly increased (Fig 1). We show that a significant fraction of these events are caused by errors introduced during PCR amplification, propagation in E. coli, and/or DNA sequencing. Our results suggest that intertelomeric and unequal sister telomere recombination events occur at a substantially lower rate than 0.3% per telomere per generation. Furthermore, our findings indicate that data obtained previously with this assay suggesting the preferential recombination of short telomeres in senescing cells may need to be re-examined [37, 41], especially considering that telomere sequencing did not detect an increase in divergence events at an artificially-induced very short telomere [35]. However, it is known that Rad52 is important to act on this very short telomere to delay senescence and Rad52 is preferentially recruited to short telomeres in telomerase-negative cells [19, 35]. In addition, recombination intermediates accumulate as telomeres shorten in tlc1Δ cells [66]. Therefore, Rad52-mediated HDR does preferentially act at short telomeres. It is important to note that our findings do not invalidate the use of the telomere sequencing assay to assess telomere recombination, but one needs to keep in mind that there is a background level of sequence divergence events that are not a result of recombination in vivo. Our data also allow us to conclude that these sequence divergence events are unrelated to the function of Rad52 during replicative senescence. It is formally possible that in the presence of Rad52, sequence divergence events (that are not due to technical artefacts) are the result of Rad52-mediated recombination events, while in the absence of Rad52, the divergence events are the result of Rad52-independent mechanisms. However, we find this possibility unlikely given the remarkably similar telomere sequence divergence profiles of est2Δ and est2Δ rad52 strains.
Replication forks have difficulty progressing through subtelomeric and telomeric sequences, causing forks to stall and collapse [67, 68]. A collapsed replication fork at a telomere would lead to a truncated telomere, and it has previously been shown that such truncations do occur in vivo, and that they are rapidly extended by telomerase [49]. In the absence of telomerase, telomere truncation events are likely repaired by BIR using the untruncated sister telomere as a template (Fig 6).
Eliminating BIR by the deletion of either POL32 or RAD52 in telomerase-null cells leads to a similar phenotype: accelerated senescence and an inability to form survivors [9, 10]. However, our data show that deletion of RAD52 in the absence of telomerase is more severe than deletion of POL32, suggesting that Rad52 has functions in addition to BIR during senescence (Fig 3). Double-strand break repair (DSBR), synthesis-dependent strand annealing (SDSA), and single-strand annealing (SSA) are all well-studied Rad52-mediated HDR mechanisms, but all of these are initiated to repair a two-ended DSB. In the absence of exogenous stress, most DSBs occur during DNA replication, likely via replication fork collapse. A fork collapse would lead to a one-ended DSB, which is converted to a two-ended DSB when a replication fork coming from the opposite direction reaches the site of the fork collapse. However, a replication fork that collapses while traversing a chromosome end would stay one-ended since there is no replication fork coming from the distal end of the telomere. The one-ended DSB can be repaired by BIR, but not DSBR, SDSA, or SSA. Thus, Rad52-mediated DSBR, SDSA, or SSA are unlikely to be involved in delaying senescence.
We find that hst3Δ hst4Δ and rtt109Δ mutants, which cause hyper- and hypo-acetylation of H3K56, respectively, also display rapid senescence in the absence of telomerase (Fig 5). Like rad52 class C mutants, hst3Δ hst4Δ and rtt109Δ strains have defects in SCR [33]. Thus, we hypothesize that, in addition to its function in BIR, Rad52 delays senescence through a mechanism involving recombination of sister chromatids. Error-free PRR utilizes a newly synthesized sister chromatid as a template to replicate past replication fork impediments, so it could be seen as a type of SCR. Rad52-mediated HDR activity has also been implicated in error-free PRR, in both a Rad5-dependent and a Rad5-independent manner [57, 58]. Rad5 localizes to a subset of telomeres during S and G2 phases, even in the absence of exogenous stress, and deletion of RAD5 in telomerase-null cells leads to accelerated senescence [22]. However, we find that rad52-Y66A, which is still proficient in Rad51-dependent BIR, and rad5Δ have additive effects in terms of telomere attrition-induced senescence (S4 Fig), suggesting that if Rad52 participates in error-free PRR to delay senescence, it does so via a mechanism separate from Rad5-dependent error-free PRR. Consistent with this view, we have reported that the Shu complex, which is required for efficient HDR and involved in Rad5-mediated error-free PRR [57], is not important for delaying senescence [69]. It has been suggested that error-free PRR proceeds via a Rad5-mediated pathway when the lesion is on the leading strand template, and a Rad52-mediated pathway when the lesion is on the lagging strand template [58]. We propose that this situation may be occurring at telomeres during senescence (Fig 6).
We believe that replication problems at chromosome ends are amplified as telomeres get shorter. est2Δ rad52Δ cells do not exhibit a growth defect immediately after the loss of telomerase (S6 Fig), indicating that telomeres need to shorten before Rad52 becomes important. As mentioned above, Rad52 and recombination intermediates accumulate at telomeres as they shorten [19, 66]. One possible explanation for increased replication problems at short telomeres is that telomere shortening triggers TERRA transcription [70], which could impede replication because of the replication fork encountering either the RNA polymerase machinery or RNA-DNA hybrids. Increased TERRA transcription and telomeric RNA-DNA hybrids both stimulate recombination at telomeres [39, 40]. Mammalian RAD51 and BRCA2, which performs many of the functions of yeast Rad52 [71], are also required for proper telomere maintenance [72, 73], indicating that the importance of HDR at telomeres is highly conserved throughout evolution.
Standard yeast media and growth conditions were used [74, 75]. All yeast strains used in this study are RAD5 derivatives of W303 [76, 77] and are listed in Table 1. Telomeres of 166 bp, 213 bp, and 230 bp, amplified by telomere PCR, were cloned into the pCR-Blunt vector from the Zero Blunt PCR Cloning Kit (Invitrogen) to generate plasmids pCC3, pCC6, and pCC2, respectively. pCC3 and pCC2 were cut with EcoRI, and the telomere-containing fragment in each was subcloned into EcoRI-cut pRS306 (ATCC) to generate plasmids pCC9 and pCC10 (two isolates containing 166 bp-long telomere sequences), and pCC7 and pCC8 (two isolates containing 230 bp-long telomere sequences). pCC9, pCC10, pCC7, and pCC8 were cut with NcoI and transformed into yeast strain W9100-12C to make CCY36, CCY37, CCY34, and CCY35, respectively. RAD52 was then deleted in CCY36 and CCY35 to generate CCY47 and CCY46, respectively.
Telomere VI-R was amplified by PCR using Phusion High-Fidelity DNA Polymerase (New England Biolabs), essentially as previously described [49]. Telomere PCR products were purified using a QIAquick Gel Extraction Kit (Qiagen), cloned using a Zero Blunt PCR Cloning Kit or a Zero Blunt TOPO PCR Cloning Kit (Invitrogen), and transformed into One Shot TOP10 Chemically Competent E. coli (Invitrogen). Individual clones were sequenced by GATC Biotech (except for Fig 2, where sequencing was performed by GENEWIZ), and the resulting data were analyzed using Sequencher software (Gene Codes). The Sequencher files are included as Supporting Information. Excel files recording telomere sequence divergence data are included as S1 Dataset (for Fig 1) and S2 Dataset (for Fig 2). For each set of sequences, the longest telomere without divergent sequence was used as a reference telomere to which all other telomeres are compared to determine whether divergence has occurred. A sequence was determined to be non-divergent if it matches perfectly to the consensus, if it contains single point mutations, or if it contains insertions or deletions of 6 nucleotides or less.
Cells used for live-cell imaging were cultured in synthetic complete media. Microscopy was performed using a DeltaVision Deconvolution Microscope (Applied Precision) with InsightSSI, an Olympus UPLS Apo 100x oil objective with 1.4 numerical aperture, and a CoolSNAP HQ2 camera.
Liquid culture senescence assays were performed as previously described [37, 79]. All senescence assays started with the sporulation of est2Δ/EST2 heterozygous diploids. With the exception of S3 Fig, senescence data were plotted with PDs on the x-axis, not time (i.e. days), because telomere shortening is a function of cell division and not time. Moreover, using PDs as a metric prevents slow growth associated with a particular mutation to be mistakenly interpreted as having an effect on senescence. For clarity, telomerase-positive control strains for each experiment are shown in separate graphs (S5B and S7 Figs). We performed an unpaired two-tailed t-test to evaluate the difference in PDs at maximum senescence between two strains.
Genomic DNA was isolated using a Wizard Genomic DNA Purification Kit (Promega), digested with XhoI restriction endonuclease, separated by agarose gel electrophoresis, transferred to a Hybond-N+ membrane (GE Healthcare), and hybridized to a telomere-specific (5′-CACCACACCCACACACCACACCCACA-3′) digoxigenin-labeled probe.
|
10.1371/journal.pntd.0002098 | Repeated Schistosoma japonicum Infection Following Treatment in Two Cohorts: Evidence for Host Susceptibility to Helminthiasis? | In light of multinational efforts to reduce helminthiasis, we evaluated whether there exist high-risk subpopulations for helminth infection. Such individuals are not only at risk of morbidity, but may be important parasite reservoirs and appropriate targets for disease control interventions.
We followed two longitudinal cohorts in Sichuan, China to determine whether there exist persistent human reservoirs for the water-borne helminth, Schistosoma japonicum, in areas where treatment is ongoing. Participants were tested for S. japonicum infection at enrollment and two follow-up points. All infections were promptly treated with praziquantel. We estimated the ratio of the observed to expected proportion of the population with two consecutive infections at follow-up. The expected proportion was estimated using a prevalence-based model and, as highly exposed individuals may be most likely to be repeatedly infected, a second model that accounted for exposure using a data adaptive, machine learning algorithm. Using the prevalence-based model, there were 1.5 and 5.8 times more individuals with two consecutive infections than expected in cohorts 1 and 2, respectively (p<0.001 in both cohorts). When we accounted for exposure, the ratio was 1.3 (p = 0.013) and 2.1 (p<0.001) in cohorts 1 and 2, respectively.
We found clustering of infections within a limited number of hosts that was not fully explained by host exposure. This suggests some hosts may be particularly susceptible to S. japonicum infection, or that uncured infections persist despite treatment. We propose an explanatory model that suggests that as cercarial exposure declines, so too does the size of the vulnerable subpopulation. In low-prevalence settings, interventions targeting individuals with a history of S. japonicum infection may efficiently advance disease control efforts.
| Approximately 1 billion people are infected with one or more helminthes – a class of parasites that can impair physical, mental and economic development. We are interested in whether there exist groups who are repeatedly infected with helminthes over time in areas where treatment is ongoing. Such individuals may be at risk of morbidity and may also serve as parasite reservoirs, making them appropriate targets for disease control programs. We followed two cohorts in rural Sichuan, China in order to evaluate whether the same individuals were repeatedly infected with the water-borne helminth, Schistosoma japonicum. Each participant was tested for infection at enrollment and two follow-up points – all infections were promptly treated. We conducted detailed interviews to assess exposures to S. japonicum. We found infections repeatedly occurred in a subgroup of individuals and this clustering of infections was only partly explained by differences in exposure. This suggests some individuals may be particularly susceptible to S. japonicum infection. Further exploration of the interplay of exposure and susceptibility suggest that as exposure declines, so too does the fraction of the population vulnerable to infection. Helminth control programs that target people with a history of infection may efficiently reduce helminth infections and morbidity.
| Recent multinational efforts to control and eliminate helminthiasis have the potential to dramatically reduce morbidity among the rural poor [1], [2]. Approximately one billion people are infected with one or more helminthes and the health impacts of these infections, including impaired growth, cognitive development and work capacity are substantial and poverty reinforcing [3]–[5]. Population-level interventions are the recommended strategy in areas where infection prevalence and morbidity are high [6], but as infections decline, how should limited disease control resources be allocated in order to sustain disease control achievements?
We are interested in whether there exist high-risk subpopulations for helminth infection, as such individuals may not only be particularly vulnerable to morbidity, they may also play a key role in sustaining transmission in regions where control efforts have reduced but not eliminated helminthiasis [7]. In many infectious disease transmission systems, a few individuals are responsible for a disproportionate number of future infections: control efforts targeting such superspreaders can efficiently reduce disease transmission compared to randomly allocated or population-based control efforts [8], [9]. In the case of helminthiasis, helminthes typically are aggregated in a population such that at any point in time, a few individuals harbor a large number of worms and therefore may be responsible for a large number of future infections [9]. If the same individuals are repeatedly infected, this suggests the presence of high-risk groups for helminthiasis – groups that may serve as persistent parasite reservoirs in the presence of on-going treatment and control efforts. Prior research suggests such high-risks groups may exist: for example, past infection with the water-borne helminth, Schistosoma sp. is a positive predictor of subsequent infection [10]–[13].
What mechanisms might promote the aggregation of infections in a few individuals? The cross-sectional clustering of helminthes in a population has largely been attributed to differential pathogen exposure – highly exposed individuals are most likely to harbor greater pathogen loads [9], [14]. If we assume an individual's exposure is relatively constant over time, we expect the same, highly exposed individuals will be repeatedly infected over time. Host susceptibility to infection may also favor repeated infections in a particular subpopulation. Host genetics play a role in susceptibility to soil-transmitted helminthiases and schistosomiasis, likely via variations in genes regulating immune function, including, in the case of Schistosoma sp., Th2 response [15]–[18]. In contrast to exposure and host-susceptibility, exposure-dependent immunity should protect highly infected individuals at a given time point from subsequent infection, resulting in a disaggregation of infections across the population over time. Age-dependent immunity should concentrate infections in vulnerable age groups, leading to time-limited membership in high-infection subpopulations.
We examined longitudinal patterns of infection with the water-borne helminth, S. japonicum, in two cohorts in order to assess the aggregation of infections in the same individuals over time and, if present, the extent to which aggregation can be attributed to exposure vs. host-susceptibility.
We followed two cohorts of rural residents in Sichuan, China drawn from hilly regions where schistosomiasis is associated with irrigated agriculture. Cohort 1 is composed of 424 individuals from 10 villages located in Xichang County, in southwest Sichuan, monitored from 2000 to 2006, a region where S. japonicum infection prevalence and intensity has historically been high. Cohort 2 is composed of 400 individuals from 27 villages in 2 counties in Sichuan province where schistosomiasis reemerged following reduction of human infection prevalence below 1%, a benchmark for schistosomiasis transmission control [19]. Individuals in the second cohort were monitored from 2007 to 2010. In each cohort, we tested all participants for S. japonicum infection at enrollment, treated all infections and conducted detailed exposure assessments. Participants were tested for incident infection at two follow-up points.
In fall 2000, we conducted S. japonicum exposure and infection surveys in 20 villages in Xichang County [20]. All residents were invited to participate in S. japonicum infection surveys (individuals age 4–60 were targeted, but infection testing was open to people of any age). A 25% random sample of residents, stratified by village and occupation, was interviewed about water contact behaviors at the same time as the infection surveys. Individuals were asked to report the frequency and duration of contact with surface water sources while conducting the following activities: washing clothes or vegetables, washing agricultural tools, washing hands and feet, playing or swimming, irrigation ditch operation or maintenance, rice planting, rice harvesting, and fishing; for each month from April to October (Supporting Information S1 in Text S1). Infection surveys were repeated in 2002 and 2006 in ten villages with high infection prevalence in 2000 (range 12.9 to 72.3%). This cohort includes all individuals from the 10 follow-up villages who completed the water contact interview and were tested for infection all three years. Infection status and intensity at enrollment did not differ between cohort members that were lost to follow-up and those with complete data, but those who were lost to follow-up reported less water contact and were younger, on average. Details of cohort selection and retention are provided in Figure S1 and Table S1 in Text S1.
In fall 2007, a cross sectional survey was conducted in 53 villages in three counties where S. japonicum reemerged following attainment of national transmission control criteria [21]. All residents age 6 to 65 were invited to participate in S. japonicum infection surveys. In May 2008, a magnitude 7.9 earthquake in Sichuan severely impacted one of the three selected counties, forcing us to limit follow-up studies to the two other counties. For efficiency, water contact behaviors were assessed using a stratified random sample of individuals based on 2007 infection status. All individuals who tested positive for S. japonicum in 2007, and, for each infected person, five people randomly drawn from the same village who tested negative for S. japonicum in 2007, were selected for participation in a survey of water contact behaviors. Interviews about water contact patterns were conducted monthly, from June to October 2008. At each interview, participants were asked to report the frequency and duration of water contact activities in the past two weeks including washing laundry, washing vegetables, washing agricultural tools, washing hands or feet, playing or swimming, ditch cleaning and repair, rice planting, rice harvesting, fishing, and collecting water for drinking or cooking. During the first interview, participants were also asked to report water contact behaviors during the May rice planting season, as, due to earthquake relief efforts, no interviews were conducted in May. Nobody reported water contact while collecting water for drinking and cooking, and this behavior was excluded from analyses. For comparability with cohort 1, washing laundry and washing vegetables were combined into a single water contact measure. Participants were tested for S. japonicum infection again in 2008 and 2010.This cohort includes all individuals who were tested for infection all three years and completed the water contact interview. As was the case for cohort 1, baseline infection status and intensity did not differ between cohort 2 members that were lost to follow-up and those with complete data, but those who were lost to follow-up reported less water contact, were more likely to be male and were younger, on average. Details of cohort selection and retention are provided in Figure S2 and Table S1 in Text S1.
As some members of cohort 2 did not complete all monthly water contact interviews, missing water contact measures were imputed using multiple imputation by chained equations [22], [23]. Multiple imputation avoids bias presented by the exclusion of incomplete cases. Imputation is based on the assumption that data are missing at random, and that missing data can be explained by other measured variables [24]. We imputed water contact minutes by month and activity using all other water contact measures, as well as age, sex, and village of residence. During the monthly interviews, participants were also asked to report the number of days they spent outside of their village and distance traveled in the past month. As travel may influence water contact patterns, travel was also included in the set of existing data used to impute missing values. The duration of water contact was imputed using predictive mean matching. Because nobody reported water contact from fishing in October or rice harvesting in June, all individuals missing these variables were assumed to have zero water contact for this exposure. Participants with one or more missing values were more likely to be younger and live in county 1, but did not otherwise differ substantially from participants with complete data (Table S2 in Text S1).Ten imputed datasets were generated. We calculated the mean of each imputed value for use in the predictive models described below. Before imputation, 5.0% of the water contact measures were missing: 71 participants (18%) did not complete all monthly interviews and 13 participants (3%) were interviewed each month but did not answer all survey questions.
All questionnaires in both cohorts were administered in the local dialect by trained staff at the Institute of Parasitic Diseases (IPD), Sichuan Center for Disease Control and Prevention and the county Anti-schistosomiasis Control stations.
During each infection survey, participants were asked to submit three stool samples, one each from three consecutive days. Each sample was analyzed using the miracidia hatching test: approximately 30 grams of stool was filtered, suspended in aqueous solution and examined for miracidia according to Chinese Ministry of Health protocols [25]. In addition, one sample from each participant was analyzed using the Kato-Katz thick smear procedure: three slides were prepared using 41.7 mg homogenized stool per slide and examined for S. japonicum eggs by trained technicians [26]. Infection intensity, in eggs per gram of stool (EPG), was calculated as the total number of S. japonicum eggs divided by the total sample weight. In 2002, only one stool sample was collected per person in cohort 1, and this sample was analyzed using both the miracidia hatching test and the Kato-Katz thick smear procedure. After each infection survey, all individuals testing positive for S. japonicum were promptly notified and provided treatment with 40 mg/kg praziquantel by health workers at the county anti-schistosomiasis control stations.
The research protocols and informed consent procedures and were approved by the Sichuan Institutional Review Board and the University of California, Berkeley, Committee for the Protection of Human Subjects. In cohort 1, all participants provided oral informed consent, documented by IPD staff, before participating in this study. Oral consent was obtained due to the high prevalence of illiteracy, and because the survey procedures used were similar to those used by IPD for schistosomiasis surveillance. In cohort 2, all participants provided written, informed consent before participating in this study. Minors provided assent and their parents or guardians provided written, informed permission for them to participate in this study.
We examined the extent to which S. japonicum infections repeatedly occur in the same individuals in regions where schistosomiasis case detection and treatment is ongoing. For each cohort we defined three time points: baseline (T0), the first follow-up infection survey (T1) and the second follow-up infection survey (T2). We estimated the ratio of the observed proportion of the population with of two consecutive infections at T1 and T2 (ODI), to the predicted proportion of the population with two consecutive infections at T1 and T2 (PDI).
The simplest model of PDI is based solely on the probability of infection at T1 and T2, such that where indicates S. japonicum infection status at time point x. Because all infections were treated at each time point, the probability of infection at Tx is the incidence of infection from T(x-1) to Tx multiplied by the elapsed time between T(x-1) and Tx, which is equal to the prevalence of infection at Tx. Note that at T0, we know the prevalence, but not the time elapsed since last treatment, which may vary by individual, and therefore can only estimate the probability of infection at T1 and T2. Our estimates of infection probability assume all infections, defined as the presence of adult S. japonicum worm pairs, are detected and successfully treated at each time point. Using this prediction model, if , this suggests that there exists a subset of individuals that are repeatedly infected with S. japonicum.
A more complex model of PDI accounts for exposure, as individuals who are repeatedly infected may be those who are most highly exposed to S. japonicum cercariae. In this case, we estimate where is S. japonicum cercarial exposure. Using this exposure-based prediction model, if , this suggests that S. japonicum infections repeatedly occur in a subset of individuals in the population for reasons not attributable to the exposure variables in the statistical model. S. japonicum cercarial exposure is determined by human behaviors that put people in contact with potentially contaminated water sources (primarily irrigation ditches and ponds), and by cercarial concentrations at the site of contact. We accounted for human behavior using questionnaire derived estimates of month- and activity-specific water contact duration. Cercarial concentration can vary over space and time due to the non-uniform distribution of the intermediate snail host and because cercarial shedding is affected by temperature, diurnal patterns and reservoir host species [27]–[29]. Currently, practical, field deployable methods for measuring cercarial concentrations are lacking. A mouse bioassay exists, in which sentinel mice are dermally exposed to surface water, then sacrificed and examined for S. japonicum worms, approximately 45 days post-exposure (allowing time for the parasite to mature inside the host). The mouse bioassay is not only resource intensive but, in low-prevalence settings, has limited sensitivity and, while new molecular methods offer promise, they have yet to be widely deployed [30], [31]. We used several proxies for cercarial concentration in our infection prediction models. We included village infection prevalence at T0, based on the assumption that villages with more infected individuals at enrollment have the potential for greater cercarial concentrations. In cohort 2, we also included county in the infection prediction model, as control measures which may impact cercarial concentration such as application of moluscicides are administered at the county level (all participants in cohort 1 are from a single county). To account for temporal variation in cercarial concentration, we included the year of infection testing. Additionally, we included age and sex to account for potential differences in the location of water contact (concentration) and the reporting of water contact activities (behavior) by age and sex.
The first step in estimating PDI requires a model that predicts infection status at a given time point based on exposure: . However, given the large number of predictor variables and the potentially complex, nonlinear relationships between exposure and infection, any single arbitrary parametric model one might choose will lead to an unknown degree of bias in the estimate of and, ultimately, PDI [32]. To minimize this problem we used a machine-learning algorithm, known as the Super learner as implemented in R [33]. In essence, this procedure estimates based on a convex combination of a number of different modeling algorithms (some simple parametric models, some highly data adaptive, generically called learners). In this case, the learners include random forests [34], k-nearest neighbor classification [35], elastic net regression [35], generalized linear models, stepwise regression and generalized boosted regression [36]. Cross-validation is used to determine the optimal combination of learners, that is the combination that maximizes the cross-validated fit. It has been shown that the Super learner estimate is asymptotically equivalent to the estimator that would come closest to the truth if the truth were known (called the Oracle selector), even if a very large number of competing models were used. In addition, in the unlikely case that the true model is a simple parametric model, then Super learner achieves nearly the same performance as a simple parametric estimation procedure (a parametric Oracle). From a practical point of view, Super learner replaces the usual ad hoc exploration of the adequacy and fit of various candidate models with a machine-based procedure that produces a robust, replicable, and theoretically defensible estimate.
We excluded from the set of exposure variables water contact variables for which <20% of the population reported any water contact. Models were fit separately for each cohort. An individual's infection probability was calculated for each year (T1 and T2) using the selected model, and the probability of two consecutive infections was calculated as the product of the infection probabilities at T1 and T2.
All estimates of observed and predicted infections were weighted to account for the stratified sampling used to assess water contact behavior. Each individual in the cohort was assigned a weight equal to the inverse probability of being sampled. Inference was estimated by calculating the probability of the observed number of consecutively infected individuals in the reweighted population (). We assumed follows a binomial distribution where is equal to the number of individuals in the reweighted population and is the probability of two consecutive S. japonicum infections in an individual.
Statistical analyses were conducted using Stata12.0 and R 2.14.1 software.
The demographic characteristics of the two cohorts, reported water contact behaviors and the distribution of S. japonicum infections at enrollment are presented in Table 1. In the 10 villages from which cohort 1 was drawn, mean S. japonicum infection prevalence among all 1,801 residents surveyed was 46.9% (12.9 to 72.3% by village) and intensity, 46.0 EPG (1.1 to 107.9 EPG by village) at enrollment (T0). In the 27 villages from which cohort 2 was drawn, mean infection prevalence among all 1,608 individuals surveyed was 10.6% (1.5 to 42.9% by village) and intensity 2.6 EPG (0 to 10.6 EPG by village). Note that in 3 villages, infections were detected by the miracidia hatching test only, no eggs were detected by the Kato-Katz method, resulting in mean village infection intensities of 0 EPG. In both cohorts, adults were generally farmers with limited formal schooling. The percent of people reporting water contact, and the average duration of water contact varied by month, activity and cohort.
Infection prevalence and intensity at follow-up was low in both cohorts (Table 2). Notably, many individuals who tested positive for S. japonicum had no detectable eggs through the Kato-Katz examination: these individuals were positive via the miracidia hatching test only. In cohort 1, 30% and 27% of the individuals that tested positive for S. japonicum infection at T1 and T2, respectively, had no detectable S. japonicum eggs on Kato-Katz examination. In cohort 2, 55% and 65% of infected individuals at T1 and T2, respectively, had no detectable S. japonicum eggs on Kato-Katz examination.
There were 21 and 20 individuals infected with S. japonicum at both T1 and T2 in cohorts 1 and 2, respectively (Table 3). Consecutive S. japonicum infections at follow-up were 3 and 7 times more common among those who were infected with S. japonicum at T0 than those who were uninfected at T0 in cohorts 1 and 2, respectively. Individuals that were infected at T1 and T2 were not demographically distinct from the cohorts as a whole. The age distributions of individuals with two consecutive infections to those with one or no infections at follow-up are similar (Figure 1). Among those with infections at T1 and T2, mean age at enrollment was 30.1 (range 5–56) and 48.2 (18–63) in cohorts 1 and 2, respectively. In cohort 1, 11 of the 21 twice-infected individuals at follow-up were female and in cohort 2, 8 of 20 were female.
The observed fraction of the population with two consecutive S. japonicum infections was 1.48 times greater than expected in cohort 1, and 5.82 times greater than expected in cohort 2 (Table 4). This concentration of repeated S. japonicum infections in the same individuals is very unlikely due to chance (p = 0.00051 and p = 6.6×10−12 in cohorts 1 and 2, respectively). When we accounted for S. japonicum cercarial exposure, the ratios declined to 1.30 and 2.06 in cohorts 1 and 2, respectively. The excess of individuals with repeated S. japonicum infection, even when accounting for exposure, is highly unlikely due to chance in cohort 2 (p = 0.00056) and unlikely due to chance in cohort 1 (p = 0.013).
In two cohorts from two geographically distinct environments, S. japonicum infections repeatedly occurred in the same individuals over time, following treatment with praziquantel. This clustering of infections occurred even when accounting for exposure, and clustering was particularly strong in cohort 2, a population with low overall infection prevalence and intensity. These findings suggest there exists a subset of individuals within the general population that is particularly vulnerable to S. japonicum infection. Alternatively, this subset of individuals may have uncured infections due to non-compliance or treatment failure. This has important implications for disease surveillance: individuals with a history of S. japonicum infection may serve as appropriate targets for infection monitoring and treatment in low-prevalence environments. In addition, our findings provide evidence for host susceptibility to helminth infections – suggesting some individuals may be more vulnerable to infection given equivalent exposures.
It is possible that individuals who are repeatedly infected with helminthes are simply the most highly exposed individuals in the population. Cercarial exposure is a well-documented determinant of S. japonicum infection [37]–[40]. We found that the ratio of observed to expected prevalence of consecutive infections exceeded unity using an exposure-blind prediction model. This ratio was lower when we included exposure in the prediction models, but still exceeded unity. This suggests some individuals may be repeatedly infected due to their high cercarial exposure, but exposure does not fully explain this phenomenon. S. japonicum exposure is challenging to assess due to the difficulties in quantifying daily human behaviors and the absence of practical methods for directly measuring cercarial concentration, and our prediction models are limited by our ability to accurately measure cercarial exposure. However, the imperfections of our exposure measures are likely offset by the use of an aggressive, data adaptive algorithm to predict S. japonicum infection using over 25 exposure variables. Over-fitting is possible when using such methods, which, in this case, would have a conservative impact on our estimates, pushing observed to expected ratios closer to unity. Therefore, exposure alone is unlikely to explain the observed concentration of repeated schistosomiasis infections in a subset of the population.
More likely, individuals who are repeatedly infected with helminthes may be those who have a sufficiently elevated combination of susceptibility and exposure. We explored the clustering of infections within certain individuals from a mechanistic perspective by postulating that an individual's worm burden, w, accumulated from exposures subsequent to successful praziquantel treatment, can be described at the end of one or more infection seasons as a result of that individual's cumulative exposure to cercariae, E, and the subsequent penetration and development of a fraction of these cercarial hits, α, into adult parasites. That is, w = αE where E is composed of two elements, water contact, S, and cercarial concentration, C. The parameter α, reflecting host susceptibility, is assumed to be a stable property of each individual in the village population and the distribution of C is assumed to be a village property shared by all inhabitants. The water contact measurements described above and cercarial bioassay data collected in conjunction with the prior studies of cohort 1 [27], [37]suggest that the population distribution of E is strongly right skewed as is generally observed to be the case for distributions of w.
Assuming that the distributions of exposure and susceptibility in a population are independent, their joint distribution is depicted in Figure 2. The marginal distribution of exposures, f(E), is for illustrative purposes shown as a negative exponential distribution since multiple cercarial hits are thought to be necessary to lead to a single adult worm. Also for illustration, the marginal distribution of susceptibility, h(α), is shown as symmetric. The line wT = αE is the threshold of infections that are epidemiologically visible which we define here as the minimum worm burden necessary to produce eggs at the lower limit of detection by a combination of the miracidia hatching test and the Kato-Katz method. The fraction of the population susceptible to infection at or above this threshold is that lying to the right of the line α = α*. That is, the probability of an exposure leading to a diagnosis of infection for an individual with an α less than α* is essentially zero given the maximum cercarial exposure in this hypothetical environment. The shaded area depicts the set of exposure-susceptibility combinations that produce detectable infections.
Specification of the two marginal distributions allows the calculation of the distribution of their product, that is, the distribution of worm burden in the population. However, the point here is that, at least in this generic example, the proportion of the population at risk for infection is less than the entire population. That is, the number of individuals susceptible to infection, , in this environment is:Where is the total population size. Hence, if is the observed number of infections, the ratio of prevalence of infection in the susceptible population to the total population is:which is always equal to or greater than unity.
Returning to the re-infection issue, suppose the population is exposed in an unchanging environment, treated annually with praziquantel at T = 0, T = 1, and T = 2, and infection assessed at the end of year 1 and year 2. Since the same population is at risk of infection with the same marginal distribution of exposure in both years, and this population is less than the entire population, the observed number of repeated infections will be greater than that expected based on infections occurring randomly in the entire population. It follows that the ratio of observed re-infections to the expected number, if distributed randomly in the entire population, is simply the square of the foregoing equation:Moreover, as the fraction of exposure-susceptibility combinations that produce infection decreases, α* and this ratio both increase. Hence, the simple model of the infection process with individual differences in susceptibility to infection, depicted in Figure 2,provides a heuristic explanation of the epidemiological finding that the ratio of observed to expected re-infections increases as prevalence of infection decreases. Clearly, more refined analyses are possible that address a more rigorous definition of α*, take distributional assumptions into account, or explore the effect of variability in individual water contact. We will further address these and related determinants of transmission in the low-risk environment via an individually-based stochastic model which will be the subject of a future report. In addition, it is possible to estimate the proportion of susceptibles in a population via a statistical innovation using a model selection procedure like SuperLearner in the context of a latent mixture model, where the susceptibility status is latent – an approach that we will pursue in the future.
The factors that govern α are not fully characterized for schistosomiasis or other helminthiases. However, there is substantial evidence that immune function, particularly the ability to mount antigen-specific IgE response, can confer host resistance to schistosomiasis as well as other helminthiases [10], [11], [41]–[43]. Immune response is likely attributable to a combination of past exposure, treatment and host genetics [16], [44]–[46]. Physical characteristics such as skin thickness may also play a role in determining host resistance or susceptibility. As these genetic and immunological pathways are further elucidated, the definition of α may be further refined.
Alternatively, it is possible that the individuals who appear to be repeatedly infected with S. japonicum do not have new infections, but instead have residual, uncured infections that persist despite treatment. Praziquantel is the primary drug used to treat schistosomiasis infections, and resistance is an ongoing concern, particularly in areas where the drug has been used extensively. In China, praziquantel has been widely administered since the 1990s through mass and targeted treatment campaigns. Currently, there is no evidence of population-level resistance to S. japonicum, S. haematobium or S. mansoni, but praziquantel resistant laboratory isolates have been identified [47]–[51]. It is possible that praziquantel kills some but not all parasites, resulting in an incomplete cure. Repeated dosing with praziquantel may enhance treatment efficacy, particularly for individuals with high infection intensities [52]. While infection intensities in our two cohorts were generally low, we cannot rule out the possibility that what appear to be repeated infections are, in fact, infections that were not cured by praziquantel treatment.
Uncured S. japonicum infection may also be the result of poor adherence to drug treatment. As schistosomiasis morbidity declines, it is possible that so too do the perceived risks of infection and willingness to take praziquantel. Praziquantel has an excellent safety record and is appropriate for mass drug distribution, even in very young populations [53] but the drug has a bitter taste and can cause transient side effects, including nausea and dizziness. In a recent survey, 33% of people said such side effects impacted their ability to work [49]. We have found a high degree of self-reported treatment adherence (>90%) in surveys of 236 people drawn from the same villages as cohort 1 (surveyed in 2007) and 686 people drawn from the same villages as cohort 2 (surveyed in 2008), but other studies have documented poor compliance with mass-treatment campaigns for helminthiasis [54], [55]. Our findings underscore the importance of continued monitoring of treatment effectiveness, including both drug resistance and population perceptions of the risks and benefits of treatment. Methods capable of distinguishing new from residual infections could advance our understanding of treatment efficacy and drug adherence.
Our findings underscore surveillance challenges in areas where worm burdens are low. While individuals with high worm burdens have the potential to contribute a large number of future infections, our prior work suggests that even modest parasite inputs are sufficient to sustain schistosomiasis transmission [7]. In China, surveillance and elimination efforts are made more complex as there are at least 40 competent mammalian host species for S. japonicum, and bovines are suspected to be key reservoirs in some areas [56]. Thus the ability to identify humans and, in the case of S. japonicum, other mammalian hosts with low-intensity helminth infections may be crucial to efforts to prevent the reemergence of helminth infections in areas where disease control efforts have successfully lowered infections and morbidity. Many of the individuals who tested positive for S. japonicum in our study had worm burdens below the limit of detection of the Kato-Katz assay, the schistosomiasis diagnostic method recommended by the World Health Organization [6]. Immunoassays generally have high sensitivity, but it can be difficult to distinguish past from current infections, which is particularly problematic when attempting to identify residual infections in regions with previously high infection prevalence and intensity [57], [58]. While new methods offer promise [59], the current lack of practical, highly sensitive diagnostics is a barrier to the long-term control of helminthiases [1], [60].
As China aims to eliminate schistosomiasis and global efforts are launched to eliminate a number of helminthiases, the success of such efforts may hinge, in part, on the ability to identify reservoirs of infection and reduce the potential of such reservoirs to generate future infections. Our findings suggest that there exist an identifiable, high-risk subpopulation for S. japonicum infection. Due to high exposure, host susceptibility or treatment failure, these individuals are potential future reservoirs of S. japonicum. Further, as infection prevalence declines, and with it, cercarial exposure, we expect the fraction of the population that is susceptible to S. japonicum infection to decline. Thus, as regions approach disease control goals, targeted interventions may prove efficient and effective. In low-prevalence regions, individuals who test positive for S. japonicum should be tested regularly and provided pharmaceutical treatment and transmission-blocking interventions such as improved household latrines [56], [61].
|
10.1371/journal.pntd.0002381 | Metabonomics Reveals Drastic Changes in Anti-Inflammatory/Pro-Resolving Polyunsaturated Fatty Acids-Derived Lipid Mediators in Leprosy Disease | Despite considerable efforts over the last decades, our understanding of leprosy pathogenesis remains limited. The complex interplay between pathogens and hosts has profound effects on host metabolism. To explore the metabolic perturbations associated with leprosy, we analyzed the serum metabolome of leprosy patients. Samples collected from lepromatous and tuberculoid patients before and immediately after the conclusion of multidrug therapy (MDT) were subjected to high-throughput metabolic profiling. Our results show marked metabolic alterations during leprosy that subside at the conclusion of MDT. Pathways showing the highest modulation were related to polyunsaturated fatty acid (PUFA) metabolism, with emphasis on anti-inflammatory, pro-resolving omega-3 fatty acids. These results were confirmed by eicosanoid measurements through enzyme-linked immunoassays. Corroborating the repertoire of metabolites altered in sera, metabonomic analysis of skin specimens revealed alterations in the levels of lipids derived from lipase activity, including PUFAs, suggesting a high lipid turnover in highly-infected lesions. Our data suggest that omega-6 and omega-3, PUFA-derived, pro-resolving lipid mediators contribute to reduced tissue damage irrespectively of pathogen burden during leprosy disease. Our results demonstrate the utility of a comprehensive metabonomic approach for identifying potential contributors to disease pathology that may facilitate the development of more targeted treatments for leprosy and other inflammatory diseases.
| Leprosy is caused by a mycobacterium that has a predilection for the skin and nerve cells, and the disease is treated with a combination of antibiotics (multidrug therapy, MDT). Nerve damage caused by the infection may lead to permanent disabilities, and can happen even during MDT and subsequent to patient release. Therefore, a more comprehensive understanding of the interaction between the leprosy bacillus and humans is mandatory in order to develop new tools for better disease control and management. Aiming to understand more about the effects of leprosy on human metabolism, we analyzed the chemical composition of sera from leprosy patients before and after MDT. Our results show that specific classes of molecules are affected by the infection, and that MDT can partially reverse these effects. In particular, lipids related to polyunsaturated fatty acid metabolism and known to play a role in the host's defense mechanisms were highly affected during the disease. A complete understanding of all the steps in this process may open new avenues for leprosy treatment with consequent prevention of neuropathy.
| Leprosy, a chronic infectious disease caused by the obligate intracellular bacterium Mycobacterium leprae, remains a major source of morbidity in developing countries [1]. The disease affects mainly the skin and the peripheral nervous system, in which the leprosy bacillus is preferentially found inside macrophages and Schwann cells (reviewed in [2]). Multidrug therapy (MDT), a combination of antibiotics that are very effective in eliminating M. leprae, was introduced by WHO in the early eighties. However, despite efforts to treat registered leprosy patients, the number of new cases reported globally remains stable and high (about 200,000/year). The disease is still considered a public health problem in several countries. In Brazil, the detection rate remains high and stable at approximately 40,000 new cases annually [1]. Moreover, nerve damage may progress during MDT itself and even subsequent to patient release, due mainly to the occurrence of acute immune-inflammatory episodes known as leprosy reactions. Therefore, new strategies and approaches need to be developed in order to decrease disease morbidity and fully eradicate leprosy as a public health problem.
Also known as Hansen's disease, leprosy manifests as a spectrum of clinical forms in correlation with the nature and magnitude of the innate and adaptive immune responses generated during infection. At one extreme of the spectrum, individuals with polar tuberculoid (TT) leprosy have few lesions and manifest a contained, self-limited infection in which scarce bacilli are detected due to the generation of a strong cellular immune response against M. leprae. At the other end, lepromatous leprosy (LL) is a progressively disseminating disease characterized by extensive bacterial multiplication within host cells and low cell-mediated immunity to the pathogen. Between these two poles are the borderline forms (characterized by their intermediate clinical and immunological patterns), commonly referred to as borderline tuberculoid (BT), borderline borderline (BB), and borderline lepromatous (BL) in accordance with their proximity to either one of the spectral extremes (reviewed in [2]).
Leprosy is a complex disease, and is essentially restricted to human beings. Despite considerable research efforts over the last decades, our understanding of the mechanisms that govern leprosy pathogenesis remains limited. The unique features of the leprosy bacillus have contributed to the slow progress in our knowledge of leprosy. One peculiar characteristic of M. leprae is its extremely long generation time, estimated to be nearly 2 weeks. This slow growth rate results in long incubation periods (2–10 years) and very slow development of pathology and clinical evolution (reviewed in [2]). In the absence of an animal experimental model that mimics the disease in humans, progress in our knowledge of leprosy pathogenesis relies on observations obtained from infected populations and on analyses of clinical samples collected directly from leprosy patients. However, continuing improvements in analytical technologies and recent developments of sensitive high-throughput techniques are now opening a new opportunity to study this ancient disease in order to suggest new strategies for leprosy prevention and treatment. Of note, techniques that identify and quantify multiple small metabolites (<1,500 Da) in complex biological samples have been recently developed, giving rise to the field of metabolomics (or metabonomics). Metabonomics has been successfully applied to different biofluids and tissue types, revealing their biochemical composition in different pathological conditions [3], [4], [5].
The complex interplay between pathogens and their hosts has profound effects on host metabolism during infection. Since the tuberculoid and lepromatous forms of leprosy constitute different responses of the host to M. leprae infection, we hypothesized that host metabolism in response to infection would be distinct in these different clinical forms of the disease. Even though M. leprae is an obligate intracellular parasite, patient plasma/serum offers an important window for detecting metabolic modulation since blood contains many molecules that are released by different tissues in response to infection. A recent metabolomic study of human serum has identified and quantified more than 4,000 metabolites generating the Human Serum Database [6].
To explore the perturbations in the human metabolome associated with M. leprae infection, we analyzed the repertoire of metabolites present in serum samples of leprosy patients. We used direct-infusion ultrahigh resolution Fourier transform ion cyclotron resonance mass spectrometry (DI-FT-ICR-MS), a powerful technique that allows the presumptive identification and relative quantification of thousands of metabolites with high mass accuracy and without the need for extensive sample preparation [7]. Our results indicate a marked modulation of omega-6 and omega-3 polyunsaturated fatty acids (PUFA) metabolism during M. leprae infection, which disappears after MDT. Effects of M. leprae infection on PUFA metabolism were confirmed by measurements through enzyme-linked immunoassays using serum, which showed significantly higher levels of prostaglandin (PG) D2 and E2 (PGD2 and PGE2), lipoxin A4 (LXA4) and resolving D2 (RvD2) in untreated leprosy patients. Moreover, high-throughput metabolic profiling of skin specimens revealed an abundance of lipase products in LL patients, such as polyunsaturated fatty acids and lysolecithin, corroborating the serum metabolome data. This study demonstrates the power of metabonomics to unravel metabolic modulation during infection and provides the opportunity to identify novel therapeutic targets and biomarkers for leprosy.
The Ethics Committee of the Oswaldo Cruz Foundation approved all procedures described in this study. All subjects, none of which were minors, provided informed written consent.
Leprosy patients (29 LL and 29 BT) were recruited on a volunteer basis from the Leprosy Outpatient Unit (Oswaldo Cruz Foundation, Rio de Janeiro, RJ, Brazil). Patients were classified with leprosy according to the criteria of Ridley and Jopling [8], and serum samples were taken before and right after MDT conclusion (without fasting). Skin biopsy specimens (6-mm punch) were also collected from LL and BT patients before treatment and were used for metabolite extraction. The baseline characteristics of each group of individuals included in the study are shown in Table 1. None of the patients were under anti-inflammatory therapy at the time of serum and biopsy specimen collection.
Serum samples were thawed and 200 µL of serum were extracted overnight at −20°C in 2-mL tubes with 750 µL of methanol/chloroform (2∶1, v/v) followed by vortexing and centrifugation at 1,500×g for 5 minutes at 4°C. It is important to note that serum samples were never thawed before the metabonomics analysis described below was performed. The supernatants were carefully transferred to new tubes and the samples were extracted once again with 500 µL of methanol/chloroform/water (2∶1∶0.8, v/v/v), vortexed, and centrifuged as described before [7]. The extracts were pooled, concentrated in a speedvac evaporator and dried under a nitrogen stream. For metabolite extraction of frozen biopsies, specimens were thawed on ice, mechanically disrupted, and extracted with chloroform/methanol/water (1∶2∶0.8, v/v/v) [9]. Samples were then partitioned with chloroform and methanol (2∶1, v/v), according to the standard procedure of Folch et al. (1957) [10]. Pellets were extracted again with acetonitrile (10 µL of acetonitrile for each mg of initial tissue) by vortexing for 10 minutes. Samples were clarified by centrifugation at 16,000×g for 5 minutes, all phases were combined and extracts were dried and saved for further analysis.
For metabolic profiling of sera, dried extracts were suspended in 70% methanol (100 µL for each µL of sample), vortexed, and cleared by centrifugation. Supernatants were collected and used as described below. Extracts were diluted 1∶3 with 70% methanol containing either 0.2% formic acid (for positive ionization mode) or 0.2% ammonium hydroxide (for negative ionization mode) and spiked with predefined amounts of an ES tuning mix solution as the internal standard for mass calibration. For metabolic profiling of skin biopsies, dried extracts were resuspended in 60% acetonitrile (100 µL per 10 mg of sample), vortexed, sonicated and cleared by centrifugation. Extracts were diluted 1∶3 in ESI standard solutions containing either 0.2% formic acid (positive ion mode) or 0.5% ammonium hydroxide (negative ion mode). Solutions were then infused, using a syringe pump (KDS Scientific, Holliston, MA), at a flow rate of 2.5 µL per minute, into a 12-T Apex-Qe hybrid quadrupole-FT-ICR mass spectrometer (Bruker Daltonics, Billerica, MA) equipped with an Apollo II electrospray ionization source, a quadrupole mass filter, and a hexapole collision cell. Data were recorded in positive and negative ion modes with broadband detection and an FT acquisition size of 1,024 kilobytes per second within an m/z range of 150 to 1,100. Other experimental parameters were: capillary electrospray voltage of 3,600 to 3,750 V, spray shield voltage of 3,300 to 3,450 V, source ion accumulation time of 0.1 second, and collision cell ion accumulation time of 0.2 second. To increase detection sensitivity, survey scan mass spectra in positive- and negative-ion modes were acquired from the accumulation of 200 (sera) or 400 (skin) scans per spectrum.
Raw mass spectrometry data were processed using a custom-developed software package, as described elsewhere [7]. Then, data analysis proceeded as previously described [3], [11]. Principal Component Analysis (PCA) was performed using the freely available software Multibase (http://www.numericaldynamics.com/). To identify differences in metabolite composition between BT and LL sera and skin samples, and sera from both groups before and after MDT, we manually selected two groups of metabolites. The first group comprised metabolites that were present in one set of samples but not the other. The second group comprised metabolites present in the two sets of samples being compared, but at different levels. To identify the metabolites in the second group, we averaged the mass intensities of metabolites in each set of samples (BT or LL, before or after MDT) and calculated the ratios between averaged intensities of metabolites from those samples. To assign possible metabolite identities to m/z values present in one group of samples but not the others as well as those m/z showing at least a 2-fold change in intensity between sets of samples, m/z of interest were queried against MassTRIX (version 2, http://metabolomics.helmholtz-muenchen.de/masstrix2/), a free-access software designed to incorporate masses into metabolic pathways using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database (http://www.genome.jp/kegg/).
Lipoxin A4 (LXA4), prostaglandin D2 (PGD2), prostaglandin E2 (PGE2), leukotriene B4 (LTB4) and resolvin D1 (RvD1) levels were measured in serum samples taken from BT (n = 20) and LL (n = 19) patients prior to and after MDT. In addition, measurements of these eicosanoids in serum samples from 10 healthy controls were taken for comparison. We used serum samples that had never been thawed and had been stored at −20°C. Measurements were performed using commercially-available kits according to the manufacturer's instructions. PGD2, PGE2, LTB4 and RvD1 enzyme-linked immunoassay (EIA) kits were purchased from Cayman Chemical (Ann Arbor, USA). The LXA4 EIA kit was from Neogen (Lexington, USA). LXA4 and RvD1 were extracted from serum samples using C18 Sep-Pak columns (Waters; Elstree, UK) before analysis, following the manufacturer's protocol.
Data were analyzed by two-tailed unpaired or paired t tests with 95% confidence intervals or Kruskall–Wallis non-parametric analysis of variance (ANOVA) and Dunn's multiple-range post hoc test, as indicated. Outliers were detected using the Grubbs' test and removed from data sets when indicated.
Serum samples were obtained from 4 BT and 4 LL patients to analyze metabolic alterations during the course of leprosy, and DI-FT-ICR-MS was used to detect and relatively quantify small metabolites in these samples. Such high-throughput analysis yielded a total of 2,565 different m/z (metabolite features) from both BT and LL groups, which were detected from combined positive and negative ion modes (Table S1). A Principal Component Analysis plot (Figure 1A) illustrates the extensive differences in metabolic composition between sera from BT and LL patients. Due to the sample size no analysis could be done on gender and age. Nevertheless, a more refined analysis was then carried out to determine the extent of the metabolic differences between BT and LL leprosy patients. To investigate which of the metabolites detected in both BT and LL samples were present at different levels in these groups, the average intensities of all metabolites were calculated and results from each of the sample groups (BT and LL) were compared. Metabolites that showed changes of 2-fold or more were used for further analyses. Based on this analysis criterion, we found that 684 of the total 2,565 metabolites were present at different levels when comparing samples from BT and LL patients (Table S1). This represents 26.7% of all detected m/z, supporting the notion that an extensive metabolic shift occurs during disease. Metabolite levels were affected to various degrees, with changes ranging from 2-fold to over 30-fold. The complete serum DI-FT-ICR-MS raw data set is shown in Tables S2 and S3.
In order to identify the metabolic pathways most significantly disturbed during leprosy, we selected metabolites detected in both BT and LL patients showing at least a 2-fold difference between them and queried the MassTRIX database (version 2, http://metabolomics.helmholtz-muenchen.de/masstrix2/) to determine putative metabolite identities. Figure 1B shows the categories of metabolites that differ between the two groups. Although many metabolic pathways were affected, our data suggest that the metabolism of omega-6 (linoleic and arachidonic acids) and omega-3 PUFAs (α-linolenic acid, EPA and DHA) are markedly modulated during M. leprae infection, with higher levels of a diverse class of bioactive lipid mediators in LL sera when compared to BT sera. The effects of M. leprae infection on specific classes of PUFAs are described in more detail below.
In order to gain further insights into the metabolic changes elicited during leprosy, sera samples from the same patients were collected before and immediately after the conclusion of MDT (six and twelve months for BT and LL patients, respectively). Total metabolites were extracted and analyzed by DI-FT-ICR-MS as described above. As an attempt to compare the metabolic profiles of the four groups of samples, we performed Principal Component Analysis (PCA) on this dataset using Multibase (http://www.numericaldynamics.com/). As can be seen from Figure 2, such analysis showed a clear separation between the BT and LL groups. Also, the PCA showed a clear separation of the LL samples before and after MDT, although the separation of BT samples before and after treatment was modest. This is in line with more extensive effects of lepromatous leprosy on host metabolism due to the high bacillary burden. Nevertheless, due to the extensive effects of lepromatous leprosy on polyunsaturated fatty acid metabolism, we focused our analysis on the effect of MDT on this metabolic pathway in both BT and LL patients. MDT caused a decrease in the levels of most potential metabolites from the arachidonic acid pathway, both in the BT and LL groups (Table 2 and Figure S4). This suggests that, although higher levels of these metabolites were generally observed in LL samples when compared to BT, these molecules were present at increased levels in leprosy patients in general, both LL and BT. In contrast, potential metabolites derived from linoleic and α-linolenic acids were mostly affected by MDT only in LL patients, returning to levels similar to those originally found in BT patients (Tables 3 and 4). The more extensive effect of MDT on the metabolic profiles of LL patients supports our initial findings that samples from LL patients show higher levels of eicosanoids and other polyunsaturated fatty acid metabolites than samples from BT patients, and this correlates well with the bacillary burden observed in these clinical forms of leprosy. As shown in Tables 2–4, with the exception of a few m/z, relative levels of most metabolites were indistinguishable when comparing BT and LL samples after MDT. In other words, MDT converted BT and LL patients to a common phenotype regarding the metabolic profiles of PUFAs.
Eicosanoids are lipid mediators that play a critical role as regulators of inflammation and the immune response generated during infection, including those caused by mycobacteria [21], [22], [23], [24], [25], [26]. Among the potential eicosanoids altered during leprosy, several of them possess the same molecular mass. In order to confirm the modulation of some of these compounds during M. leprae infection, levels of PGE2, PGD2, LXA4 (m/z [M-H]− 351.21782) as well as LTB4 (m/z [M-H]− 335.22282) were screened by EIAs. Circulating levels of these mediators were determined in leprosy patients (BT, n = 25; LL, n = 25) and compared with their levels in healthy controls (n = 10). While no differences in LTB4 levels were detected between different sample groups, the levels of PGD2 and PGE2 were significantly higher in LL patients when compared to BT (Figure 3), thus confirming the original observation that m/z 351.21782 was found in higher levels in LL serum by DI-FT-ICR-MS analysis (Table 2). Next, to reinforce the notion that the altered production of eicosanoids observed in leprosy patients results from an active modulation by the M. leprae infection, serum concentrations of PGE2, PGD2 and LTB4 in sera from BT and LL patients were measured at the conclusion of MDT and compared with the levels observed before treatment. By comparing pre- and post-MDT serum samples taken from the same patients, we observed significantly higher PGD2 levels in the BT group after the conclusion of MDT, in contrast to the heterogeneous behavior of this mediator observed in LL patients (Figure 4A). After treatment, the PGD2 levels were similar between LL and BT (Figure S5). Regarding levels of PGE2, a decrease was observed in most LL patients, although 4 of them showed higher levels after treatment (Figure 4B). In contrast, no changes in PGE2 levels were observed in most BT patients after the conclusion of MDT, although 3 patients showed a decrease in its levels (Figure 4B). Even after treatment, PGE2 levels were significantly higher in LL versus BT patients (Figure S5). LTB4 levels tended to decrease both in LL and BT patients, although 2 LL patients showed higher levels after conclusion of MDT (Figure 4C). Finally, as seen previously in the context of untreated patients, no differences between LTB4 serum levels in LL versus BT patients were detected after conclusion of MDT (Figure S5).
We also measured the levels of LXA4 in serum samples and found that concentrations of this lipid mediator were significantly altered in leprosy patients when compared to healthy controls. LXA4 is likely the major contributor to m/z [M-H]− 351.21782, followed by PGE2. While PGD2, PGE2 and LTB4 serum levels were below 0.6 ng/mL in most samples from leprosy patients, particularly in untreated LL patients, LXA4 levels were much higher, ranging from 2 to 17 ng/mL. As shown in Figure 3D, significantly higher levels of LXA4 were detected in both BT and LL patients when compared to the controls, but no significant difference was found between these two groups. However, after treatment, serum LXA4 levels in BT and LL patients returned to normal (Figure S5). The decrease in LXA4 levels in LL and BT sera after the conclusion of MDT can be clearly seen in paired pre- and post-MDT serum samples taken from the same patients (Figure 4D). LXA4 concentrations showed a statistically significant decrease after MDT, with a consistent behavior in all analyzed sera. These data suggest that LXA4 is a major contributor of m/z [M-H]− 351.21782 and point to a more predominant role of LXA4 during leprosy.
Resolvins, including D and E series resolvins, are endogenous lipid mediators generated during the resolution phase of acute inflammation from the omega-3 polyun- saturated fatty acids docosahexaenoic acid (DHA) and eicosapentaenoic acid (EPA), having potent anti-inflammatory and pro-resolution actions in several animal models of inflammation. In order to confirm that the omega-3 polyunsaturated fatty acid metabolism is disturbed during leprosy, levels of RvD1 in sera from leprosy patients were measured by EIA. Circulating levels of this mediator were determined in leprosy patients (BT, n = 20; LL, n = 19) and compared with their levels in healthy controls (n = 6). Interestingly, the results were similar to those observed for LXA4, which also has anti-inflammatory and pro-resolution action. Levels of RvD1 were found to be significantly different in leprosy patients when compared to healthy controls (Figure 5A), returning to normal levels after treatment in both BT and LL patients (Figure 5B). The decrease in RvD1 levels in LL and BT sera after the conclusion of MDT can also be seen in paired pre- and post-MDT serum samples taken from the same patients (Figure 5C and D). However, there was no difference between levels of RvD1 between BT and LL sera before MDT (Figure 5A), in contrast to the profile observed for the m/z 375.21792 in the DI-FT-ICR analysis (Table 4). Thus, the difference observed in the metabonomic study could be due to other compounds with the same m/z such as RvD2–4 or others.
To expand the metabolite profiles generated with serum samples, we performed a metabolomic analysis of human skin biopsies from 4 cases of LL and 4 cases of BT, looking for alterations in PUFA metabolism at the site of M. leprae infection. To do so, we extracted metabolites from the biopsies and analyzed them through DI-FT-ICR-MS, as described above. The complete skin DI-FT-ICR-MS raw data set is shown in Tables S4 and S5. Almost 2,000 metabolites were detected, and their relative abundance was compared between LL and BT lesions. Among the list of m/z detected, we selected metabolites showing at least a 2-fold difference between samples from BT and LL patients. As shown in Table 5, m/z potentially corresponding to docosapentaenoic acid (DPA), DHA, AA, linoleic acid/9-cis,11-trans-octadecadienoate, 1-acyl-sn-glicero-3-phosphocholine (lysolecithin), lecithin and plasmenic acid were present in higher levels in LL lesions. In contrast, the mass 376.2226 Da, which corresponds to several potential metabolites of the arachidonic acid pathway, was present in higher levels in BT lesions. Although definitive metabolite identity cannot be determined using this method, our results suggest that phospholipids and products of PLA2 activity accumulate in LL lesions, correlating with the higher levels of potential phospholipids and free unsaturated fatty acids and their derivatives observed in the serum of these patients. Of note, potential DPA (330.2559 Da) levels were about 50 times higher in LL lesions when compared to BT lesions (Table 5). DPA is a 22-carbon PUFA with anti-inflammatory properties derived from an elongation step of EPA, abundantly present in macrophages treated with EPA [27].
Although leprosy is one of mankind's oldest diseases, the interplay between the human body and M. leprae remains poorly understood. Research in leprosy lacks laboratory tools that can be used to predict susceptibility to the disease and disease progression, which are critical for an improved management of patients through the use of more rational therapeutic approaches. Among the branches of “omics”, the recent development of high-throughput techniques that allow the simultaneous identification and quantification of small metabolites from different tissues and biofluids is emerging as a powerful approach to investigate the modulation of host metabolism during infection, with the perspective to disclose potential contributors to disease pathology. Herein, we have applied a metabonomics analysis of serum samples from leprosy patients to the comparison of host metabolism regulation during infection in two distinct clinical forms of the disease (lepromatous versus tuberculoid). Extensive differences in metabolic composition during leprosy were observed, supporting the notion that a unique metabolic shift occurs during disease. Moreover, serum composition of infected patients converged to a similar profile after conclusion of treatment, indicating that the differences observed resulted from M. leprae infection. When analyzing the metabolic pathways affected by M. leprae, a robust increase in the levels of potential AA metabolites was observed in LL patients in comparison to BT patients. However, MDT caused a decrease in the levels of most potential metabolites from the arachidonic acid pathway, both in the BT and LL groups. This suggests that, although higher levels of these metabolites were generally observed in LL samples when compared to BT, these molecules were present at increased levels in leprosy patients in general, both LL and BT. One caveat of our metabonomics study is the fact that only a limited number of samples was available for analysis. Therefore, an extensive statistical analysis was not feasible and the results of the metabonomics experiments must be taken with caution. Nevertheless, this approach is very useful in an exploratory mode and many aspects that we have previously investigated using this methodology were confirmed using other techniques [3], [11]. In order to ameliorate this issue, we used EIAs to measure the concentrations of a few molecules of interest in the serum of other leprosy patients and healthy controls. Higher levels of PGD2 and PGE2 in LL sera when compared to BT sera were confirmed through EIAs. We also found high levels of LXA4 in both LL and BT patients in comparison with healthy individuals. Of note was the enrichment of masses that may correspond to omega-3 PUFAs and their biologically-active, anti-inflammatory and pro-resolving Rvs, PD1 and MaR derivatives, some of which were also detected in LL skin lesions. Higher levels of RvD1 were detected by EIA in leprosy patients (both LL and BT), and decreased to normal levels after treatment. To our knowledge, this is the first study reporting the levels of LX and RvD1 during leprosy. The main conclusion of this study is that PUFA metabolism is markedly regulated during M. leprae infection, potentially contributing to multiple aspects of the immunopathogenesis of leprosy.
The finding of higher levels of potential free PUFAs both in sera and skin lesions of LL patients, and of lysophosphatidylcholine in LL lesions suggests a high lipid turnover in these lesions. These data agree with previous studies showing a higher expression of host PLA2 and PLC in LL patients [28] and of the high PL activity detected in M. leprae preparations [29]. PGE2 levels were significantly higher in untreated LL patients, returning to levels similar to BT patients after the conclusion of MDT. Accordingly, increased cyclooxygenase-2 expression has been observed in biopsies from LL patients [26], [30]. PGE2 is the main cyclooxygenase product produced by macrophages, and it supports acute local inflammation, being at a first moment pro-inflammatory and at the same time immunosuppressive, because it inhibits cell-mediated immunity by selectively inhibiting Th1 cytokines (IFN-γ and IL-2) and suppressing IL-12 production in monocytes and dendritic cells (as well as the expression of its receptor), without interfering with the production of the Th2 cytokines IL-4 and IL-5. Overproduction of PGE2 is observed in Th2-associated diseases (asthma, atopic dermatitis) (reviewed in [31]), which is the case of LL leprosy, where humoral immune responses are unable to control the infection. The observed increase in PGE2 levels in sera from LL patients agrees with previous studies of PGE2 in M. leprae, where it was observed in animal models (nude mice) that infected macrophages obtained from footpad granulomas produced high levels of PGE2, which was associated with a down-regulation of macrophage and T-cell functions [21]. These functions were restored when PGE2 biosynthesis was inhibited, either in vivo, when infected mice were subjected to a diet deficient in essential fatty acids, or in vitro, by treatment of cultured cells with indomethacin [21], [22]. Human monocytes obtained from LL patients showed a high production of PGE2 [24], [32], and other studies showed that the lipid droplets induced in macrophages and Schwann cells by M. leprae are sites for PGE2 biosynthesis. Moreover, COX-2 was detected in lipid droplets present in nerve and dermal lesions of LL patients, suggesting that they constitute sites of PGE2 production in vivo [25], [26].
Recent studies indicate that PGE2 may have different effects during the course of inflammation. At early stages, as previously described, PGE2 presents a pro-inflammatory activity (reviewed in [31]). However, with the progress of the inflammatory process, it was observed that PGE2 decreases the production of 4-series LTs through the inhibition of 5-lipoxygenases, and regulates the transcription of 15-lipoxygenase in neutrophils, switching the production of LTs to LXs (reviewed in [20]). Indeed, it has been recently shown that PGE2 serves as a feedback inhibitor essential for limiting chronic inflammation in autoimmune arthritis [33]. Furthermore, PGE2 inhibits the synthesis of the pro-inflammatory cytokines TNF-α and IL-1 by macrophages (reviewed in [31]). PGE2 may undergo a non-enzymatic dehydration reaction, forming the cyclopentenone PGA2 and its isomerization products PGC2 and PGB2. Cyclopentenone PGs have reported anti-inflammatory activity, through activation of PPAR, specifically PPAR-α and PPAR-δ in the case of PGA2 (reviewed in [34]). Interestingly, m/z 333.20712, which may correspond to PGA2, PGB2 and PGC2, was detected in LL but not BT sera, probably as a consequence of the higher availability of PGE2 in LL. Therefore, PGE2, in conjunction with its cyclopentenone PG derivatives, may play an immunosuppressive and anti-inflammatory role in LL.
Regarding PGD2, it is also a pro-inflammatory eicosanoid, and it elicits inflammatory and vascular responses through interaction with the D prostanoid receptor 1 (DP) and chemoattractant receptor-like molecule expressed on Th2 cells (CRTH2). PGD2 is capable of inducing chemotaxis of eosinophils, basophils, and Th2 cells, stimulating the production of IL-4, IL-5, and IL-13 in the latter [35], and thus eliciting a Th2 response, typical of LL immunopathology. Other studies of PGD2 synthase (PGDS) expression showed a drop in its biosynthesis after the beginning of the inflammatory process, reaching its lowest point at the peak of inflammation, and returning to normal levels as the inflammation resolved, indicating a role of PGD2 in the promotion of the resolution process. Similarly to PGE2, PGD2 can undergo spontaneous dehydrations, leading to the formation of 15-deoxy-Δ12,14-PGJ2 (15d-PGJ2), which can also act via DP. However this PG acts mainly via intracellular receptors, activating PPAR-γ and inhibiting nuclear factor kappa B (NF-κB). 15d-PGJ2 has anti-inflammatory and pro-resolution effects, inhibiting the secretion of IL-6, IL-1β, IL-12 and TNF-α from macrophages, and downregulating the production of inducible nitric oxide synthase (iNOS) [36]. 15d-PGJ2 is a very unstable molecule; its intermediate Δ12-PGJ2 is formed by the dehydration of PGD2 catalyzed by human serum albumin, which may bind and stabilize Δ12-PGJ2, as well as 15d-PGJ2 [36]. In our metabonomics analysis, no significant hits for 15d-PGJ2 were observed in BT and LL sera. However, the levels of compounds with an m/z potentially corresponding to Δ12-PGJ2 (m/z [M-H]− 333.20712) were significantly higher in LL patients, and were reduced after MDT.
As mentioned above, the higher levels of LXA4, the predominant endogenously-generated LX, in leprosy patients suggested by the metabonomics analysis were confirmed by EIA. LXs are trihydroxytetraene-containing AA metabolites that are produced by at least 3 distinct LO pathways, involving interactions among diverse cell types, including leukocytes, epithelia, endothelia, and platelets. LXA4 and/or its aspirin-triggered isomer, 15-epi-LXA4 have a number of reported in vitro activities, including: (a) inhibition of neutrophil chemotaxis, adherence, transmigration, and activation; (b) suppression of the production of diverse chemokines by epithelial cells and leukocytes; (c) inhibition of IL-12 production by dendritic cells; (d) upregulation of monocyte chemotaxis and ingestion of apoptotic neutrophils; and (e) suppression of MMP production, while stimulating production of tissue inhibitors of MMPs. In vivo, LXA4 has been shown to have broad counter-regulatory properties, suppressing proinflammatory responses (preventing neutrophil-mediated damage, promoting the resolution of neutrophil-mediated inflammation), Th2-polarized responses (inhibiting inflammation and airway hyperresponsiveness in experimental asthma), and Th1 responses (suppressing immunopathology during infection with Toxoplasma gondii) alike [20], [37]. Moreover, LXA4 stimulates phagocytosis and IL-10 production in macrophages [38], a phenotype characteristic of foamy macrophages present in LL lesions [25].
Our metabonomics data on omega-3 PUFAs are sustained by a recent serum metabonomic analysis on leprosy patients, which showed a significant raise in the levels of EPA and DHA in sera from high-BI patients [39]. Also, hits that may correspond to DHA and DPA (a 22-carbon derivative of EPA) were detected in higher levels in skin lesions of LL patients when compared to BT lesions, reinforcing these data. Moreover, the remarkable differences in the levels of several potential omega-3 PUFA metabolites observed in leprosy patients before and after MDT, point to the participation of these bioactive lipid mediators in the immunopathology of leprosy. The anti-inflammatory properties of omega-3 PUFAs have been recently shown to be mediated, at least in part, by a new family of pro-resolving lipid mediators that include Rvs, PD1 and MaR (reviewed in [20]). Our metabonomics data showed the decrease of m/z that may correspond to RvE1 and RvE2, RvD1–4, RvD5–6, PD1, as well as MaR1 after treatment. Indeed, high levels of RvD1 were found by EIA in serum samples of leprosy patients, which returned to normal levels after treatment. Moreover, an m/z that corresponds to DPA was found in levels 50 times higher in skin biopsies of LL when compared to BT lesions.
Lipid mediators are produced in a temporally orchestrated fashion during inflammation. During the initial phases of inflammation, pro-inflammatory eicosanoids such as PGE2, PGD2 and LTB4 are generated. With time, a class-shift occurs towards anti-inflammatory and pro-resolving mediators (LXA4, 15d-PGJ2, Rvs, PD1 and MaR) that switch the inflammatory response off and restore homeostasis. Resolution of inflammation and return to homeostasis is actively mediated by these compounds and the failure of resolution is considered as one of the causes of chronic inflammatory diseases such as age-related macular degeneration, asthma, lupus erythematosus, atherosclerosis, chronic pulmonary disease, inflammatory bowel disease, multiple sclerosis, rheumatic arthritis and cancer [40]. In all of these cases, LX deficiency in association with high levels of pro-inflammatory mediators has been implicated in disease pathogenesis. Thus, LXA4 and its more stable synthetic analogues, as well as Rvs, PD1 and MaR and their agonists have emerged as novel therapeutic candidates via accelerated resolution of inflammation for the management of a broad range of disorders with an inflammatory component, including type 2 diabetes and cardiovascular diseases [41], [42]. On the other hand, production of LXA4 early during inflammation was shown to delay resolution and, in the case of infection, promote pathogen persistence in the host. This is the case for infections with M. tuberculosis and M. marinum, where an imbalance between LXA4 and pro-inflammatory eicosanoids (PGE2 and LTB4) during the early stages of infection has been shown to favor pathogen survival and multiplication [43].
Interestingly, a recent study on metabolic profiling of sera from tuberculosis (TB) patients also provided evidence for anti-inflammatory metabolic changes in this disease [44]. The authors found increased levels of kynurenine, the product of tryptophan catabolism by indoleamine 2,3 dioxygenase 1 (IDO1), in patients with active TB. This was significantly correlated with similarly increased abundance of the immunosuppressive stress hormone cortisol.
The metabonomics analysis presented herein discloses potential host tolerance mechanisms to M. leprae infection. Recently, the concept of disease tolerance as a defense strategy to infection has been introduced in the field of animal immunity (reviewed in [45]). While the immune system protects from infections primarily by detecting and eliminating the pathogen, tolerance does not directly affect pathogen burden, but rather, decreases immunopathology caused by the pathogens or the immune responses against them. Particularly the lepromatous pole of leprosy seems to be an excellent model to study disease tolerance in humans. Clinical data indicate that LL patients have developed tolerance mechanisms that allow them to survive with minimal pathology, despite the high bacterial burden. In LL patients, failure of the immune system to kill or inhibit M. leprae allows the mycobacteria to reproduce to very high numbers reaching multiple tissues and organs in a systemic infection. Heavy bacteremia is often observed in these patients but, in contrast to other bacterial infections, no symptoms of septicemia are observed. Moreover, a subtype of LL, known as diffuse LL, “pretty leprosy” or Lucio leprosy, appears in the earlier stages of disease as uniformly diffused, shiny infiltrations of all the skin of the body, without any actual lesions [46]. Increased tolerance to tissue damage can be achieved, in general, through tissue protection and repair. It is, therefore, reasonable to speculate that the higher levels of LXA4, and PGE2 levels, in association with the omega-3 PUFAs DHA, EPA, RvD1, and other potential Rvs, PD1 and MaR detected in leprosy patients may contribute to the molecular mechanisms that restrain the inflammatory responses in LL and at the same time favor M. leprae growth and persistence in the host. Indeed, the ameliorative effects of LXA4 and omega-3 PUFA metabolites have been reported in animal models of sepsis and through the observation of their inhibitory effects on the inflammatory response to endotoxin in humans (reviewed in [42], [47], [48]). Although the role of these resolving lipid mediators is well established in acute infections, more detailed studies on chronic infections are needed to establish the function of these mediators in determining disease outcome. Deciphering the molecular details of tolerance mechanisms in leprosy may pave the way to new prevention and management strategies of leprosy reactions as well as new treatments for many human maladies, including infectious, inflammatory and autoimmune diseases.
|
10.1371/journal.pgen.1004826 | Genome-Wide Analysis of leafbladeless1-Regulated and Phased Small RNAs Underscores the Importance of the TAS3 ta-siRNA Pathway to Maize Development | Maize leafbladeless1 (lbl1) encodes a key component in the trans-acting short-interfering RNA (ta-siRNA) biogenesis pathway. Correlated with a great diversity in ta-siRNAs and the targets they regulate, the phenotypes conditioned by mutants perturbing this small RNA pathway vary extensively across species. Mutations in lbl1 result in severe developmental defects, giving rise to plants with radial, abaxialized leaves. To investigate the basis for this phenotype, we compared the small RNA content between wild-type and lbl1 seedling apices. We show that LBL1 affects the accumulation of small RNAs in all major classes, and reveal unexpected crosstalk between ta-siRNA biogenesis and other small RNA pathways regulating transposons. Interestingly, in contrast to data from other plant species, we found no evidence for the existence of phased siRNAs generated via the one-hit model. Our analysis identified nine TAS loci, all belonging to the conserved TAS3 family. Information from RNA deep sequencing and PARE analyses identified the tasiR-ARFs as the major functional ta-siRNAs in the maize vegetative apex where they regulate expression of AUXIN RESPONSE FACTOR3 (ARF3) homologs. Plants expressing a tasiR-ARF insensitive arf3a transgene recapitulate the phenotype of lbl1, providing direct evidence that deregulation of ARF3 transcription factors underlies the developmental defects of maize ta-siRNA biogenesis mutants. The phenotypes of Arabidopsis and Medicago ta-siRNA mutants, while strikingly different, likewise result from misexpression of the tasiR-ARF target ARF3. Our data indicate that diversity in TAS pathways and their targets cannot fully account for the phenotypic differences conditioned by ta-siRNA biogenesis mutants across plant species. Instead, we propose that divergence in the gene networks downstream of the ARF3 transcription factors or the spatiotemporal pattern during leaf development in which these proteins act constitute key factors underlying the distinct contributions of the ta-siRNA pathway to development in maize, Arabidopsis, and possibly other plant species as well.
| Mutations in maize leafbladeless1 (lbl1) that disrupt ta-siRNA biogenesis give rise to plants with thread-like leaves that have lost top/bottom polarity. We used genomic approaches to identify lbl1-dependent small RNAs and their targets to determine the basis for these polarity defects. This revealed substantial diversity in small RNA pathways across plant species and identified unexpected roles for LBL1 in the regulation of repetitive elements within the maize genome. We further show that only ta-siRNA loci belonging to the TAS3 family function in the maize vegetative apex. The TAS3-derived tasiR-ARFs are the main ta-siRNA active in the apex, and misregulation of their ARF3 targets emerges as the basis for the lbl1 leaf polarity defects. Supporting this, we show that plants expressing arf3a transcripts insensitive to tasiR-ARF-directed cleavage recapitulate the phenotypes observed in lbl1. The TAS3 ta-siRNA pathway, including the regulation of ARF3 genes, is conserved throughout land plant evolution, yet the phenotypes of plants defective for ta-siRNA biogenesis are strikingly different. Our data leads us to propose that divergence in the processes regulated by the ARF3 transcription factors or the spatiotemporal pattern during development in which these proteins act, underlies the diverse developmental contributions of this small RNA pathway across plants.
| Small RNAs are important regulators of development, particularly in plants, where many of the abundant and conserved microRNAs (miRNAs) target transcription factors that direct or reinforce cell fate decisions [1]. Consequently, mutations in genes required for miRNA processing or function condition defined developmental defects. Likewise, plants defective for the biogenesis of trans-acting short interfering RNAs (ta-siRNAs) show distinctive patterning defects due to the deregulation of key developmental targets [1]. ta-siRNAs are generated in response to miRNA activity via one of two possible mechanisms, referred to as the “one-hit” and “two-hit” pathways. In both pathways, a single miRNA-guided cleavage event triggers the conversion of target transcripts into long double stranded RNAs by RNA-DEPENDENT RNA POLYMERASE6 (RDR6) and SUPPRESSOR OF GENE SILENCING3 (SGS3), and sets the register for the subsequent production of phased 21-nt siRNAs by DICER-LIKE4 (DCL4) [2]-[4]. In the one-hit pathway, transcripts targeted by a single, typically 22-nt, miRNA will generate ta-siRNAs downstream of the miRNA cleavage site, whereas transcripts producing ta-siRNAs via the two-hit pathway harbor two binding sites for 21-nt miRNAs and the ta-siRNAs are processed upstream of the cleaved 3′ miRNA target site [5]–7. Analogous to miRNAs, a subset of the phased ta-siRNAs act at the post-transcriptional level to repress the expression of genes involved in development or other cellular processes.
The phenotypes conditioned by mutations affecting ta-siRNA biogenesis vary greatly across species. In Arabidopsis, such mutants exhibit a relatively subtle phenotype, developing downward curled leaves that are weakly abaxialized and undergo an accelerated transition from the juvenile to the adult phase [2]–[3]. These defects result from misregulation of the AUXIN RESPONSE FACTOR ARF3, which is targeted by TAS3-derived ta-siRNAs, termed tasiR-ARFs [8]–[10]. Biogenesis of the TAS3 ta-siRNAs follows the two-hit model and involves a subspecialized pathway, which requires the unique association of miR390 with its effector AGO7 to trigger siRNA production [11]. Localized expression of TAS3 and AGO7 confines tasiR-ARF biogenesis to the adaxial/upper most cell layers of developing leaves, which then limits accumulation of ARF3 to the abaxial/lower side [12]. The developmental defects of sgs3, rdr6, and dcl4 are phenocopied by mutations in AGO7 and TAS3A, as well as by expression of tasiR-ARF- insensitive ARF3 transgenes [2], [8]–[9], [13], indicating that the contribution of ta-siRNAs to Arabidopsis development is primarily mediated by tasiR-ARFs.
In contrast to Arabidopsis, loss of AGO7 activity in Medicago results in the formation of highly lobed leaves [14], and mutants defective for ta-siRNA biogenesis components in rice and tomato exhibit severe defects in meristem maintenance, mediolateral blade expansion, and adaxial-abaxial leaf polarity [15],[16]. Likewise, mutations in maize lbl1 and ragged seedling2 (rgd2), which encode the orthologs of SGS3 and AGO7, respectively, have severe effects on meristem function and leaf development [17]–[18]. lbl1 mutants, in particular, develop radial, fully abaxialized leaves (Fig. 1A).
Importantly, while the TAS3 ta-siRNA pathway is evolutionarily conserved, the number and nature of phased siRNA loci vary greatly between plant species [19]. In Arabidopsis, three TAS families have been described in addition to TAS3. TAS1-, TAS2-, and TAS4-derived ta-siRNAs are generated via the one-hit model following miR173- or miR828-directed cleavage, and function in the regulation of members in the MYB transcription factor and PPR families [4], [20]–[21]. Each of these pathways has been identified in other plant species, but their evolutionary origin appears to lie within the eudicots [19]. Depending on the species, variation is seen in the genes targeted by the ta-siRNAs derived from these TAS loci, as well as in the miRNA that triggers their biogenesis [22], [23]. In addition, apparent species-specific TAS pathways may exist, as novel TAS loci with unique targets have been identified in tomato and the moss Physcomitrella patents [24], [25]. Moreover, genome-wide small RNA analyses in a number of plant species have uncovered clusters of phased siRNAs, generated primarily via the one-hit model that, unlike the ta-siRNAs, are proposed to act in cis. These are generally referred to as phasiRNAs and are processed from non-coding transcripts, such as in the panicles of the grasses [26]–[28], or from protein-coding genes, including members of the NB-LRR, MYB, and PPR gene families [19]. The set of genes regulated by phased siRNAs thus varies widely across plant species.
The basis for the phenotypes of maize ta-siRNA biogenesis mutants remains unclear. In light of the tremendous diversity in phased siRNAs seen across plant species, it is conceivable that TAS loci, other than the four known TAS3 genes [17], exist in maize. Moreover, phased siRNAs other than the tasiR-ARFs may target genes with roles in development, and contribute to the defects seen in lbl1 mutants. To assess these possibilities and to obtain a comprehensive view of LBL1-dependent siRNAs active in the maize vegetative apex, where the mutant phenotype manifests itself, we compared the small RNA content between wild-type and lbl1 shoot apices. This revealed unexpected contributions of LBL1 to the regulation of transposons, particularly the gyma class of LTR-retrotransposons. Interestingly, in contrast to other plant species, we found no evidence for the existence of phased siRNAs generated via the one-hit model. Our analyses identified nine TAS loci all belonging to the TAS3 family. Data from RNA deep sequencing and PARE analysis present the ARF3 genes as the only LBL1-dependent small RNA targets with a role in development. Consistent with this finding, plants expressing a tasiR-ARF insensitive arf3a transgene recapitulate the phenotype of lbl1 mutants. These findings underscore the importance of the tasiR-ARF - ARF3 regulatory module to maize development, and indicate that diversity in TAS pathways and their targets cannot fully account for the phenotypic differences conditioned by ta-siRNA biogenesis mutants across plant species. Instead, divergence in the gene networks downstream of the ARF3 transcription factors or the spatiotemporal pattern in which these tasiR-ARF targets act emerge as a testable hypotheses to explain the diverse contributions of the ta-siRNA pathway to development in maize, Arabidopsis, and possibly other plant species as well.
To understand the basis for the lbl1 phenotype, we compared the small RNA content between vegetative apices, comprising the shoot apical meristem (SAM) and up to five leaf primordia, of two-week old B73 and lbl1-rgd1 seedlings. Three independent biological replicates were analyzed for each genotype. Approximately 92% of the 18- to 26-nt reads in both sets of libraries mapped to the unmasked B73 reference genome (S1 Table). Of the mapped reads, 43–48%, corresponded to unique small RNAs suggesting that, despite the highly repetitive nature of the maize genome, close to half of the small RNAs expressed in the vegetative apex are distinct. The small RNA size distribution profiles in both wild-type and lbl1 resemble those described previously for maize [29]–[30], suggesting that LBL1, in contrast to components of the heterochromatic siRNA pathway [29], has a relatively subtle effect on the overall small RNA population (S1 Figure). However, consistent with a role for SGS3 proteins in the biogenesis of 21-nt secondary siRNAs [2], [4], the 21-nt small RNA population is slightly reduced in lbl1 compared to wild-type. In addition, an unexpected modest reduction in the 22- and 24-nt small RNA fractions is seen in lbl1 (S1 Figure).
To more precisely define the effects of LBL1 on small RNA biogenesis, we identified genomic loci that differentially accumulate 21-, 22-, or 24-nt small RNAs in wild-type versus lbl1 apices (S2 Figure). Consistent with a relatively subtle effect of lbl1 on the overall small RNA population, this identified 79, 172, and 209 loci that generate 21-, 22-, or 24-nt small RNAs, respectively, that are significantly changed (q-value<0.05) at least 2-fold between wild-type and lbl1 (Fig. 1B; S1 Dataset). A small subset of these (11/79), correspond to low copy genic regions that generate significantly fewer 21-nt small RNAs in lbl1, properties predicted for phasiRNA and ta-siRNA loci (Fig. 1C; S1 Dataset). Indeed, the four previously described TAS3 genes, tas3a-d [17], are among these loci. Five additional genes appear non-coding, whereas the remaining two correspond to arf3a and arf3d (Fig. 1C; S1 Dataset). With the exception of arf3a, each of these loci also generate 22-nt LBL1-dependent siRNAs and five differentially accumulate 24-nt small RNAs, albeit generally to substantially lower levels than seen for the corresponding 21-nt small RNAs (Fig. 1C). One additional gene (GRMZM2G093276), encoding a zinc/iron transporter, also accumulates 22-nt small RNAs that are significantly reduced in lbl1. Finally, six predicted protein-coding genes of unknown function generate low levels of 24-nt small RNAs that are lost or significantly reduced upon mutation of lbl1. Thus, a total of 18 distinct low copy genic regions generate small RNAs in an apparent LBL1-dependent manner. These include the four known maize TAS loci, tas3a-d, and the remaining present candidate novel phasiRNA or TAS loci that are active in the vegetative apex and may contribute to the developmental defects resulting from mutation of lbl1.
To further discern whether additional phased siRNA loci are active in the vegetative maize apex, we developed a pipeline that scans the genome for clusters of siRNAs showing a regular phasing of 21, 22, or 24 nucleotides (S2 Figure). This pipeline includes a phasing score calculation (P-score) that identifies clusters in which the majority of small RNAs produced are phased. With a P-score threshold (P≥25) that has been shown to identify 7 of the 8 TAS loci in Arabidopsis [21], [31], we identified 16 phased 21-nt siRNA clusters, 102 phased 22-nt siRNA clusters, and 8 phased 24-nt siRNA clusters (S2 Dataset). However, combining this analysis with the differential small RNA accumulation data described above (S1 Dataset), showed that small RNA levels in most of the clusters are not changed significantly between wild-type and lbl1. In fact, small RNA accumulation at just 8 of the phased siRNA clusters is changed in the lbl1 mutant (S3A Figure). A closer inspection of the remaining clusters indicates that these correspond primarily to repetitive regions in the genome. Moreover, these siRNA clusters are typically embedded within large windows, frequently spanning over 5 kb, that generate an uncharacteristically high number of small RNA reads, which inflates the P-score (see Materials and Methods). As such, it is unlikely that these clusters correspond to new TAS loci or other miRNA-triggered phased siRNA loci. Instead these loci, particularly the LBL1-independent 22-nt siRNAs, appear to be processed from long hairpin RNAs or overlapping antisense transcripts (S3B-C Figure). In Arabidopsis, natural antisense transcripts are processed by DCL1 into 21-nt small RNAs [32], [33], whereas long hairpin RNAs are targeted by multiple DCL enzymes to give rise to variably sized small RNAs [34]. The preferential processing of such double stranded RNAs into 22-nt siRNAs presents a possible basis for the uncharacteristically high overall abundance of 22-nt small RNAs in maize (S1 Figure), and suggests diversification in the action of DCL family members between Arabidopsis and maize.
Of the eight phased siRNA clusters whose small RNA levels are significantly changed in lbl1, one generates 24-nt phased siRNAs (S3A Figure; S2 Dataset). However, its small RNA levels are increased in lbl1, again making it unlikely that this cluster corresponds to a novel phased secondary siRNA locus. Thus, in the vegetative apex, only seven loci generate phased siRNAs with a P-score≥25 in an LBL1-dependent manner. These phased siRNAs are 21-nt in size and are derived from low copy genic regions predicted to generate non-coding transcripts. Importantly, the four previously described maize TAS3 loci, tas3a-d [17], are among these seven loci (S2 Dataset). The three additional loci, GRMZM5G806469, GRMZM2G082055, and GRMZM2G512113, present novel maize phased siRNA loci that could contribute to the developmental defects resulting from mutation of lbl1.
A closer analysis of the three novel phased siRNA loci indicates that these represent new members of the TAS3 family. Transcripts from these loci contain two miR390 binding sites and have the potential to generate small RNAs homologous to tasiR-ARFs (Fig. 2A-D; S4A-G Figure). As mentioned above, two additional low copy regions in the genome (GRMZM2G155490 and GRMZM2G588623) generate 21-nt LBL1-dependent small RNAs from predicted non-coding transcripts (Fig. 1C; S1 Dataset). While both loci did not pass the P-score filter in the phased siRNA analysis (S2 Dataset), a closer analysis indicates that both contain two miR390 binding sites (S4D, F Figure). Both loci generate relatively few small RNAs in vegetative apices with many being out of phase (S4 D, F Figure), presenting a likely explanation for the observed low P-score. Supporting this, a similar P-score analysis of 21-nt small RNAs in Arabidopsis failed to detect the confirmed TAS4 ta-siRNA locus due to the low abundance of its reads [21]. Data from PARE (parallel analysis of RNA ends; [35]) libraries generated from B73 apices, which allow the detection of small RNA-directed mRNA cleavage products, confirms that the 3′ miR390 target site in transcripts generated from all nine loci is cleaved (Fig. 2A-B). As such, the new loci were named tas3e to tas3i (Table 1).
The number of potential ta-siRNAs generated from these loci varies from 9 to 13 per strand (Table 1). However, not all predicted ta-siRNAs are detected in the vegetative apex; in fact, only 99 of the 194 possible ta-siRNAs are present in our datasets (Fig. 2A-D; S4A-G Figure). Moreover, our analyses reveal that while not all ta-siRNAs are 21-nt in size, only the 21-nt long small RNAs are phased, and this class is more abundant than the longer, or out of phase, small RNAs (Fig. 2C; S4A-G Figure). This indicates that DCL4 processing is occasionally out of phase and/or a different DCL enzyme takes over. Intriguingly, besides the generally low ta-siRNA levels coming from tas3f and tas3h, the potential tas3f-derived tasiR-ARF was not detected in our small RNA libraries, and tas3h lacks the potential to generate this small RNA altogether. Although it is conceivable that tas3f and tas3h have other biological roles, it seems that, with respect to the vegetative apex, these loci are diverging and losing their function in the tasiR-ARF pathway.
A similar analysis of the remaining seven low copy regions accumulating significantly fewer 22- and 24-nt small RNAs in lbl1 (Fig. 1C; S1 Dataset), make it unlikely that these represent new phased siRNA loci. Target prediction and PARE analysis indicate that none of these genes are targeted by the known maize miRNAs or ta-siRNAs (see below). Interestingly, similar analyses of phased siRNA clusters in rice and Brachypodium inflorescences did detect 24-nt phased siRNA loci in addition to 21-nt phased siRNAs [26]–[28]. miR2275, which triggers the biogenesis of the rice and Brachypodium 24-nt phased siRNAs, is not detected in the maize vegetative apex. This miRNA is, however, present in maize inflorescence tissues [26], implying that some of the observed diversity may reflect tissue-specificity of small RNA pathways. Likewise, many of the 21-nt phased siRNAs identified in rice and Brachypodium inflorescences form a family distinct from the TAS3 loci whose biogenesis is triggered by miR2118, which in maize accumulates specifically in the inflorescences [26].
In addition, no loci generating phased siRNAs via the “one-hit” model were identified in our analysis, despite the presence of 22-nt small RNAs shown to serve as triggers in this process [6], [7]. The maize apex appears unique in this regard, as all similar genome-wide analyses of small RNAs in seedling tissues in other species identified both types of phased siRNA loci [21], [31], [36], [37]. Perhaps such miRNAs are not loaded into maize AGO proteins or, alternatively, loading of these miRNAs fails to trigger a reprogramming of the RNA-induced silencing complex required to trigger secondary siRNA biogenesis [7]. Taken together, the above data indicate that, in the maize vegetative apex, phased secondary siRNAs are generated from nine loci, all belonging to the TAS3 family. This rules out a possible contribution of novel TAS families to the distinctive phenotype of lbl1 mutants.
Considering that the lbl1 phenotype is not explained by the presence of novel TAS loci, we next asked whether TAS3-derived ta-siRNAs other than the tasiR-ARFs contribute to the developmental defects seen in lbl1. To test this possibility, we constructed degradome libraries from B73 apices, and used these in target prediction [4] and PARE analyses to identify potential targets for all 99 TAS3-derived ta-siRNAs detected in our small RNA libraries. Allowing for a maximum score of 4.5, well above the maximum score of 3.0 obtained for the tasiR-ARF ARF3 duplexes, and filters comparable to Zhai et al. [31], PARE analysis confirmed 18 cleavage sites in 11 target genes (S3 Dataset). Five of the TAS3 transcripts are among the verified tasiRNA targets, indicating possible feedback regulation in the tasiRNA pathway. The remaining targets include GRMZM2G018189, which encodes an uncharacterized protein homologous to AtSLT1 that mediates salt tolerance in yeast [38], as well as the five members of the maize ARF3 gene family (Fig. 3A-C; Table 2). The ARF3 genes are targeted by multiple closely related tasiR-ARFs, and account for 8 of the validated cleavage sites (Figs. 2D; 3A; S3 Dataset).
To assess a possible contribution of the ta-siRNA targets to the lbl1 seedling defects, we next determined whether their expression levels are changed in the mutant. The same tissue samples from wild-type and lbl1 vegetative apices were used to identify differentially expressed genes by RNA deep sequencing (S2 Table; S3 Dataset). 1116 genes show differential expression in lbl1 compared to wild-type (fold change ≥2, q-value<0.05). Of the verified ta-siRNA targets, transcript levels for arf3a and arf3c-e are significantly increased in lbl1 (Table 2). Misregulation of these abaxial determinants is consistent with the lbl1 leaf polarity defects [17], but, unexpectedly, expression for arf3b is unchanged in the mutant. qRT-PCR analysis of RNA isolated from lbl1-rgd1 apices similarly shows that transcript levels for arf3a and arf3c-e are significantly increased, by approximately 2-fold, whereas expression of arf3b remains unchanged (Fig. 3B). The latter is correlated with a relatively low number of PARE signatures at arf3b (S3 Dataset), and suggests that this arf3 member may not substantially add to the adaxial-abaxial polarity defects seen in lbl1. Likewise, transcript levels for GRMZM2G018189 are not significantly changed in the mutant, and the number of PARE signatures precisely at the predicted cleavage site is low. Along with its predicted role in salt detoxification, this makes a contribution of GRMZM2G018189 to the lbl1 phenotype unlikely.
Interestingly, arf3a and arf3d generate 21-nt small RNAs, which are significantly downregulated in lbl1 (Fig. 1C; S1 Dataset). Both ARF3 genes contain two tasiR-ARF target sites and the majority of the LBL1-dependent siRNAs map between these sites (Fig. 3A; S3 Dataset). This suggests that tasiR-ARF-mediated regulation of arf3a and arf3d triggers the biogenesis of third tier secondary siRNAs also via the two-hit model, and implies positive feedback in the regulation of ARF3 expression. Although the significance of such feedback regulation remains to be established, considering the role of tasiR-ARFs in limiting the activity of ARF3 abaxial determinants to the lower side of leaf primordia, positive feedback could be important to reinforce adaxial cell fate, and to sharpen or maintain the boundary between adaxial and abaxial domains [10], [17].
In contrast to dicot species [21], [31], [36], [37], the miR390-dependent TAS3 ta-siRNA pathway, thus, is the only phased secondary siRNA pathway active in the maize shoot apex. This ta-siRNA pathway, including the regulation of ARF targets, is conserved throughout land plant evolution [4], [5], [10]–[17], although substantial diversity has accumulated over evolutionary time. The tasiR-ARFs show sequence divergence, and not all maize TAS3 genes generate this biologically active ta-siRNA (Fig. 2D; S4F Figure). In addition, like the Physcomitrella patens AP2 targets [24], GRMZM2G018189 presents a possible novel maize TAS3 ta-siRNA target. Despite such divergence in the TAS3 pathway, the ARF3 genes form the prime ta-siRNA targets also in maize, suggesting that the developmental phenotype of lbl1 results, at least in part, from a failure to correctly regulate their expression.
A requirement for LBL1 in the production of phased siRNAs explains, however, only a subset of the small RNA level changes identified in the vegetative apex of lbl1 mutants. A further 68, 161, and 198 loci generating 21-, 22-, or 24-nt small RNAs, respectively, show a significant change (q-value<0.05) in small RNA accumulation of at least 2-fold between wild-type and lbl1 apices (S1 Dataset, Fig. 4A-B), and these could conceivably contribute to the developmental defects of lbl1 mutants. To assess this possibility and to gain insight into the possible function of these LBL1-dependent siRNAs, we first determined whether genes differentially accumulating small RNAs in lbl1 show a corresponding change in transcript levels. In addition to the ARF3 and TAS3 loci discussed above, 23 protein coding genes within the maize filtered and working gene sets show a significant difference in 21-, 22-, or 24-nt small RNA accumulation between wild-type and lbl1 apices (Fig. 4A-B; S1 Dataset). However, only GRMZM2G093276, which encodes a ZIP zinc/iron transport protein, shows a significant increase in transcript levels in lbl1 that is correlated with a decrease in siRNAs at the locus (S4 Dataset). Plants overexpressing ZIP proteins show a reduction in plant height and axillary bud outgrowth [39], [40], but such plants are otherwise morphologically normal. As such, a contribution of ZIP deregulation to the severe polarity defects of lbl1 mutants seems speculative.
In addition, levels of eight miRNAs are significantly changed in lbl1 (Fig. 4B; S1 Dataset). However, even though LBL1 is necessary for the proper spatiotemporal pattern of miR166 accumulation [17], its levels overall appear not significantly changed in the mutant. Unexpectedly, miR156, which represses the juvenile to adult phase transition [41], is upregulated in lbl1. Arabidopsis ta-siRNA biogenesis mutants exhibit an accelerated transition from the juvenile to the adult phase [2], [9], whereas increased levels of miR156 in lbl1 might imply a delayed vegetative phase change. It is conceivable that rather than contributing to the phenotype, the changes in miR156 levels are a consequence of the lbl1 phenotype. Similarly, while miR169, miR528, and miR529 are more abundant in lbl1, transcript levels for their targets remain unchanged in the mutant. The abaxialized leaf phenotype of lbl1 can however not account for the increased accumulation of miR390, as this small RNA is expressed on the adaxial side of leaf primordia [42]. Instead, upregulation of miR390 in lbl1 is consistent with feedback regulation in the TAS3 ta-siRNA pathway [42] that is not expected to further impact the phenotype of lbl1.
The remaining loci differentially accumulating small RNAs in lbl1 correspond primarily to DNA and retrotransposons (Fig. 4A-B; S1 Dataset). Interestingly, nearly all retrotransposons showing a significant change in 21- and 22-nt small RNA levels generate siRNAs preferentially or exclusively in the mutant (Fig. 4A-B). Moreover, more than 75% of these upregulated siRNA loci belong to the gyma class of Gypsy-like LTR retroelements (S1 Dataset), indicating a role for LBL1 specifically in the silencing of this class of retrotransposons. Recent studies in Arabidopsis revealed that ta-siRNA biogenesis components can act in a hierarchical manner to the transcriptional gene silencing pathway to silence repetitive regions in the genome [43]–[46]. While most repeats in the Arabidopsis genome are repressed at the transcriptional level through the action of PolIV/V, RDR2, and DCL3-dependent 24-nt heterochromatic siRNAs [47], in instances where this canonical silencing pathway is lacking or perturbed, 21- and 22-nt secondary siRNAs trigger the post-transcriptional repression of repetitive elements [43]–[46]. However, the observations presented here reveal an additional layer of small RNA-mediated transposon regulation. The data implies that when LBL1 activity is lost, a subset of gyma retroelements become targets for yet another small RNA pathway generating 21- and 22-nt LBL1-independent siRNAs. Unlike the repeat derived siRNAs generated by ta-siRNA pathway components in Arabidopsis [46], [48], the 21- and 22-nt LBL1-independent siRNAs map to the long terminal repeats (LTR) of the gyma elements. These seem to maintain the repression of these repeats, as gyma transcript levels are unchanged in the mutant.
Differentially expressed 24-nt small RNAs are also largely derived from retrotransposons and DNA transposons (Fig. 4A-B; S1 Dataset). For many of these repeats, siRNA levels are reduced in lbl1, consistent with the hypothesis that LBL1 also functions in the biogenesis of heterochromatic siRNAs associated with transcriptional gene silencing. While a role in RNA-dependent DNA methylation has been proposed for other distantly related members in the Arabidopsis SGS3-like protein family, SGS3 itself is not considered part of this subgroup [49]-[51]. Moreover, whether the SGS3-like proteins affect the accumulation of 24-nt siRNAs remains controversial. Consistent with a role for LBL1 in the production of 24-nt heterochromatic siRNAs, reduced expression of lbl1 in extended transition stage leaves is correlated with demethylation and reactivation of MuDR transposons [52]. Importantly, the loci accumulating fewer 24-nt siRNAs in lbl1 are distinct from the gyma retrotransposons generating increased levels of 21- and 22-nt small RNAs. This implies multiple contributions for LBL1 in the repression of repetitive elements in the genome: one via production of 24-nt siRNAs, and a distinct, not yet fully understood, role in the regulation of gyma retroelements. It also supports the presence of greater complexity in small RNA-mediated transposon silencing pathways, and that such alternate pathways may act preferentially on a select subset of transposon families [43], [45].
To assess the contribution of repeat-derived small RNAs to the lbl1 defects, we next asked whether any non-phased, differentially expressed 21- or 22-nt small RNAs act in trans to regulate developmental genes at the post-transcriptional level in a manner analogous to miRNAs or ta-siRNAs. Assuming that all siRNAs are loaded into an AGO effector complex, we performed a target prediction and PARE analysis for those 21- and 22-nt differentially expressed small RNAs with three or more reads in either the wild-type or lbl1 libraries. With a maximum target score of 4.5 [4], this identified eight genes targeted by the same gyma-derived siRNA (S3 Dataset). However, in contrast to the verified ta-siRNA targets (see above), these genes are not or scarcely expressed (<1 RPKM) in the vegetative apex and retain normal expression upon mutation of lbl1.
We further considered that the epigenetic regulation of transposable elements can influence expression of adjacent genes [53]. We, therefore, tested whether any of the repetitive or intergenic regions differentially accumulating 24-nt small RNAs are positioned in close proximity to a gene within the maize filtered or working gene sets. 106 of the 185 windows differentially accumulating 24-nt siRNAs are positioned within 10 kb of an annotated protein coding gene, but only three genes show a significant difference in transcript levels between wild-type and lbl1 apices (S5 Dataset). GRMZM2G027495 and GRMZM2G016435 show increased expression in lbl1, even though 24-nt siRNAs levels at the adjacent repetitive and intergenic regions are also upregulated. This correlation is opposite to what is expected if silencing at the repeat region spreads into the adjacent gene, suggesting that expression of these genes is indirectly affected upon mutation of lbl1. The third gene, GRMZM2G089713 encodes for SHRUNKEN1. Its expression is significantly increased in lbl1 consistent with a downregulation in 24-nt siRNAs at the adjacent repeat region. However, as SHRUNKEN1 modulates starch levels [54], a contribution to the developmental defects in lbl1 is not immediately obvious.
Taken together, these analyses reveal unexpected crosstalk between small RNA pathways, with LBL1 making multiple unique contributions to the regulation of repeat-associated siRNAs, in addition to functioning in the biogenesis of ta-siRNAs. It is conceivable that, due to the highly repetitive nature of the maize genome, additional silencing pathways were co-opted to maintain genome integrity. The repeat-derived LBL1-dependent siRNAs are, however, unlikely to contribute substantially to the developmental defects of lbl1 mutants. Instead, the data predicts that the essential role for lbl1 in development reflects its requirement for the biogenesis of TAS3-derived tasiR-ARFs and the correct regulation of their ARF3 targets.
The finding that loss of tasiR-ARF activity and the correct regulation of its ARF3 targets underlie the developmental defects of lbl1 mutants is unexpected. Mutations in rgd2 (ago7), which is likewise required for tasiR-ARF biogenesis, are reported to yield a phenotype distinct from lbl1. Plants homozygous for rgd2-R develop narrow strap-like leaves, but these maintain adaxial-abaxial polarity [18]. Based on this difference in phenotype, LBL1 was proposed to have functions other than in tasiR-ARF biogenesis that contribute to maize leaf polarity. The described rgd2-R allele, however, results from a transposon insertion in the first large intron of the gene, presenting the possibility that it is not a complete loss-of-function allele. Moreover, the rgd2-R phenotype has been characterized primarily in the Mo17 inbred background, and natural variation present between B73 and Mo17 is known to affect the phenotype of developmental mutants. We therefore introgressed the rgd2-Ds1 allele (Fig. 5A), which contains a Ds-transposon insertion in the essential PIWI domain and is predicted to completely disrupt protein activity, into the B73 background. The phenotype of rgd2-Ds1 in B73 is more severe than that described for rgd2-R, giving rise to seedlings with reduced thread-like leaves that often arrest shortly after germination (Fig. 5B-C). The thread-like rgd2-Ds1 leaves lack marginal characters, including the saw tooth hairs and sclerenchyma cells, as well as a ligule, macrohairs, and bulliform cells that characterize the adaxial epidermis. In transverse sections, such leaves show a radial symmetry with photosynthetic and epidermal cells surrounding a central vascular bundle (Fig. 5D-E). These defects closely resemble the phenotype of lbl1-rgd1 [17], [55], consistent with our finding that the developmental defects of lbl1 reflect a loss of tasiR-ARF activity. In fact, the phenotype of the rgd2-Ds1 null allele appears slightly more severe than that of lbl1-rgd1, and this is correlated with a more pronounced effect on ARF3 expression. Transcript levels for arf3a and arf3c-e are significantly increased up to 3.5 fold in rgd2-Ds1 seedling apices (Fig. 5F). In addition, expression of arf3b is changed significantly in rgd2-Ds1, albeit less than two-fold.
The fact that mutations in lbl1 and rgd2 condition comparable adaxial-abaxial leaf polarity defects supports the finding that LBL1 contributes to development through the biogenesis of TAS3-derived ta-siRNAs. To confirm that this requirement lies specifically in the production of tasiR-ARFs and the downstream regulation of the ARF3 abaxial determinants, we generated transgenic lines that express either a native or tasiR-ARF-resistant version of arf3a. The latter (arf3a-m) harbors silent mutations in each of the two tasiR-ARF target sites (Fig. 6A). Most plants expressing the native arf3a cDNA from the endogenous arf3a regulatory regions (arf3a:arf3a) are phenotypically normal. Only occasionally do such plants develop slightly narrower leaves and these become less evident as the plant matures. In contrast, transgenic plants expressing the tasiR-ARF insensitive arf3a:arf3a-m transgene displayed pronounced vegetative and reproductive abnormalities (Fig. 6B-I). arf3a:arf3a-m seedlings resemble seedlings homozygous for the weak lbl1-ref allele, and develop half leaves and thread-like abaxialized leaves, as well as leaves with ectopic blade outgrowths surrounding abaxialized sectors on the upper leaf surface [55]; (Fig. 6B-F). However, as such plants matured, their phenotypes became progressively more severe and resembled the phenotypes of lbl1-rgd1. Mature arf3a:arf3a-m plants have a dramatically reduced stature (Fig. 6G-H), and exhibit developmental abnormalities in both male and female inflorescences that result in complete sterility (Fig. 6I). Thus, misregulation of arf3a alone is sufficient to recapitulate phenotypic defects seen in lbl1 and rgd2 mutants [17], [18], [55]. The initial milder phenotypes of arf3a:arf3a-m plants are potentially explained by the fact that only arf3a expression is affected in these plants, as ARF3 has been shown to condition dose-dependent phenotypes in Arabidopsis [10]. These data demonstrate a conserved role for the ARF3 transcription factors in promoting abaxial fate, and confirm our findings from genome-wide analysis that LBL1 contributes to development through the biogenesis of TAS3-derived tasiR-ARFs and the regulation of their ARF3 targets. Moreover, the severe defects conditioned by mutations in lbl1 and rgd2 thus underscore the importance of the tasiR-ARF ARF3 regulatory module to maize development.
The present work reveals substantial diversity in small RNA pathways across plant species, both in the regulation of repeat-associated siRNAs and the spectrum of phased siRNAs. Only TAS loci belonging to the TAS3 family are active in the maize vegetative apex. In other plant species for which genome-wide small RNA analyses were completed, additional phased siRNA loci belonging to either the one-hit or two-hit sub-families were identified [21], [24], [26]–[28], [31], [36], [37], [56]. Some of the observed diversity may reflect variation in small RNA pathways across different tissue types, but our results indicate that essential steps in the one-hit phased siRNA pathway may function distinctly in the maize seedling apex. Whether this reflects a broader difference between monocots and dicots could be resolved by a similar in depth analyses of phased siRNAs in vegetative apices of e.g. rice and Brachypodium.
Our data further shows that loss of tasiR-ARF mediated regulation of ARF3 genes is responsible for the developmental phenotypes of ta-siRNA biogenesis mutants in maize. Mutants affecting ta-siRNA biogenesis display phenotypes that differ widely from species to species. The TAS3 ta-siRNA pathway, including the regulation of ARF3 targets, is conserved throughout land plant evolution, but the population of phased siRNAs and their targets otherwise vary extensively [4], [5], [10]–[19]. Our findings indicate that this diversity in TAS pathways cannot fully account for the phenotypic differences of ta-siRNA biogenesis mutants. As in maize, the developmental defects of Arabidopsis and Medicago ta-siRNA biogenesis mutants can be mimicked by overexpression of a tasiR-ARF insensitive allele of ARF3 [8]–[9], [14]. Nonetheless, in contrast to the severe polarity phenotype of lbl1 leaves [17], Arabidopsis and Medicago ta-siRNA biogenesis mutants exhibit relatively subtle defects in leaf development, giving rise to downward-curled and highly lobed leaves, respectively [2], [12], [14], [57]. The fact that ARF3 proteins act as repressors of the auxin response [58] may be crucial to understanding these diverse phenotypes. Through their effect on the pattern and level of ARF3 accumulation, the ta-siRNA pathway allows the auxin response to be modulated in a precise spatiotemporal manner. While the TAS3 ta-siRNA pathway itself is highly conserved, its expression in time and space seems to vary across organisms. tasiR-ARFs act in the incipient maize leaf to polarize ARF3 expression and establish adaxial-abaxial polarity, whereas tasiR-ARF biogenesis in Arabidopsis and Medicago is delayed until later in primordium development [10], [12], [14], [42]. Moreover, the nature and wiring of auxin responsive gene networks regulated by the ARF3 transcriptional repressor may vary between plants. Indeed, the polarity network in Arabidopsis and maize appears to be wired differently, as reflected in the distinct redundancies between polarity determinants in these species [10], [59]. As such, divergence in the gene networks downstream of the ARF3 transcription factors or the spatiotemporal pattern in which these tasiR-ARF targets act emerge as a testable hypotheses to explain the diverse contributions of the ta-siRNA pathway to development in maize, Arabidopsis, and possibly other plant species as well.
Families segregating the lbl1-rgd1 allele [17] introgressed at least three times into B73 were grown in growth-chambers at 16 hour 28°C/light and 8 hour 24°C/dark cycles. Shoot apices including the meristem and up to 5 leaf primordia were dissected from 2 week-old plants in triplicate. Total RNA was prepared using the mirVana RNA isolation kit (Life Techologies), and 1ug per sample used to generate small RNA libraries using the small RNA-seq kit (Illumina). RNA 3′ and 5′ adapters were ligated in consecutive reactions with T4 RNA ligase. Ligated RNA fragments were primed with an adapter-specific RT primer and reverse transcribed with Superscipt II reverse transcriptase (Life Technologies) followed by eleven cycles of amplification with adapter specific primers. Resulting cDNA libraries were separated on a 6% TBE gel and library fragments with inserts of 18-25p excised. Recovered cDNA libraries were validated by QC on an Agilent Bioanalyzer HiSens DNA chip (Agilent Technologies Inc.) and were sequenced for 50 cycles on the Illumina GAIIx according to Illumina protocols with one sample per lane. The same RNA samples from two biological replicates were also used for RNA deep sequencing by Macrogen Inc, Korea.
Trimmed reads 18 to 26 nt in length were aligned to the maize B73 RefGen_v2 genome (release 5a.57) using Bowtie v0.12.7 [60]. While the observations presented in this study are robust across a wide range of mapping parameters, the specific data presented uses the following standard filtering criteria [28], [31]: only perfectly matched reads were considered and, taking into consideration the characteristics of previously described miRNA and ta-siRNA loci, a maximum of 20 alignments per read were reported. Reads matching known structural RNAs (rRNAs, tRNAs, sn-RNAs and sno-RNAs) from Rfam 10.0 [61] were removed from further analysis. As expected, considering only uniquely mapped reads eliminated several of the developmentally important ta-siRNAs and miRNAs. Whereas allowing 100 alignments per read in both the phased and LBL1-dependent siRNA analyses identified some additional and distinct repeat loci without impacting the overall conclusions.
Using similar criteria as previously described [24], [62], the abundances of small RNA reads in each individual library were calculated using non-overlapping 500 nt windows. Any windows with fewer than 10 reads total, across all libraries were removed from further analysis. For the remaining windows, edgeR [63] was used to model the counts distribution using a negative binomial model with common dispersion estimate. Differentially expressed loci were defined as windows with at least a 2-fold difference in abundance between the wild-type and lbl1 samples, and an adjusted P-value<0.05, corrected according to the method of Benjamini and Hochberg [64]. Differentially expressed 24-nt small RNAs derived from the lbl1 introgression interval were excluded from further analysis. Differential accumulation of total reads in the 18-26 nt size classes between wild-type and lbl1 were calculated using a two-tailed t-test, as the millions of reads sequenced in each size class would follow an approximately normal distribution.
Mapped reads from the three wild-type libraries were normalized by the number of genome-matched reads in the library and pooled into a single database. Reads matching the forward and reverse strand were merged, adjusting for the 2-nt 3′ overhangs generated by DICER processing. Using a sliding window of 500 bp, genomic regions containing at least 5 reads of 21-nt with a phasing distance of exactly 21-nt were identified as candidate phased clusters. To identify those candidate clusters in which the majority of small RNAs are phased, we modified the phasing score calculation of De Paoli et al. [65] to the following: where n is the number of phase cycle positions occupied by at least one small RNA read, C is the number of phase cycles that fit within a 500 bp window (23 for 21nt small RNAs, 22 for 22-nt small RNAs, 20 for 24-nt small RNAs), P is the abundance of reads in each phase cycle, and U is the abundance of out of phase reads in the window. Analogous to previous studies [28], [31], a stringent threshold of P-score≥25 was used to identify phased siRNA clusters.
Target predictions for ta-siRNAs and lbl1-dependent siRNAs were made using Target Finder, allowing a maximum score = 4.5. Scoring was assigned as described previously [4]. PARE libraries were generated from B73 apex tissues as described previously [35]. PARE analysis was performed according to Zhai et al. [31], with the following minor modification: two windows flanking each predicted target site were defined and the total abundance of PARE tags matching to (1) a small window “WS” of 5 nt (cleavage site ±2 nt), and (2) a large window “WL” of 31 nt (cleavage site ±15 nt) calculated. Cleavage sites were filtered to retain only those for which WS/WL≥0.75 and WS≥4.
Reads were trimmed and aligned to the B73 RefGen_v2 genome (release 5a.57) and the annotated exon junctions (version 5b_AGPv2) using TopHat v1.2.0. The majority (78–80%) of the trimmed reads were uniquely mapped and further analyzed for read counts per genes in the two replicates for each wild-type and lbl1 mutant samples. RPKM values were determined using the Cufflinks package Cuffdiff v1.0.3 with default parameters, except that the “minimum alignment count” (-c) was set to 50 [66]. Differentially expressed transcripts were selected as those showing a 2-fold change in expression with a Cuffdiff determined, BH-corrected P-value<0.05. To analyze expression of repetitive elements, RNAseq reads were allowed to map up to 50 times to genome. Any multi-mapping reads were weighted 1/n, where n is the number of possible alignments. A counts table containing the total reads mapped to each repeat locus in each wild-type and lbl1 replicate was analyzed as above.
The MultiSite Gateway System (Invitrogen) was used to create the arf3a:arf3a and the arf3a:m-arf3a constructs. The coding sequences of arf3a (GRMZM2G030710_T01), and the arf3a regulatory regions (2.838 kb promoter and 1.059 kb 3′UTR) were amplified and cloned into pDONR entry vectors. To generate the tasiR-ARF insensitive arf3a version, mutations were introduced at the two tasiR-ARF target sites using megaprimer PCR mutagenesis. Entry clones for each of the two constructs were combined in a single MultiSite Gateway LR recombination reaction with the pTF101 Gateway-compatible maize transformation vector. Positive clones were transferred into Agrobacterium tumefaciens strain EHA101 and transformed into maize by the Iowa State University Plant Transformation Facility.
The rgd2-Ds1 allele [18] was introgressed six times into B73, and segregating families were grown in growth-chambers at 16 hour 28°C/light and 8 hour 24°C/dark cycles. Leaves from mutant and wild-type plants were fixed and embedded as described [67]. Paraplast blocks were sectioned at a thickness of 10 µM and stained with Safranin-O and Fast Green according to Johansen's method.
Shoot apices were dissected from 12 day-old seedlings of lbl1-rgd1, rgd2-Ds1, and their respective non-mutant siblings. Total RNA was prepared using Trizol reagent (Invitrogen) and treated with DNase I (Promega). cDNA from 500 ng of RNA per sample was synthesized using Superscript III First-Strand Synthesis System (Invitrogen) according to manufacturer's protocol. Gene-specific primers were designed (sequence available upon request) for use with iQ SYBR Green Supermix (BioRad) in qPCR. The specificity of all amplification products was determined using dissociation curve analyses. Relative quantification (RQ) values were calculated based on at least three biological and two technical replicates using the 2−ΔCt method, with the ΔCt of glyceraldehyde-3-phosphate dehydrogenase (gapc) as normalization control, taking into consideration the efficiencies of each primer pair as described [68], [69].
High throughput sequencing data, both raw and processed files, has been submitted to the Gene Expression Omnibus and is available upon publication at accession number GSE50557.
|
10.1371/journal.pcbi.1000230 | Optimal Learning Rules for Discrete Synapses | There is evidence that biological synapses have a limited number of discrete weight states. Memory storage with such synapses behaves quite differently from synapses with unbounded, continuous weights, as old memories are automatically overwritten by new memories. Consequently, there has been substantial discussion about how this affects learning and storage capacity. In this paper, we calculate the storage capacity of discrete, bounded synapses in terms of Shannon information. We use this to optimize the learning rules and investigate how the maximum information capacity depends on the number of synapses, the number of synaptic states, and the coding sparseness. Below a certain critical number of synapses per neuron (comparable to numbers found in biology), we find that storage is similar to unbounded, continuous synapses. Hence, discrete synapses do not necessarily have lower storage capacity.
| It is believed that the neural basis of learning and memory is change in the strength of synaptic connections between neurons. Much theoretical work on this topic assumes that the strength, or weight, of a synapse may vary continuously and be unbounded. More recent studies have considered synapses that have a limited number of discrete states. In dynamical models of such synapses, old memories are automatically overwritten by new memories, and it has been previously difficult to optimize performance using standard capacity measures, for stronger learning typically implies faster forgetting. Here, we propose an information theoretic measure of storage capacity of such forgetting systems, and use this to optimize the learning rules. We find that for parameters comparable to those found in biology, capacity of discrete synapses is similar to that of unbounded, continuous synapses, provided the number of synapses per neuron is limited. Our findings are relevant for experiments investigating the precise nature of synaptic changes during learning, and also pave a path for further work on building biologically realistic memory models.
| Memory in biological neural systems is believed to be stored in the synaptic weights. Numerous computational models of such memory systems have been constructed in order to study their properties and to explore potential hardware implementations. Storage capacity and optimal learning rules have been studied both for single-layer associative networks [1],[2], studied here, and for auto-associative networks [3],[4]. Commonly, synaptic weights in such models are represented by unbounded, continuous real numbers.
However, in biology, as well as in potential hardware, synaptic weights should take values between certain bounds. Furthermore, synapses might be restricted to have a limited number of synaptic states, e.g. the synapse might be binary. Although binary synapses might have limited storage capacity, they can be made more robust to biochemical noise than continuous synapses [5]. Consistent with this, experiments suggest that synaptic weight changes occur in steps. For example, putative single synapse experiments show that a switch-like increment or reduction to the excitatory post-synaptic current can be induced by pairing brief pre-synaptic stimulation with appropriate post-synaptic depolarization [6],[7].
Networks with bounded synapses have the palimpsest property, i.e. old memories decay automatically as they are overwritten by new ones [8]–[15]. In contrast, in networks with continuous, unbounded synapses, storing additional memories reduces the quality of recent and old memories equally (see section Comparison to continuous, unbounded synapses). Forgetting of old memories must in that case be explicitly incorporated, for instance via a weight decay mechanism [16],[17]. The automatic forgetting of discrete, bounded synapses allows one to study learning in a realistic equilibrium context, in which there can be continual storage of new information.
It is common to use the signal-to-noise ratio (SNR) to quantify memory storage in neural networks [2],[18]. The SNR measures the separation between responses of the network; the higher the SNR, the more the memory stands out and the less likely it will be lost or distorted. When weights are unbounded, each stored pattern has the same SNR. Storage capacity can then be defined as the maximum number of patterns for which the SNR is larger than some fixed, minimum value.
However, for discrete, bounded synapses performance must be characterized by two quantities: the initial SNR, and its decay rate. Ideally, a memory has a high SNR and a slow decay, but altering learning rules typically results in either 1) an increase in memory lifetime but a decrease in initial SNR [18], or 2) an increase in initial SNR but a decrease in memory lifetime. Optimization of the learning rule is ambivalent because an arbitrary trade-off must be made between these two effects. In this paper we resolve this conflict between learning and forgetting by analyzing the capacity of synapses in terms of Shannon information. We describe a framework for calculating the information capacity of bounded, discrete synapses, and use this to find optimal learning rules.
We model a single neuron, and investigate how information capacity depends on the number of synapses and the number of synaptic states. We find that below a critical number of synapses, the total capacity is linear in the number of synapses, while for more synapses the capacity grows only as the square root of the number of synapses per neuron. This critical number is dependent on the sparseness of the patterns stored, as well as on the number of synaptic states. Furthermore, when increasing the number of synaptic states, the information initially grows linearly with the number of states, but saturates for many states. Interestingly, for biologically realistic parameters, capacity is just at this critical point, suggesting that the number of synapses per neuron is limited to prevent sub-optimal learning. Finally, the capacity measure allows direct comparison of discrete with continuous synapses, showing that under the right conditions their capacities are comparable.
The single neuron learning paradigm we consider is as follows: at each time-step during the learning phase, a binary pattern is presented and the synapses are updated in an unsupervised manner with a stochastic learning rule. High inputs lead to potentiation, and low inputs to depression of the synapses. Note that if we assume that the inputs cause sufficient post-synaptic activity, the learning rule can be thought of as Hebbian: high (low) pre-synaptic activity paired with post-synaptic activity leads to potentiation (depression). After the learning phase, the neuron is tested with both learned and novel patterns, and it has to perform a recognition task and decide which patterns were learned and which are novel. Alternatively, one can do a (supervised) association task in which some patterns have to be associated with a high output, and others with a low output. This gives qualitatively similar results (see Associative learning below).
More precisely, we consider the setup depicted in Figure 1. A neuron has n inputs, with weights wa, a = 1,…,n. At each time-step it stores a n-dimensional binary pattern with independent entries xa. The probability of a given entry in the pattern being high is given by the sparseness parameter p. We set the value of x for the low input state equal to −p, and the high state to q = (1−p), so that the probability density for inputs is given by P(x) = qδ(x+p)+pδ(x−q). Note that 〈x〉 = 0. Using the expression for the SNR below, it can be shown that this is optimal, c.f. [2]. We assume that , as the case is fully analogous.
Each synapse occupies one of W states. The corresponding values of the weight are assumed to be equidistantly spaced around zero, and are written as a W–dimensional vector, e.g. for a 3-state synapse s = {−1,0,1}, while for a 4-state synapse s = {−3,−1,1,3}. In numerical analysis we sometimes saw an increase in information by varying the values of the weight states, although this increase was always small. The state of any given synapse at a given time is described stochastically, by a probability vector π. Each entry of π is the probability that the synapse is in that state (and hence the weight of the synapse takes the corresponding value in the weight look-up table s).
Finally, we note that this setup is of course an abstraction of biological memory storage. For instance, biological coding is believed to be sparse, but the relation between our definition of p and actual biological coding sparsity is likely to be complicated. Our model furthermore assumes plasticity at each synapse and for every input. In some other models it has been assumed that only a subset of the inputs can cause synaptic changes [14]. Our model could in principle include this by defining null inputs that do not lead to plasticity at all. This would lead to two sparsity parameters: the proportion of inputs that induce plasticity and the proportion of plasticity-inducing inputs that lead to actual strengthening of the synapse.
Storage capacity depends on the W×W learning matrices M+ and M−. To find the maximal storage capacity we need to optimize these matrices, and this optimization depends on sparseness, the number of synapses, and the number of states per synapse. Because these are Markov transition matrices, their columns need sum to one, leaving W(W−1) free variables per matrix.
We have studied pattern storage using discrete, bounded synapses. Learning rules for these synapses can be defined by stochastic transition matrices [18],[19]. In this setup an SNR based analysis provides two contradictory measures of performance: the quality of learning (the initial SNR), and the rate of forgetting [18]. With our single measure of storage capacity based on Shannon information, learning rules can be optimized. The optimal learning rule depends on the number of synapses n and the coding sparseness p, as well as on the number of states W. Our analysis was restricted to about 8 states per synapse, although we have no reason to believe that extrapolation to larger numbers would not hold.
Given optimal learning we find two regimes for the information storage capacity: 1. When the number of synapses is small, information per synapse is constant and approximately independent of the number of synaptic states. 2. When the number of synapses is large, capacity per synapse increases linearly with W but decreases as . The critical n that separates the two regimes is dependent on sparseness and the number of weight states. The optimal learning rule for regime 2 has band-diagonal transition matrices, and in the dense case (p = 1/2), these take a particularly simple form, see Equation 16. Capacity of order in the large n limit has been reached in other studies of bounded synapses [10],[21], but has not been exceeded to our knowledge. It remains a challenge to construct a model that does better than this.
The implications for biology depend on the precise nature of single neuron computation. If a neuron can only compute the sum of all its inputs then we might conclude the following. As synapses are metabolically expensive [22], biology should choose parameters such that the number of synapses per neuron does not exceed the critical number much. Although there are currently no accurate biological estimates for either the number of weight states, or the sparsity, for binary synapses with p = 0.005, the critical number of synapses is close to the number of synapses (∼10,000) per neuron in the hippocampus (see Figure 2). However, if the neurons can do compartmentalized processing so that the dendrite is the unit of computation [23], then one could think of this model as representing a single dendrite, and we could conclude that the number of synapses per dendrite might be optimized for information storage capacity. For binary synapses with p = 0.005 choosing the number of synapses to be several hundred is also close to optimal.
Furthermore, our results predict that when synapses are binary, coding is sparse, and learning is optimized, that at equilibrium about 67% of synapses should occupy the low state. This is not far off the experimental figure of 80% [7].
We have directly compared discrete to continuous synapses. For few synapses and dense coding, binary synapses can store up to 0.11 bits of information, which is comparable to the maximal capacity of continuous synapses. However, for sparse coding and many synapses per neuron, the capacity of binary synapses is reduced. Hence, if one considered only information storage, one would conclude that, unsurprisingly, unbounded synapses perform better than binary synapses. However, in unbounded synapses, weight decay mechanisms must be introduced to prevent runaway, so the information storage capacity is necessarily reduced in on-line learning [16],[17]. In contrast, for bounded, discrete synapses with ongoing potentiation and depression, such as those considered here, old memories undergo “graceful decay” as they are automatically overwritten by new memories [8],[9],[12],[13],[15]. Thus discrete, bounded synapses allow for realistic learning with a good capacity.
Finally, it is worth noting that although using Shannon information is a principled way to measure storage, it is unclear whether for all biological scenarios it is the best measure of performance, c.f. [24]. The information can be higher when storing very many memories with a very low SNR, than when storing just a few patterns very well. This might be undesirable in some biological cases. However, if many neurons work in parallel on the same task, it is likely that all information contributes to performance, and thus the total bits per synapse is a useful measure.
To obtain the information capacity numerically, we used Matlab and implemented the following process. For a given number of synaptic states, number of synapses and sparsity, we used Matlab's fminsearchbnd to search through the parameter space of all possible transfer matrices M+ and M−. That is, all matrix elements were constrained to take values between 0 and 1, and all columns were required to sum to 1. For each set of transfer matrices we first obtained the equilibrium weight distribution π∞ as the eigenvector with eigenvalue 1 of the matrix M. Then we computed the means and variances of the output for learned and unlearned patterns from Equations 2 and 1, and further used that . Equations 6 and 3 then gave the information stored about the pattern presented at each time-step. To calculate the total information, this was summed over sufficient time-steps.
In particular, in the case of many weight states (large W) and sparse patterns, the maximization would sometimes get stuck in local maxima. In those cases we did multiple (up to 50) restarts to make sure that the solution found was truly optimal.
Our results can also be compared to the so-called cascade model, which was recently proposed to have high SNR and slow memory decay [10]. In order to compare the cascade model to our results, we created a six-state cascade model using learning matrices that only had transitions according to the state diagram in [10]. These transition rates were then optimized. We found that the information capacity of the optimized cascade model was always larger than a two-state model, but always lower than our six state model with transfer matrix Equation 16. Only when the number of synapses was small (and the information became equal to the integral over the SNR), did the two-state, six-state and cascade models give identical performance. For a higher number of states the results could be different, but this study suggests that, at least for a small number of states, the cascade model is sub-optimal with respect to Shannon information capacity.
Finally, we explored how well the Gaussian approximation worked. We calculated the full multinomial distribution of the total input h and applied an optimal threshold. Because of a combinatorial explosion, this was only feasible for up to 100 synapses. When the information was maximized this way, the information increased to about 0.3 for n = 1 binary synapses storing dense patterns, but for more than n = 10 synapses the results were indistinguishable from the presented theory.
|
10.1371/journal.pntd.0002670 | Deletion of the NSm Virulence Gene of Rift Valley Fever Virus Inhibits Virus Replication in and Dissemination from the Midgut of Aedes aegypti Mosquitoes | Previously, we investigated the role of the Rift Valley fever virus (RVFV) virulence genes NSs and NSm in mosquitoes and demonstrated that deletion of NSm significantly reduced the infection, dissemination, and transmission rates of RVFV in Aedes aegypti mosquitoes. The specific aim of this study was to further characterize midgut infection and escape barriers of RVFV in Ae. aegypti infected with reverse genetics-generated wild type RVFV (rRVF-wt) or RVFV lacking the NSm virulence gene (rRVF-ΔNSm) by examining sagittal sections of infected mosquitoes for viral antigen at various time points post-infection.
Ae. aegypti mosquitoes were fed an infectious blood meal containing either rRVF-wt or rRVF-ΔNSm. On days 0, 1, 2, 3, 4, 6, 8, 10, 12, and 14 post-infection, mosquitoes from each experimental group were fixed in 4% paraformaldehyde, paraffin-embedded, sectioned, and examined for RVFV antigen by immunofluorescence assay. Remaining mosquitoes at day 14 were assayed for infection, dissemination, and transmission. Disseminated infections were observed in mosquitoes as early as three days post infection for both virus strains. However, infection rates for rRVF-ΔNSm were statistically significantly less than for rRVF-wt. Posterior midgut infections in mosquitoes infected with rRVF-wt were extensive, whereas midgut infections of mosquitoes infected with rRVF-ΔNSm were confined to one or a few small foci.
Deletion of NSm resulted in the reduced ability of RVFV to enter, replicate, and disseminate from the midgut epithelial cells. NSm appears to have a functional role in the vector competence of mosquitoes for RVFV at the level of the midgut barrier.
| Rift Valley fever virus (RVFV) is a mosquito-borne virus endemic to Africa. Outbreaks of RVFV have resulted in devastating morbidity and mortality in livestock and humans. A novel RVFV vaccine strain has been developed in which two virulence genes, NSs and NSm, have been deleted from the RVFV genome. Previously, we demonstrated that deletion of NSm also significantly reduced the ability of Aedes aegypti mosquitoes to transmit RVFV. The objective of this study was to track the spread (dissemination) of wild type RVFV (rRVF-wt) and RVFV lacking the NSm virulence gene (rRVF-ΔNSm) through different tissues in the mosquito body over time by staining lengthwise slices of infected mosquitoes with fluorescent antibody specific to RVFV. We found that midgut infections in mosquitoes exposed to rRVF-wt were extensive, whereas midgut infections in mosquitoes infected with rRVF-ΔNSm were confined to only one or a few small foci. Our data provide supporting evidence that the NSm virulence gene has a functional role in mosquitoes by helping RVFV establish an infection in, and escape from, the midgut.
| Rift Valley fever virus (RVFV) (family Bunyaviridae, genus Phlebovirus) is a zoonotic, mosquito-borne virus endemic to Africa. Human illness is typically febrile, but 1–2% of cases develop more severe disease including hepatitis, encephalitis, retinitis, vision loss, jaundice, severe anemia, neurologic manifestations, renal failure, and hemorrhagic fever [1]–[3]. A hallmark of RVFV outbreaks are “abortion storms” among sheep and cattle, with devastating mortality rates in newborn and young animals [4]–[5]. Transmission of RVFV is mosquito-borne. RVFV is registered as a Category A overlap select agent with both the U.S. Department of Agriculture Animal and Plant Health Inspection Service and the Centers for Disease Control and Prevention due to its biothreat potential and ability to cause significant economic losses to the livestock industry as well as substantial human morbidity and mortality [6]–[7]. The recent spread of RVFV to the Arabian Peninsula [8]–[9] serves as a precedent for the potential for this virus to be introduced into the United States or Europe, and highlights the need for preparedness and development of a safe and efficacious human and veterinary vaccine.
The RVFV genome is tripartite, negative sense RNA segments. Of the three segments, the small (S) segment codes for the nucleoprotein and the nonstructural NSs protein, the medium (M) segment encodes the two structural glycoproteins, Gn and Gc, as well as two nonstructural proteins (NSm and NSm-Gn) and the large (L) segment codes for the viral RNA-dependent RNA polymerase. Nonstructural protein NSs is known to inhibit IFN-β, promote degradation of PKR, and suppress host transcription [10]–[12] while NSm is involved in suppression of virus-induced apoptosis [13]. Both NSs and NSm are recognized to function as virulence factors, however, neither NSs nor NSm are individually required in cell culture for efficient virus replication, assembly, or maturation [13]–[16].
A number of vaccines have been developed against RVFV. However these vaccines have been plagued with many problems including poor immunogenicity, difficulties in manufacturing, and post-vaccination abortions and teratogenesis in livestock [4], [17]–[18]). In response to these problems, Bird et al. [19]–[20] developed a novel vaccine based on the deletion of NSs and NSm. Utilizing a reverse genetics system, infectious wild type and deletion mutant RVF viruses were reconstituted in cell culture from three plasmids encoding antigenomic copies of the S, M, and L segments; deletion mutant viruses lacked the NSs gene on the S segment, and/or the NSm gene on the M fragment [19]–[20]. The double deletion mutant RVFV was demonstrated to be protective and immunogenic in rats, mice, and sheep, without producing clinical illness in these animals [19]–[20]. Due to the enhanced safety profile of this vaccine candidate it was recently excluded from the Select Agent regulations and reclassified by the NIH Recombinant Advisory Committee and the CDC as requiring BSL-2 safety precautions [21]. Deletion of NSm alone retained some ability to cause lethal hepatic and neurologic disease in Wistar-Furth rats and has been developed as an animal model for human delayed onset encephalitic disease [22].
The non-essential nature of NSm for growth in vertebrate cell culture or to mammalian pathogenesis prompted investigations into the role of this protein in RVFV infection and replication in the mosquito vector [23]. Indeed, deletion of NSm greatly reduced the infection, dissemination, and transmission rates of RVFV in Aedes aegypti mosquitoes and infection rates in Culex quinquefasciatus mosquitoes [23]. The specific aim of this study was to further characterize midgut infection and escape barriers of RVFV in Ae. aegypti infected with reverse genetics-generated wild type RVFV (rRVF-wt) or RVFV lacking the NSm virulence gene (rRVF-ΔNSm) by examining sagittal sections of infected mosquitoes for viral antigen by immunofluorescence at various time points post-infection.
The Ae. aegypti Rexville D mosquito strain used was an isofemale line derived from a population of Ae. aegypti collected as larvae in San Juan, Puerto Rico (Rexville) in 1991 [24]. Mosquitoes were double-caged in screened paperboard pint containers inside environmental chambers at 28°C and approximately 95% relative humidity. Reverse genetics-generated viruses, rRVF-wt and rRVF-ÄNSm, were used in this study [19], [22].
To maximize infectivity to mosquitoes, freshly-harvested rRVF-wt and rRVF-ÄNSm virus strains were used in the infectious blood meal. Three days prior to the infectious blood-feed, one T-75 flask each of Vero cells was inoculated with either rRVF-wt or rRVF-ÄNSm at a multiplicity of infection (MOI) of 0.1. On Day 3 post-infection, cell-culture supernatant was harvested and clarified for use in the infectious blood meal. Because differences in virus concentration may affect mosquito vector competence [23], we attempted to equalize the virus titers of rRVF-wt and rRVF-ΔNSm in the mosquito blood meals. RNA was extracted from clarified supernatant from flasks containing freshly-grown virus and quantified by qRT-PCR using novel primers and a probe targeting the polymerase gene: 4108F (5′-TTT AGA GAC CGT TTG AAC ATA CC-3′), 4217R (5′-GCA ATG CGC AAC AAT ATT TCT-3′) and probe, 4161P (5′FAM-TC CAG AGG TGC TCT ATC GGG CTC C-3′). The observed difference in RNA copies/mL between rRVF-wt and rRVF-ΔNSm were corrected by diluting the rRVF-wt virus supernatant in Dulbecco's modified Eagle's media by a factor of 2.8 prior to preparing the infectious blood meals.
The infectious blood meals were prepared by mixing two parts washed defibrinated calf blood with two parts virus and one part FBS+10% sucrose. A virus-negative blood meal contained cell culture media in place of virus-positive cell supernatant. Blood was warmed to 37°C in a water bath. Adult 8- to 10-day-old Ae. aegypti mosquitoes starved for 27-hours were administered an infectious RVFV blood meal containing either rRVF-wt or rRVF-ΔNSm on blood-soaked cotton balls. Screened pint cups containing 100–150 female Ae. aegypti were placed inside plastic bins inside a 28°C environmental chamber. One blood-soaked cotton ball was placed on each carton for 25 minutes. Five hundred microliters of each blood meal and 500 µl of virus seed brought to 20% FBS were frozen at −80°C for later quantification. Following the blood meal, mosquitoes were anesthetized by freezing at −20°C for 1 min, and fully engorged mosquitoes were sorted over ice inside of a glove box; only fully-engorged mosquitoes were used for the experiment. Engorged mosquitoes were placed into screened 3.8L paperboard cartons and supplied with 5% sugar solution. Paperboard cartons were placed inside a 30.5 cm3 metal cage inside the environmental chamber for double containment. On days 0, 1, 2, 3, 4, 6, 8, 10, 12, and 14 post-infection, between 10 and 16 mosquitoes from each experimental group were harvested for pathology and frozen at −80°C. Remaining mosquitoes at day 14 were processed for vector competence and analyzed statistically as described by Crabtree et al. [23]. Plaque titrations were performed on Vero cells as described by Miller et al. [25], with the second overlay on day three. All work involving manipulations with infectious virus or infected mosquitoes was performed in BSL3+ containment.
The complete protocol used for fixing and antibody staining of paraffin-embedded sections of mosquitoes is described in detail by Kading et al. [26]. Spot slides of rRVF-wt-infected and uninfected Vero cells, as well as sections of rRVF-wt- and rRVF-ΔNSm-infected and uninfected Ae. aegypti from various time points were tested simultaneously served as positive and negative controls [26]. Embedding and sectioning was performed by Colorado HistoPrep. Mosquitoes were arranged vertically, four per section, with two sections per slide and 10 slides per block. The methodology for the staining of control spot slides and head squashes was described previously [27]. Mouse anti-RVFV strain ZH501 hyperimmune ascetic fluid diluted 1∶2500 was used as a primary antibody for immunofluorescence assays on control spot slides and head squashes; a dilution of 1∶1600 was used on paraffin-embedded sections. Goat anti-mouse IgG-Alexa 488 (Invitrogen, Baltimore, MD) diluted 1∶2000 served as the secondary antibody conjugate.
To quantify and compare virus dissemination to different mosquito tissues over time, the dissemination index was employed [28]–[29]. The dissemination index is based on the infection status of particular tissues that are nearly always infected in specimens with a disseminated infection. The presence or absence of viral antigen was scored in the following six tissues and the number of antigen-positive tissues was divided by six to give the dissemination index: ommatidia; fat body in the head, thorax and abdomen; salivary glands; and thoracic ganglia. A dissemination index of 1.0 indicated that all examined tissues were positive, whereas an index of zero indicated that dissemination had not yet occurred. Scatterplots of dissemination indices were generated in Prism. Overlapping points were nudged horizontally for visibility.
Aedes aegypti mosquitoes received an artificial blood meal containing either 7.6 log10 pfu/mL rRVF-wt or 7.9 log10 pfu/mL rRVF-ΔNSm. Mosquitoes were harvested for pathology daily for the first five days and subsequently every other day until 14 days following the infectious blood meal (Table 1). Up to 12 mosquitoes per time point for each virus were submitted for paraffin-embedding and sectioning. On day 14, infection, dissemination and transmission rates were assayed for the remaining 25 mosquitoes infected with rRVF-wt and 30 mosquitoes infected with rRVF-ΔNSm (Table 2). Of the 21/25 mosquitoes remaining infected with rRVF-wt 14 days post-infection, the average titer was 5.5±0.30 log10 pfu. In contrast, 7/30 mosquitoes exposed to rRVF-ΔNSm were infected 14 days after the infectious blood meal; these mosquitoes had a titer of 2.87 log10 ±0.58 pfu. Accounting for virus uptake on day 0, these rates of sustained infection differed significantly (OR = 15.2, 95% CI 4.0–57.7). The difference in these log10 titers was 2.7 (95% CI 2.1–3.2), statistically significantly different from 0. Notably, the average titer in a mosquito exposed to rRVF-wt increased from 5.0 log10 pfu in the blood meal to 5.5 log10 pfu on day 14, whereas the average virus titer in a mosquito exposed to rRVF-ΔNSm dropped from 5.8 log10 pfu in the blood meal to 2.87 log10 pfu on day 14.
The dynamics of infection between rRVF-wt and rRVF-ΔNSm was distinctly different between the two virus strains. Posterior midgut infections in mosquitoes infected with rRVF-wt were extensive (Fig. 1 A and B) whereas infection of the posterior midgut in mosquitoes infected with rRVF-ΔNSm was confined to one or a few small foci (Fig. 1C and D). Disseminated infections were observed in mosquitoes by three days post infection for both viruses (Table 1). However, accounting for number of days post infection, infection rates for rRVF-ΔNSm were statistically significantly less than that of rRVF-wt (odds ratio (OR) = 0.20, 95% CI = 0.10–0.41) (Table 1). When modeling the dissemination rates, a statistically significant interaction was found between days post infection and virus type (OR = 0.53, 95% CI = 0.37–0.76). Since the dissemination rate of rRVF-wt increased with time but only three dissemination events were documented among mosquitoes infected with rRVF-ΔNSm, the difference between the two viruses also increased as days post infection increased. Similar results were found for the dissemination rates of only those mosquitoes that became infected (OR = 0.48, 95% CI = 0.32–0.72) (Table 1).
Other than the midgut, RVFV antigen was visible in a variety of tissues including the thoracic and abdominal fat body (Fig. 1A,C,D), thoracic ganglia (Fig. 2A), salivary glands (not shown), intussuscepted foregut (not shown), the ommatidia, fat body and nervous tissue of the head (Fig. 2B), tracheal cells of the ovaries (Fig. 2C), ovariole sheath and follicular epithelium (Fig. 2D), and in five specimens, the hindgut (not shown). RVFV antigen was not observed in the Malpighian tubules (Fig. 1A) or flight muscles. The time of dissemination varied from mosquito-to-mosquito but ranged between three to eight days post infection in mosquitoes exposed to rRVF-wt (Table 1, Fig. 3). Dissemination to other tissues was rapid for both rRVF-wt and rRVF-ΔNSm once the virus escaped from the midgut, as evidenced by the majority of specimens having a dissemination index of close to either zero, indicating no dissemination, or one, indicating all counted tissues were infected (Fig. 3).
Infection, dissemination and transmission rates of Ae. aegypti exposed to rRVF-wt were all significantly higher than those of Ae. aegypti exposed to rRVF-ΔNSm (Table 2). None of the day 14 mosquitoes infected with rRVF-ΔNSm developed a disseminated infection or transmitted virus that was detectable by plaque titration.
Through transmission experiments and histological examination of infected mosquitoes, we have demonstrated that Ae. aegypti mosquitoes infected with RVFV lacking the NSm nonstructural protein gene have significantly lower infection, dissemination, and transmission rates than mosquitoes infected with wild-type RVFV [23, and this study] The barriers to infection in mosquitoes infected with rRVF-ΔNSm appear to be in the ability of the virus to replicate in and escape from epithelial cells of the posterior midgut. Small foci of RVFV antigen were visible in the midgut epithelial cells two days post infection for mosquitoes infected with the rRVF-ΔNSm deletion mutant virus. By day 14 post infection, rRVF-ΔNSm infection rates were not over 80%, and virus had disseminated from the midgut in only three mosquitoes (n = 81, 3.7%) (Table 1). We are confident that these three disseminations are not artifacts, as evidenced by the specific cell-associated staining depicted in Figure 1 and the clarity of our positive and negative controls (data not shown). In contrast, midgut infections in mosquitoes exposed to rRVF-wt were easily observable one day post-infection, and by histological examination of specimens, 100% infection and dissemination rates were observed on days 10, 12, and 14 post-infection (Table 1). An infection rate of 21/25 (84%) was also determined at day 14 by plaque assay for rRVF-wt-exposed mosquitoes in the transmission experiment (Table 2), for a combined day 14 infection rate of 31/35, or 88.6%. Replication of rRVF-ΔNSm was reduced to small foci in the midgut, compared to the extensive infections present in the midguts of mosquitoes infected with rRVF-wt (Fig. 1). Therefore, the absence of NSm resulted in a reduction in the ability of RVFV to establish an infection, and escape from the midgut.. These results are consistent with the findings of Crabtree et al. [23], who found that while infection and dissemination were severely inhibited by deletion of NSm, it was not blocked completely. In that study, a single mosquito (n = 129) exposed to rRVF-ΔNSm developed a disseminated infection and transmitted virus [23].
On a cellular level, it is not yet known how NSm promotes the ability of RVFV to establish an infection in and disseminate from the mosquito midgut. The biological function of NSm has remained largely obscure, as this protein is not necessary for viral growth in cell culture [14], [30], nor for pathogenesis in a mammalian host [22]. Won et al. [13] demonstrated that NSm suppressed virus-induced apoptosis through inhibiting STP-induced caspase 8 and caspase 9 activities. This antiapoptotic activity occurred in the absence of other viral proteins. More recently, Engdahl et al. [31] identified several murine proteins that interacted with the NSm protein of RVFV, providing additional clues regarding the role of NSm in RVFV infection. The strongest protein-protein interactions were found with the cleavage and polyadenylation specificity factor subunit 2 (Cpsf2) which functions in pre-mRNA 3′ end processing and formation, the peptidyl-prolyl cis-trans isomerase (cyclophilin)-like 2 protein (Ppil2) which has multiple functions spanning intracellular protein trafficking and regulation of chemotactic responses such as cell-mediated immunity and inflammation, and the 25 kDa synaptosome-associated protein (SNAP-25) which is involved in vesicle docking, membrane fusion and exocytosis, and Ca2+-dependent neurotransmission in neuronal cells of the brain [31]. It is unclear how these results translate to the role of NSm in the midgut of a mosquito vector. RNAi and autophagy have been recognized as anti-arboviral innate immune responses in insects [32]–[33]. The involvement of NSm in counteracting one or both of these processes seems reasonable, and warrants further investigation.
Our results regarding the timing of dissemination and distribution of RVFV antigen to various mosquito tissues are also consistent with those of previous studies of RVFV infections in mosquitoes. We observed virus dissemination from the posterior midgut by three days post infection for both rRVF-wt and rRVF-ΔNSm. Early dissemination has previously been reported for RVFV. Faran et al. [34] found that 6% of Culex pipiens L. infected with RVFV ZH-501 had a disseminated infection as early as 12 hours post infection, 9% had disseminated by 24 h and 22% had disseminated by 48 h. These results were confirmed by Romoser et al. [29] who observed RVFV dissemination in Cx. pipiens one day post infection. Detailed data for RVFV infections in Aedes species are less complete, but Romoser [28] recorded that 87.5% of Ae. mcintoshi Huang mosquitoes infected with RVFV (Kenyan strain C6/36—6/12/86) had a disseminated infection by three days post infection. The mechanism for such rapid dissemination into the hemocoel was not clear, although it reportedly occurred prior to replication in the midgut epithelium [34].
Tissue tropisms of RVFV in Ae. aegypti with disseminated infections were also consistent with those reported for other mosquito species. Every tissue in which we detected RVFV antigen has been reported as commonly infected in Cx. pipiens [29] and Ae. mcintoshi [28]. Further, the sporadic timing of dissemination and the rapidity with which RVFV infected other tissues is also congruent with what has been reported previously [28]–[29]. We observed virus dissemination between days three and eight post-infection for rRVF-wt, with all specimens having a disseminated infection by day 10 post infection. For Cx. pipiens and Ae. mcintoshi, the earliest dissemination was observed one and three days post-infection, respectively, and occurred throughout the observation periods, with some specimens of each species not having disseminated infections until as late as 21 days after the infectious blood meal [28]–[29]. However in this study and in those previous, dissemination indices tended to cluster either at zero or close to one, indicating that dissemination to the various mosquito tissues was rapid once the virus entered the hemocoel. Tissue tropisms between rRVF-wt and rRVF-ΔNSm were not observably different.
We frequently also observed RVFV antigen associated with tracheal cells in the ovaries (Fig. 2C). The observation of RVFV in the tracheal system is not new. Romoser et al. [35] reported that RVFV could infect the trachea and tracheoles, and provided supporting evidence that the trachea could serve as a conduit for virus dissemination between the midgut epithelium and the hemocoel. This observation has also been reported for Ae. mcintoshi in Kenya [36]. Tracheal conduits have also been hypothesized to facilitate virus dissemination from the midgut in dengue 2 virus (DENV-2) infections in Ae. aegypti [37]. In that study, DENV-2 antigen was detected in portions of the tracheal system in approximately 35% of mosquitoes between days two and seven post-infection. Presence of viral antigen in the tracheal system was primarily in the abdominal cavity, and was strongly correlated with virus dissemination from the midgut between days two and five post-infection [37]. While this phenomenon was commonly observed for the Chetumal strain of Ae. aegypti, DENV-2 viral antigen was only rarely associated with the trachea of infected Rexville D Ae. aegypti, the strain used in this study. Infection of the ovarian tissues and the potential for vertical transmission through the tracheal cells in the ovaries is not known.
In conclusion, we have provided histological and virological evidence for the reduction in infection, dissemination, and transmission rates of RVFV lacking the NSm gene. Deletion of NSm results in the reduced ability of RVFV to replicate in and disseminate from the midgut epithelial cells. This report together with the report by Crabtree et al. [23] comprise the first description of a functional role for NSm in the vector competence of mosquitoes for RVFV.
|
10.1371/journal.pntd.0002148 | Antigenicity and Protective Efficacy of a Leishmania Amastigote-specific Protein, Member of the Super-oxygenase Family, against Visceral Leishmaniasis | The present study aimed to evaluate a hypothetical Leishmania amastigote-specific protein (LiHyp1), previously identified by an immunoproteomic approach performed in Leishmania infantum, which showed homology to the super-oxygenase gene family, attempting to select a new candidate antigen for specific serodiagnosis, as well as to compose a vaccine against VL.
The LiHyp1 DNA sequence was cloned; the recombinant protein (rLiHyp1) was purified and evaluated for its antigenicity and immunogenicity. The rLiHyp1 protein was recognized by antibodies from sera of asymptomatic and symptomatic animals with canine visceral leishmaniasis (CVL), but presented no cross-reactivity with sera of dogs vaccinated with Leish-Tec, a Brazilian commercial vaccine; with Chagas' disease or healthy animals. In addition, the immunogenicity and protective efficacy of rLiHyp1 plus saponin was evaluated in BALB/c mice challenged subcutaneously with virulent L. infantum promastigotes. rLiHyp1 plus saponin vaccinated mice showed a high and specific production of IFN-γ, IL-12, and GM-CSF after in vitro stimulation with the recombinant protein. Immunized and infected mice, as compared to the control groups (saline and saponin), showed significant reductions in the number of parasites found in the liver, spleen, bone marrow, and in the paws' draining lymph nodes. Protection was associated with an IL-12-dependent production of IFN-γ, produced mainly by CD4 T cells. In these mice, a decrease in the parasite-mediated IL-4 and IL-10 response could also be observed.
The present study showed that this Leishmania oxygenase amastigote-specific protein can be used for a more sensitive and specific serodiagnosis of asymptomatic and symptomatic CVL and, when combined with a Th1-type adjuvant, can also be employ as a candidate antigen to develop vaccines against VL.
| Life-long immunity to leishmaniasis in recovered patients has inspired the development of vaccines against disease. The present study aimed to evaluate a non-described hypothetical Leishmania amastigote-specific protein, identified by an immunoproteomic approach in L. infantum, attempting to select a new candidate antigen for specific serodiagnosis and a vaccine against visceral leishmaniasis (VL). The recombinant protein (rLiHyp1) was recognized by antibodies from sera of asymptomatic and symptomatic canine visceral leishmaniasis (CVL), but presented no cross-reactivity with sera of vaccinated dogs, those with Chagas' disease or healthy animals. In addition, the rLiHyp1 plus saponin was able to induce a Th1 response, which was based on the production of high levels of IFN-γ, IL-12, and GM-CSF after in vitro stimulation in BALB/c mice. The protective efficacy of rLiHyp1 plus saponin was evaluated in mice challenged with L. infantum promastigotes. Challenged and vaccinated mice showed significant reductions in the number of parasites in all evaluated organs, and the protection was associated with a Th1-type response. Therefore, the present study reveals a new potential candidate for the improvement of serodiagnosis of CVL, as well as an effective vaccine candidate against VL.
| Visceral leishmaniasis (VL) caused by Leishmania donovani and L. infantum/L. chagasi represents an important disease in the world, leading to nearly 500,000 new cases and 50,000 deaths annually. The first choice of treatment for all forms of leishmaniasis is still based on the use of the parenteral administration of pentavalent antimonials; however, increased parasite resistance and several side effects reported by patients have been important problems [1], [2]. Liposomal amphotericin B is effective but is too expensive for most patients [3]. Results from clinical trials using oral miltefosine are encouraging; however, therapy is linked to both potential toxicity and teratogenicity, and should not be given to childbearing-age women [4]. Therefore, the development of new strategies to prevent leishmaniasis has become a high priority.
The evidence of life-long immunity to leishmaniasis has inspired the development of prophylactic vaccination protocols against the disease, but few have progressed beyond the experimental stage. Most experimental vaccines have focused on the mouse model for cutaneous leishmaniasis. Several studies have demonstrated the Type-1 cells mediated immunity-dependence for protective responses against disease. Moreover, Th1 cells response has also been correlated with protection against VL [5]. In this context, the protective immunity in murine VL primarily depends on an IL-12-driven Th1 cells response, leading to an increased IL-2 and IFN-γ production. Substantial uptake of inducible NO syntase by IFN-γ generates NO from splenic and liver cells, thereby controlling parasite multiplication in these organs [6], [7]. By contrast, TGF-β, IL-10, and IL-13 represent major disease promoting cytokines, leading in turn to the suppression of the Th1 response [8]. Low levels of IL-4 commonly enhance vaccine-induced protection by indirectly increasing IFN-γ production by T cells [9].
In recent decades, the majority of studies have focused on Leishmania promastigote antigens for vaccine development [10], [11], [12], [13]; however, amastigote antigens also seem to be appropriate targets for the immune responses elicited by vaccines, given that after a few hours of infection and during the active disease, this parasite stage becomes exposed to the host immune system [14]. In addition, the fact that promastigotes can be easily cultured in vitro, as opposed to axenic amastigotes, has hampered the identification of amastigote-specific stage antigens [15].
In the present work, a hypothetical amastigote-specific L. infantum protein, LiHyp1 (XP_0014689 41.1), which has been identified by an immunoproteomic approach, was recognized by antibodies present in sera samples of dogs with asymptomatic and symptomatic VL [16]. The amastigote-specific Leishmania protein gene (LiHyp1) is predicted to encode a protein with a theoretical molecular weight of 36.6 kDa. An in silico sequence comparison revealed that LiHyp1 belongs to the super-oxygenase family in Leishmania, and is an alkylated DNA repair protein. The ability of the recombinant protein (rLiHyp1) to induce protection against infection with virulent L. infantum promastigotes was assessed in BALB/c mice. The results showed that rLiHyp1 was antigenic and specifically recognized by canine VL (CVL) sera, whereas a Th1 response, induced by immunization of a combination of rLiHyp1 and saponin, was able to confer protection against L. infantum. This protection correlated with a Leishmania antigens-specific and IL-12-dependent IFN-γ production, mediated mainly by CD4 T cells, as well as by a diminished production of parasite-specific IL-4 and IL-10.
Thus, the present study demonstrates that this unique amastigote-specific protein, a member of the super-oxygenase family in Leishmania, can be a new candidate for the improvement of serodiagnosis of CVL and, when associated with a Th1-type adjuvant, to develop an effective vaccine against VL.
Experiments were performed in compliance with the National Guidelines of the Institutional Animal Care, and Committee on the Ethical Handling of Research Animals (CEUA) from the Federal University of Minas Gerais (Law number 11.794, 2008), with code number 043/2011. Sera samples used in this study were kindly provided by Prof. Fernando Aécio de Amorim Carvalho (Department of Pharmacology and Biochemistry, UFPI) and Prof. Maria Norma Melo (Department of Parasitology, Institute of Biological Sciences, UFMG).
Female BALB/c mice (8 weeks age) were obtained from the breeding facilities of the Department of Biochemistry and Immunology, Institute of Biological Sciences, UFMG, and were maintained under specific pathogen-free conditions. Experiments were carried out using the L. infantum (MOM/BR/1970/BH46) strain. Parasites were grown at 24°C in Schneider's medium (Sigma, St. Louis, MO, USA), supplemented with 10% heat-inactivated fetal bovine serum (FBS, Sigma), 20 mM L-glutamine, 200 U/mL penicillin, and 100 µg/mL streptomycin, at pH 7.4. The soluble L. infantum antigenic extract (SLA) was prepared from 1×1010 stationary-phase promastigote cultures (5–7 day-old), as described [17]. Parasites were kindly provided by Prof. Maria Norma Melo.
The LiHyp1 (XP_001468941.1) nucleotide and amino acid sequences used in this study were obtained from the National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov). The local alignment of the LiHyp1 sequence against the available complete genomes of other organisms was performed by BLAST. Analyses of basic physical and chemical properties, as well as phylogenetic analysis, were performed in a TriTrypDB database (http://tritrypdb.org). The recombinant protein (rLiHyp1) was obtained after having cloned a DNA L. infantum fragment containing the LiHyp1 coding region. Initially, genomic DNA was extracted by a phenol:chloroform extraction, as described [17], and it was used as a template. Forward (5′-GAAGGATCCAGCATGTCTATCGTGTCGAG-3′) and reverse (5′-GGAAAGCTTCGCTTGCGGCGTCACGTGAGC-3′) primers were designed according to the DNA sequence of the ORF described in the L. infantum genome sequence database (LinJ.35.1290). The PCR product was cloned into the pGEM-T easy vector confirmed by sequencing and transferred to the pET21a expression vector (Novagen), using the BamHI and HindIII restriction enzymes included in the primers for this purpose (underlined). Recombinant plasmid was transformed into Escherichia coli BL21 (DE3). The recombinant protein expression was performed by adding 0.5 mM IPTG (isopropyl-β-D-thiogalactopyranoside, Promega, Montreal, Canada) for 4 h at 37°C, when cells were lysed by a homogenizer after five passages through the apparatus. The product was centrifuged at 13.000× g for 20 min at 4°C, while the rLiHyp1, containing a tag of 6× residues of histidine, was purified under non-denaturing conditions, using a 5 mL HIS-Trap column (GE Healthcare Life Science) attached to an FPLC (GE Healthcare Life Science) system. After purification, the recombinant protein was passed through a polymyxin-agarose column (Sigma) to remove residual endotoxin content.
To evaluate the antigenicity of rLiHyp1, sera samples from healthy (n = 37), vaccinated with Leish-Tec® (n = 18), T. cruzi experimentally infected (n = 18), asymptomatic (n = 19) and symptomatic L. infantum-infected dogs (n = 15) were used. All animals were considered symptomatic when three or more of the following symptoms were present: loss of weight, alopecia, adenopathy, onychogryphosis, hepatomegaly, conjunctivitis and exfoliate dermatitis on the nose, tail and ear tips; and asymptomatic when they were free from clinical symptoms. In the infected animals, the diagnosis of the disease was defined when amastigotes were seen in Giemsa stained smears of bone marrow aspirates or promastigotes were identified on culture of peripheral blood or bone marrow aspirates. Sera were considered positive when tested by indirect immunofluorescence. A titration curve was performed to determine the best protein concentration and antibody dilution to perform ELISA. Plates (Falcon) were sensitized with rLiHyp1 (1.0 µg/well) or SLA (0.5 µg/well) for 18 h at 4°C. Free binding sites were blocked with a PBS-Tween 20 0.05% (PBST) and 5% casein solution for 2 h at 37°C. After five washes with PBST, plates were incubated with 100 µL of canine sera for 1 h at 37°C. Serum samples were diluted 1∶200 in PBST and 0.5% casein. After, plates were washed seven times with PBST and incubated with 1∶10.000 anti-dog IgG antibody (Sigma, St. Louis, USA) horseradish peroxidase conjugated. The reaction was developed through incubation with H2O2, along with orto-phenylenediamine and citrate-phosphate buffer pH 5.0, for 30 min in the dark, and was stopped by adding H2O2 2 N. Optical densities were read at 492 nm in an ELISA microplates spectrophotometer (Molecular Devices, Spectra Max Plus, Concord, Canada).
For immunoblotting experiments, the recombinant protein was submitted to a 10% SDS-PAGE and blotted onto a nitrocellulose membrane (0.2 µm pore size, Sigma, St. Louis, USA). Membranes were blocked with PBST and 5% casein solution, and were incubated for 2 h at 37°C before the first incubation with a pool of sera samples of asymptomatic CVL, diluted 1∶100 in PBST. Peroxidase conjugated anti-dog IgG (1∶5.000) was used as a second antibody (Sigma). Reactions were revealed by adding chloronaphtol, diaminobenzidine, and H2O2.
Mice (n = 8, per group) were vaccinated subcutaneously in their left hind footpad with 25 µg of rLiHyp1 associated with 25 µg of saponin (Quillaja saponaria bark saponin, Sigma), with adjuvant or only diluent (PBS). Three doses were administered at 2-week intervals. Four weeks after the final immunization, animals (n = 4, per group) were euthanized for the analysis of the immune response elicited by vaccination. At the same time, the remaining animals were infected subcutaneously in the right hind footpad, with virulent 1×107 stationary-phase promastigotes of L. infantum, when 10 weeks after the animals were euthanized, and the liver, spleen, bone marrow (BM), and the paws' draining lymph nodes (dLN) were collected to determine parasite burden and evaluation of the immune response.
The liver, spleen, BM, and dLN were collected for parasite quantification, following a limiting-dilution protocol [18]. Briefly, the organs were weighed and homogenized using a glass tissue grinder in sterile PBS. Tissue debris was removed by centrifugation at 150× g, and cells were concentrated by centrifugation at 2000× g. Pellets were resuspended in 1 mL of Schneider's insect medium supplemented with 20% FBS. Two hundred and twenty microliters were plated onto 96-well flat-bottom microtiter plates (Nunc, Nunclon®, Roskilde, Denmark) and diluted in log-fold serial dilutions in supplemented Schneider's medium with a 10−1 to 10−20 dilution. Each sample was plated in triplicate and read 7 days after the beginning of the culture at 24°C. Pipette tips were discarded after each dilution to avoid carrying adhered parasites from one well to another. Results are expressed as the negative log of the titer (i.e., the dilution corresponding to the last positive well) adjusted per microgram of tissue.
Splenocyte cultures and cytokine assays were performed before infection and at 10th week after challenge, as described [17]. Briefly, single-cell preparations from spleen tissue were plated in duplicate in 24-well plates (Nunc) at 5×106 cells per mL. Cells were incubated in DMEM medium (non-stimulated, background control), or separately stimulated with SLA (25 µg mL−1) or rLiHyp1 (20 µg mL−1), at 37°C in 5% CO2 for 48 h. IFN-γ, IL-4, IL-10, IL-12, and GM-CSF levels were assessed in the supernatants by a sandwich ELISA provided in commercial kits (BD OptEIA TM set mouse IFN-γ (AN-18), IL-12 and GM-CSF; Pharmingen, San Diego, CA, USA; and Murine IL-4 and IL-10 ELISA development kits; PeproTech®, São Paulo, Brazil); following manufacturer's instructions. In order to block IL-12, CD4, and CD8 mediated T cell cytokine release, spleen cells of mice vaccinated with rLiHyp1 plus saponin and challenged with L. infantum were in vitro stimulated with SLA (25 µg mL−1), and incubated in the presence of 5 µg mL−1 of monoclonal antibodies (mAb) against mouse IL-12 (C17.8), CD4 (GK 1.5), or CD8 (53-6.7). Appropriate isotype-matched controls – rat IgG2a (R35-95) and rat IgG2b (95-1) – were employed in the assays. Antibodies (no azide/low endotoxin™) were purchased from BD (Pharmingen, San Diego, CA, USA).
The statistical analysis was made using the GraphPad Prism software (version 5.0 for Windows). Statistical analysis with the vaccinated and/or infected mice was performed by one-way analysis of variance (ANOVA), using the Bonferroni's post-test for multiple comparisons of groups. Receiver Operating Characteristic (ROC) curves were used to analyze the data obtained using sera samples of dogs. Statistical analysis between CVL and the control groups were performed by one-way ANOVA using Tukey's multiple comparison test. Differences were considered significant when P<0.05. Data of shown in this study are representative of two independent vaccination' experiments, which presented similar results.
In the present study, a putative member of the super-oxygenase family in Leishmania was fused as a recombinant protein to an N-terminal 6× histidine-tag and expressed in E. coli. The recombinant protein (rLiHyp1) was purified by nickel affinity chromatography (Fig. 1A), and tested for serodiagnosis of CVL. Initially, a pool of sera of asymptomatic dogs was able to recognize the rLiHyp1 by immunoblotting analysis, as seen in Fig. 1B.
Sera were individually tested in ELISA against rLiHyp1, and the results indicated that all sera samples of symptomatic dogs, and 18 out of 19 samples of asymptomatic CVL animals were able to recognize the recombinant protein. In contrast, antibodies from T. cruzi-infected, Leish-Tec® vaccinated or healthy dogs did not react with the rLiHyp1 protein (Fig. 1C). To determine the diagnostic performance of rLiHyp1 for CVL, Receiver-Operating Characteristic (ROC) curves were constructed to determine area under curve (AUC) and sensitivity and specificity values in the experiments. In the results, it was observed that the performance of rLiHyp1 proved to be highly effective in order to identify sera samples of symptomatic and asymptomatic CVL, and also to differentiate them in relation to the other sera samples employed in this study (Fig. 1D and 1E, respectively).
The immunogenicity of the rLiHyp1 was evaluated in BALB/c mice, 4 weeks after the last vaccine dose. Following in vitro stimulation with rLiHyp1, spleen cells from vaccinated mice significantly produced higher levels of IFN-γ, IL-12, and GM-CSF than those secreted by spleen cells from control mice (saline and saponin groups). No increase in IL-4 and IL-10 production could be observed in any experimental group, after stimulation with rLiHyp1 (Fig. 2A). The ratio between IFN-γ/IL-4 and IFN-γ/IL-10 levels; as well as between IL-12/IL-4 and IL-12/IL-10 levels showed that vaccinated animals presented an elevated Th1 immune response after rLiHyp1-stimulus (Fig. 2B and 2C, respectively). In addition, mice vaccinated with rLiHyp1 plus saponin presented an rLiHyp1-specific humoral response, with the predominance of IgG2a isotype (Fig. 2D).
Next, the present study analyzed whether the immunization with the rLiHyp1 plus saponin was able to induce protection against L. infantum. The infection was followed up over a 10-weeks period, when the parasite burden in the liver, spleen, BM, and dLN was determined. Significant reductions in the number of parasites were observed in the different evaluated organs of vaccinated mice, as compared with those that received only saline or saponin (Fig. 3). In this context, vaccinated mice with rLiHyp1 plus saponin showed significant reductions in the parasite load in liver (3.8- and 3.3-log reductions, Fig. 3A), spleen (3.7- and 3.5-log reductions, Fig. 3B), BM (3.0- and 3.0-log reductions, Fig. 3C), and dLN (3.9- and 3.6-log reductions, Fig. 3D), in comparison to the saline and saponin groups, respectively. Attempting to determine the influence of immunization with rLiHyp1 plus saponin on the L. infantum specific killing effectors functions in the spleen of infected mice, nitrite was assayed as an indicator of nitric oxide (NO) production in spleen cells. The nitrite production was significantly higher in mice vaccinated with rLiHyp1 plus saponin after stimulation with SLA, as compared to the control groups that produced minimum amounts of this product (data not shown).
The production of cytokines in the supernatants of spleen cells cultures stimulated with rLiHyp1 and SLA after challenge was analyzed to determine the immunological correlates of protection induced by rLiHyp1. The spleen cells from mice vaccinated with rLiHyp1 plus saponin produced higher levels of SLA-specific IFN-γ, IL-12 and GM-CSF cytokines than those secreted by spleen cells from control groups, 10 weeks after infection (Fig. 4A). In contrast, the SLA-driven production of IL-4 and IL-10 showed that vaccination with rLiHyp1 plus saponin induced no production of these cytokines in the vaccinated and infected animals.
The contribution of CD4 and CD8 T cells and the dependence of IL-12 production for the SLA-specific IFN-γ response from the spleen cells of mice immunized with rLiHyp1 plus saponin and challenged with L. infantum were evaluated. The IFN-γ production was completely suppressed using anti-IL-12 or anti-CD4 monoclonal antibodies in the spleen cells cultures (Fig. 4B). The addition of anti-CD8 antibodies to the cultures also decreased the production of IFN-γ, as compared to the cell cultures without treatment (1.881±139 pg/mL before, and 1.533±110 pg/mL after including anti-CD8 antibodies); however, the production of this cytokine proved to be higher than that produced by the use of the anti-CD4 monoclonal antibody.
As observed before challenge, the ratio between IFN-γ/IL-4 and IFN-γ/IL-10, and between IL-12/IL-4 and IL-12/IL-10 indicated that vaccinated mice developed a specific Th1 immune response, which was maintained after infection in these animals (Fig. 4C and 4D, respectively). In this study, very low levels of anti-SLA antibodies could be observed in the sera of all mice groups challenged with L. infantum. However, it was possible to detect that vaccinated and infected mice presented SLA-specific IgG2a antibodies that were significantly higher than the obtained IgG1 levels (Fig. 4E).
Different Leishmania proteins with antigenic properties were recently identified by an immunoproteomic approach applied to L. infantum promastigotes and amastigote-like proteic extracts [18], including hypothetical proteins of the parasites. The fact that antibodies present in the sera of infected dogs recognized these hypothetical proteins indicates that they are expressed by parasites during active infection, and are antigenic to the host's immune system. In this context, the DNA encoding one of these Leishmania hypothetical proteins, which was specifically recognized by antibodies in the amastigote-like antigenic extracts, was cloned and expressed in E. coli and tested for its antigenicity and prophylactic properties. Immunoblotting and ELISA analyses demonstrated that the recombinant LiHyp1 protein (rLiHyp1) was specifically recognized by antibodies present in the sera of dogs with symptomatic and asymptomatic VL, yet it presented no cross-reactivity with the sera of dogs vaccinated with a Brazilian recombinant vaccine, Leish-Tec®, or with animals experimentally infected with T. cruzi, demonstrating, besides antigenicity its potential for improvement of CVL serodiagnosis. Dogs are also reservoirs for T. cruzi parasites in endemic areas for VL transmission in Brazil, and seroprevalences of anti-T. cruzi antibodies, which may cross react with Leishmania antigens, of 21.9% and up to 57.0% have been reported [19], [20].
In a previous study, it was demonstrated that sera from dogs naturally infected with L. infantum displayed reactivity with Leishmania ribosomal proteins (LRP) through Western-Blot analysis. A comparison between LRP and SLA showed that LRP had a similar sensitivity in ELISA, but higher specificity than the SLA-based assays in the diagnosis of CVL [21]. However, the technical purification of LRP is complex and labor-intensive. In contrast, the production of recombinant proteins is less complex and allows obtain large amounts of proteins, when compared to production of semi-purified extracts of the parasites (like LRP). In this context, the use of the rLiHyp1 protein in the serodiagnosis of CVL is attractive.
Amastigote antigens have been far less tested as vaccine candidates against VL [15]. Therefore, a vaccine that is able to elicit immune responses against intracellular amastigotes of Leishmania may present advantages not only for prophylactic, but also for therapeutic vaccines. In this context, the immunization with rLiHyp1 plus saponin was able to induce a predominant Th1 immune response, which was characterized by an in vitro rLiHyp1-specific production of IFN-γ, IL-12 and GM-CSF, combined with the presence of very low levels of IL-4 and IL-10. After infection, mice immunized with rLiHyp1 plus saponin, when compared to control groups, displayed significant reductions of the number of parasites in all evaluated organs (liver, spleen, BM, and dLN), which correlated a specific rLiHyp1- and SLA-dependent IFN-γ production in the spleen, one of the main cytokines implicated in the acquired immunity against infection with Leishmania [22], [23], [24]. The CD4 T cells proved to be the major source of IFN-γ in the protected mice, since depletion of these cells in cultures of spleen cells stimulated with SLA significantly abrogated this response. Similarly, in the vaccinated mice, IFN-γ production proved to be IL-12-dependent. Although previous reports have shown that the activation of both CD4 and CD8 T cells subsets may be important for the killing of parasites in mice vaccinated with different parasite recombinant antigens [25], [26], the present study's data suggest that CD8 T cells may contribute in a less extension to the induction of IFN-γ mediated response elicited by the rLiHyp1 plus saponin formulation. Besides production of IFN-γ, these cells may contribute to infection control by their direct cytotoxic effect on infected cells, as previously demonstrated in other experimental conditions [24]. Altogether, the present study indicates that immunization with rLiHyp1 plus saponin primed BALB/c mice for an rLiHyp1-specific Th1 response that was sustained after L. infantum infection challenge.
The present study also showed that the protection in BALB/c mice against L. infantum is associated with a significant decrease in the production of macrophage deactivating cytokines, like IL-4 and IL-10. Very low levels of Leishmania-specific IL-10 were detected after the stimulation of spleen cells from vaccinated mice, 10 weeks after infection. In contrast, spleen cells from both control mice groups showed a significantly higher production of this cytokine. Indeed, control of the parasite-mediated IL-10 response in vaccinated mice may be critical for protection, since this cytokine is considered to be the most important factor for VL progression after infection with viscerotropic Leishmania species in IL-10 deficient mice [12],[27],[28], or in mice treated with an anti-IL-10 receptor antibody [29]. In BALB/c mice, the IL-4-dependent production of IgG1 antibodies is associated with disease progression due to some Leishmania species, including L. amazonensis [30], but it is not confirmed in L. infantum or L. donovani [31],[32]. Nonetheless, in BALB/c mice vaccinated with recombinant A2 protein or LRP plus saponin, the protection against cutaneous or visceral leishmaniasis have been also correlated with a decrease in Leishmania-specific IL-4 and IL-10 mediated response [12], [17], [22], [33].
Spleen cells from vaccinated mice, as compared to the control groups, produced higher levels of rLiHyp1- and SLA-specific GM-CSF, a cytokine related with macrophage activation and resistance in murine models against different intracellular pathogens, including L. major [34], L. donovani [35], and L. chagasi ( = L. infantum) [12]. It has also been shown that the immunization of humans with a crude Leishmania antigenic preparation using this cytokine as an adjuvant commonly induces a parasite-specific Th1 response [36], and that the administration of a therapeutic vaccine containing some Leishmania antigens plus GM-CSF could be correlated with the cure of lesions in the muco-cutaneous leishmaniasis [37].
A critical aspect for Leishmania vaccines development refers to the pre-clinical model chosen for initial screening of vaccine candidates. Although sand fly transmitted infection in hamsters more closely resemble the natural transmission and the human disease, this infection model requires specific laboratory conditions and trained personnel staff, which are not widely available, hindering its general use as a first step for testing vaccine efficacy against VL [38]. In contrast, BALB/c mice infected with L. donovani or L. infantum is one the most widely studied murine models for VL, and is therefore naturally selected over other models for this purpose [17], [39], [40], [41]. Murine models have also allowed the characterization of the immune mechanisms required to develop organ-specific immune response against Leishmania [41], [42]. Therefore, the evaluation of the parasite burden in different organs is an important marker of vaccine efficacy against VL in these models. In a recent study, it was demonstrated that the subcutaneous route of inoculation of L. infantum in BALB/c mice induces a faster infection development in the animals and higher parasite burden in different tissues as compared to the intravenous challenge [41]. In this context, the subcutaneous route was selected to evaluate the efficacy of rLiHyp1 plus saponin vaccine against L. infantum. In addition, in comparative studies, it was found that protection afforded by vaccination might be improved in animals challenged by intradermal/subcutaneous route as compared to those receiving an intravenous challenge [43], [44]. Nevertheless, additional studies may well be carried out in order to extend the observations present herein of the protective efficacy of rLiHyp1 plus saponin vaccination to other infection models and experimental conditions.
In conclusion, the present study's data indicated that a Leishmania amastigote-specific protein, member of the super-oxygenase family, LiHyp1, is antigenic in the CVL, and also conferred protection in BALB/c mice against L. infantum. Protection correlated with the CD4 T cells response characterized by high IFN-γ, IL-12, and GM-CSF, and low IL-4 and IL-10 levels. Therefore, the LiHyp1 protein constitutes a new and promising antigen candidate for serodiagnosis and vaccine development against VL.
|
10.1371/journal.pgen.1004544 | Chromatin Insulator Factors Involved in Long-Range DNA Interactions and Their Role in the Folding of the Drosophila Genome | Chromatin insulators are genetic elements implicated in the organization of chromatin and the regulation of transcription. In Drosophila, different insulator types were characterized by their locus-specific composition of insulator proteins and co-factors. Insulators mediate specific long-range DNA contacts required for the three dimensional organization of the interphase nucleus and for transcription regulation, but the mechanisms underlying the formation of these contacts is currently unknown. Here, we investigate the molecular associations between different components of insulator complexes (BEAF32, CP190 and Chromator) by biochemical and biophysical means, and develop a novel single-molecule assay to determine what factors are necessary and essential for the formation of long-range DNA interactions. We show that BEAF32 is able to bind DNA specifically and with high affinity, but not to bridge long-range interactions (LRI). In contrast, we show that CP190 and Chromator are able to mediate LRI between specifically-bound BEAF32 nucleoprotein complexes in vitro. This ability of CP190 and Chromator to establish LRI requires specific contacts between BEAF32 and their C-terminal domains, and dimerization through their N-terminal domains. In particular, the BTB/POZ domains of CP190 form a strict homodimer, and its C-terminal domain interacts with several insulator binding proteins. We propose a general model for insulator function in which BEAF32/dCTCF/Su(HW) provide DNA specificity (first layer proteins) whereas CP190/Chromator are responsible for the physical interactions required for long-range contacts (second layer). This network of organized, multi-layer interactions could explain the different activities of insulators as chromatin barriers, enhancer blockers, and transcriptional regulators, and suggest a general mechanism for how insulators may shape the organization of higher-order chromatin during cell division.
| Chromatin insulators mediate specific long-range DNA interactions required for the three dimensional organization of the interphase nucleus and for transcription regulation, but the mechanisms underlying the formation of these interactions is currently unknown. In this manuscript, we investigate the molecular associations between different protein components of insulators (BEAF32, CP190 and Chromator) by biochemical and biophysical means, and develop a novel biophysical assay to determine what factors are necessary and essential for the formation of long-range DNA interactions (LRI). Importantly, we show that CP190 and Chromator are able to mediate LRIs between specifically-bound BEAF32 nucleoprotein complexes. This ability of CP190 and Chromator to establish LRI requires specific contacts between BEAF32 and their C-terminal domains, and dimerization through their N-terminal domains. In particular, the BTB/POZ domains of CP190 form a strict homodimer. We propose a general model for insulator function in which BEAF32/dCTCF/Su(HW) provide DNA specificity, whereas CP190/Chromator are responsible for the physical interactions required for long-range contacts. This network of organized, multi-layer interactions could explain the different activities of insulators, and suggest a general mechanism for how insulators may shape the organization of higher-order chromatin during cell division.
| The physical organization of eukaryotic chromosomes is key for a large number of cellular processes, including DNA replication, repair and transcription [1]–[6]. Chromatin insulators are genetic elements implicated in the organization of chromatin and the regulation of transcription by two independent modes of action: ‘enhancer blocking’ insulators (EB insulators) interfere with communications between regulatory elements and promoters, whereas ‘barrier’ insulators prevent the spread of silenced chromatin states into neighboring regions [7]–[9]. Recently, insulator elements have been implicated in chromosome architecture and transcription regulation through their predicted binding to thousands of sites genome-wide. For instance, insulators were shown to regulate transcription of distinct gene ontologies, to separate distinct epigenetic chromatin states, and to recruit H3K27me3 domains to Polycomb bodies [10]–[13].
In Drosophila, five insulator families have been identified, that differ by their DNA-binding protein (insulator binding protein, or IBP): Suppressor of Hairy-wing [Su(Hw)] [14], boundary element-associated factor (BEAF32) [15], Zeste-white 5 (Zw5) [16], the GAGA factor (GAF) [17], and dCTCF [18], a distant sequence homologue of mammalian CTCF. Two BEAF32 isoforms exist (BEAF32A and BEAF32B). In this paper, we will only consider BEAF32B (which will be referred to as BEAF32) as: (i) BEAF32B represents more than 95% of the binding peaks detected by chip-seq in cell lines [11], (ii) BEAF32A binding does not play a role in the insulating function of BEAF [19], and (iii) BEAF32A expression is not essential for the development of embryos in adult flies [20]. IBPs are often necessary but not sufficient to ensure insulation activity at a specific locus, and several insulator co-factors have been shown to be additionally required. Particularly, Centrosomal Protein 190 (CP190) [21], a protein originally described for its ability to bind to the centrosome during mitosis [22], was shown to play a crucial role in the insulation function of various IBPs [10], [23], [24].
Insulator proteins often associate in clusters of overlapping binding sites more often than would be expected by chance, suggesting that these factors often bind as a complex to the same genetic locus. For instance, BEAF32, dCTCF and CP190 binding sites most often cluster with at least another factor (∼70, ∼77 and >90%, respectively) [25]. In addition, insulators show a large compositional complexity, as demonstrated by the frequencies of binding of different combinations of insulator associated proteins: CP190 associates with its most common partner BEAF32 (∼50%), but also to a lesser extent to dCTCF and Su(HW) (25 and 20%, respectively), while BEAF32, dCTCF, and CP190 cluster together in >15% of CP190 binding sites [10], [25], [26]. This compositional complexity may be key to understanding the locus-specific functions of insulators.
A critical feature of Drosophila and vertebrate insulators is their ability to form specific long-range DNA interactions (hereafter LRIs) [27]–[34]. Three-dimensional loops have been implicated in all levels of chromatin organization ranging from kb-size loops to larger intra-chromosomal loops hundreds of kb in size [6], [35], [36]. To date, it is unclear what factors provide the physical interactions required for the formation and regulation of LRIs. In addition to binding the specific insulator sequences, IBPs have been proposed to be sufficient to bridge two distant DNA molecules [10], [37]. However, other factors such as CP190, Mod(mdg4), or cohesin have been implicated in the formation of LRIs [10], [38]–[40].
The observation that most CP190 binding sites co-localize with insulator binding proteins (>90%) [10], [25] prompted the hypothesis that CP190 is a common regulator of different insulator classes [10], [40]. CP190 is composed of a BTB (bric-a-brac, tramrack, and broad complex)/POZ (poxvirus and zinc-finger) domain, four predicted C2H2 zinc-finger motifs, and an E-rich, C-terminal region. Importantly, CP190 has been recently shown to preferentially mark chromatin domain barriers [13]. These barriers are also heavily bound by other insulator proteins, such as BEAF32, dCTCF and to a lesser extent Su(HW), and have been shown to often form LRIs [12]. Overall, these data suggest a role for CP190 in participating in the three dimensional folding of the genome by the formation of long-range interactions.
Surprisingly, a second factor, called Chromator, was also shown to be overrepresented at physical domain barriers [13]. During mitosis, Chromator forms a molecular spindle matrix with other nuclear-derived proteins (Skeletor and Megator) [41]. In contrast, during interphase Chromator localizes to inter-band regions of polytene chromosomes [42], [43] and plays a role in their structural regulation as well as in transcriptional regulation [44]. Chromator can be divided into two main domains, a C-terminal domain containing a nuclear localization signal, and an N-terminal domain containing a chromo-domain (ChD) required for proper localization to chromatin during interphase [45].
Here, we investigate the molecular associations between different components of insulator complexes (BEAF32, CP190 and Chromator) by biochemical and biophysical means. We developed a unique assay to determine what factors are necessary and essential for the formation of long-range DNA interactions, and show that BEAF32 is necessary but not sufficient to bridge long-range interactions. In contrast, addition of CP190 or Chromator is sufficient to mediate LRI between specifically-bound BEAF32 nucleoprotein complexes. This ability of CP190 and Chromator to establish LRI requires specific contacts between BEAF32 and their C-terminal domains, and dimerization through their N-terminal domains. In particular, the BTB/POZ domains of CP190 form a strict homodimer, and its C-terminal domain interacts with several IBPs. We propose a general model for insulator function in which BEAF32/dCTCF/Su(HW) provide DNA specificity (first layer proteins) whereas CP190/Chromator are responsible for the physical interactions required for long-range contacts (second layer). The multiplicity of interactions between insulator binding and associated proteins could thus explain the different activities of insulators as chromatin barriers, enhancer blockers, and transcriptional regulators.
BEAF32 co-localizes genome-wide with CP190 and Chromator, but the molecular mechanisms underlying this co-localization are unknown. To investigate whether this observed co-localization was due to direct protein-protein interactions, we heterologously expressed and purified BEAF32, CP190, Chromator and several protein subdomains. BEAF32 was expressed as a MBP (Maltose-Binding Protein) fusion protein (Figure 1A–B), since wild-type BEAF32 was mainly insoluble. CP190, Chromator, their C-terminal domains (CP190-C and Chromator-C, respectively), and CP190-BTB/POZ were heterologously expressed as His-tagged fusions (Figure 1A–B, and Materials and Methods). After purification, proteins were >95% pure and were specifically recognized by the corresponding antibodies (Figure 1B, and Materials and Methods).
A typical example of co-localization of these factors can be found at the Tudor-SN locus, a genomic region that shows a strong localization pattern for BEAF32, CP190, and Chromator but not for dCTCF or Su(HW) (Figure 1C), and contains six specific binding sites for BEAF32 (CGATA motifs) [19]. To directly test whether BEAF32 was able to specifically bind to this genomic site, we PCR-amplified a 447 bp DNA fragment from Tudor-SN that contained six CGATA motifs (hereafter DNAtudor, Figure 1C). First, we used an electric mobility shift assay (EMSA) in which a plasmid containing the DNAtudor insertion was restricted and used as a substrate (Figure 2A). The restriction reaction produced three different DNA fragments of 750, 1627 and 4025 bp, the second of which contained the 447-bp DNAtudor insertion, and was the only DNA fragment harboring specific CGATA motifs. The specific binding of factors to these different DNA fragments was assessed by quantifying the disappearance of unbound DNA species, as bound species often produced smeared bands due to rapid association/dissociation of proteins from DNA at low affinities and due to the low resolution of the gel matrix. The binding of BEAF to DNA was specific, as only the DNAtudor-containing band was preferentially shifted by addition of BEAF32 (Figure 2A).
Secondly, to quantify the affinity and specificity of DNA binding by BEAF32, we implemented a fluorescence anisotropy-based assay that directly measures the binding of proteins to DNA. The binding of proteins, such as BEAF32, to short fluorescently-labeled DNA fragments decreases the rotational diffusion of the DNA molecule and increases the fluorescence anisotropy of the attached fluorophore (Figure 2B) [46]. BEAF32 binds with a moderate apparent affinity to non-specific DNA (58 bp DNA fragment with no CGATA motif, hereafter DNANS), and the binding isotherm can be well described by a simple single-site model (Eq.1, Text S1, KD = 165±30 nM, Figure 2C). In contrast, BEAF32 binds to a specific DNA fragment of the same length (58 bp; DNA fragment containing three CGATA motifs from Tudor-SN, hereafter DNAS) with a higher affinity and displaying a degree of cooperativity (Figure 2C). The binding isotherm cannot be fitted by a single-site model, thus we turned to a Hill model (Eq. 2, Text S1) with a resulting apparent affinity of KD = 68±5 nM and a Hill coefficient of n = 3±0.4. In addition, the change in fluorescence anisotropy signal was larger for DNAS (32±2 anisotropy units) than for DNANS (12±5 anisotropy units), indicating that BEAF32-DNAS makes a larger complex. Overall, these results indicate a cooperative binding of BEAF32 to CGATA motifs, suggesting oligomerization of BEAF32 at genomic sites containing multiple CGATA motifs. These results were consistent with competitive inhibition experiments (Supplementary Figure S1A). Equivalent fits of the DNANS binding isotherm to a Hill model produced a Hill coefficient of 0.9±0.3, consistent with cooperative binding of BEAF32 to DNAS being due to the presence of CGATA motifs.
CP190 contains a BTB/POZ domain and is predicted to possess four classical C2H2 zinc-finger motifs that could be involved in direct DNA binding. It is unclear whether CP190 can directly associate to DNA, or rather relies on its binding to other factors to target specific binding sites [21], [47]. To address this question, we investigated the ability of CP190 to bind to Tudor-SN. This locus displays CP190 binding by Chip-chip [25], [48] (Figure 1C and Supplementary Table S6) and may thus contain moderate affinity sites for CP190. By EMSA, we observed that CP190 associated equally well to all DNA fragments, with no specificity shown for the DNAtudor-containing fragment (Figure 2D). Next, we tested the binding specificity of CP190 by fluorescence anisotropy, using two different dsDNA fragments (DNAS and DNANS). DNAS should contain the potential CP190 moderate affinity sites giving rise to the in vivo binding of CP190 to Tudor-SN, while DNANS is a DNA fragment of the same length but with a random sequence serving as a control for specificity. In agreement with EMSA, fluorescence anisotropy experiments showed moderate DNA binding affinity but no specificity (KD = 109±5 nM, n = 2±0.3 for CP190 on both DNAS and DNANS, Figure 2E). These results are supported by competition experiments (Supplementary Figure S1B), and are in agreement with similar experiments showing that CP190 fails to show any specificity when using a dsDNA fragment containing the predicted binding sequence of CP190 [25] (Supplementary Figure S6). Overall, these results are consistent with the specificity of in vivo binding of CP190 to Tudor-SN being mediated by other factors.
Next, we tested whether the C-terminal domain of CP190 was involved in the ability of CP190 to bind DNA non-specifically by determining the DNA binding properties of CP190-C, a protein construct that contains neither BTB/POZ nor the zinc-finger motif (Figure 1A). CP190-C was not able to bind DNAS (Figure 2H), consistent with the non-specific association of CP190 to DNA being mediated by the N-terminal domain of CP190. Binding competition experiments of pre-bound BEAF32-DNAS are inconsistent with CP190-BTB/POZ being involved in DNA binding (Supplementary Figure S1D), but further experiments will be required to determine the contribution of the different domains in the N-terminus of CP190 to DNA association. In addition, we cannot exclude the possibility that other factors or post-translational modifications may partially affect the mechanism of DNA binding by CP190. However, the ubiquitous co-localization of CP190 with factors displaying specific DNA-binding activities (BEAF32, dCTCF, Su(HW)) (>90%) [25] suggests that the presence of CP190 at specific loci is mediated in most cases by other proteins. From these experiments, we cannot exclude the possibility that CP190 may bind specifically to other genomic sites.
The ability of Chromator to associate to DNA has not been described so far, although its association to chromatin has been suggested to require its Chd-containing N-terminal domain (Yao et al, 2012). Despite the presence of high affinity in vivo sites for Chromator in Tudor-SN, our EMSA and fluorescence anisotropy experiments showed that Chromator binds DNA non-specifically (Figure 2F–G) and with a lower affinity than BEAF32 or CP190 (KD = 360±30 nM and n = 2±0.2, see Figure 2G and Supplementary Figure S1C). Chromator-C did not present any DNA binding activity (Figure 2H), suggesting that Chromator binding to DNA requires its N-terminal domain or uncharacterized post-transcriptional modifications.
Next, we investigated whether BEAF32 directly interacts with CP190 and Chromator by using several complementary approaches. First, we employed co-immunoprecipitation (co-IP) to detect protein-protein interactions with heterologously purified proteins. A guinea pig-anti-Chromator-antibody was covalently linked to a column and a mix of purified BEAF32 and Chromator were incubated in the column for 60 min, eluted and analyzed by western blotting (Figure 3A, see full bands of all co-IPs in Supplementary Fig. S2C–I). Western-Blot analysis of the elution clearly showed the specific interaction between BEAF32 and Chromator (Figure 3A, middle column), whereas neither BEAF32 nor Chromator were found to bind to an IgG-antibody column (Figure 3A, right column). Importantly, BEAF32 did not bind to an anti-Chromator column in the absence of Chromator (Supplementary Figure S2B). To investigate what domain of Chromator is involved in interactions with BEAF32, we performed co-IP experiments in which a mix of BEAF32 and Chromator-C were incubated in a column covalently bound by antiChromator antibody, and the elution analyzed by western blotting. Interestingly, BEAF32 is specifically retained in the Chromator-C column, consistent with BEAF32/Chromator interactions being mediated by the C-terminal domain of Chromator (Figure 3B). Additionally, Chromator is retained in a CP190 column, an interaction that seems to be specifically mediated by CP190-BTB/POZ (Supplementary Figure S2K).
Similar co-IP experiments were performed to test putative BEAF32-CP190 interactions. A rabbit-anti-CP190-antibody was covalently linked to a resin and incubated with a purified mix of BEAF32 and CP190 or CP190-C. BEAF32 binds efficiently to both CP190 and CP190-C (Figure 3C–D), but failed to interact with CP190-BTB/POZ (Supplementary Figure S7). BEAF32 is not recognized by CP190 antibodies and was not retained by an anti-CP190 column (Supplementary Figures S2A–B). These results indicate that BEAF32/CP190 interactions are mediated by the C-terminal domain of CP190, although we cannot discard an additional contribution of the zinc-finger domains of CP190 to this interaction. Interestingly, both BEAF32 and CP190 were retained in an anti-Chromator antibody column (Figure 3E), consistent with binary interactions between BEAF32 and CP190/Chromator and with interactions between CP190 and Chromator.
To test whether these interactions are physiologically relevant, we performed Co-IP experiments using S2 nuclear extracts (see Materials and Methods). Interactions between BEAF32, Chromator and CP190 were clearly detected while either using anti-Chromator or anti-CP190 (Figures 3F–G, respectively) antibodies. Overall, these results suggest that BEAF32, Chromator and CP190 are part of the same molecular complex. However, further work is necessary to determine the architecture and stoichiometry of this complex.
Next, we investigated whether interactions among BEAF32, CP190 and Chromator lead to the formation of higher-order DNA interactions. First, we used EMSA to test whether BEAF32-CP190/Chromator sub-complexes bind to the 447 bp DNAtudor fragment (Figure 1C). BEAF32 binding to DNAtudor (Figure 4A, Lane 1, band I) produced a discrete shift corresponding to a BEAF32/DNAtudor complex (Figure 4A, lane 2, band II). Consistent with previous results, neither CP190/Chromator (as the concentrations used here were lower than the KD), nor their C-terminal fragments were able to bind DNAtudor under these conditions (Figure 4A, band I, lanes 3, 5, 7 and 9, respectively). Interestingly, a second band with lower electrophoretic mobility appeared only when BEAF32 and either CP190 or Chromator were simultaneously present (Figure 4A, lanes 4 and 6, band III). Furthermore, this complex did not form when CP190 was replaced by CP190-C (Figure 4A, compare lanes 4 and 8), and exhibited a similar intensity when Chromator-C was used instead of full-length Chromator (lane 10 in Figure 4A). Control experiments where BEAF32 was replaced by MBP showed that the formation of the protein-DNA complexes leading to bands II and III required the presence of BEAF32 (Figure 4B). Band II thus corresponds to a complex formed by BEAF32 and DNAtudor, while band III indicates the presence of specific interactions between DNA-bound BEAF32 and CP190, Chromator, and Chromator-C.
To characterize the different complexes formed by BEAF32, CP190 and Chromator, we turned to fluorescence correlation spectroscopy (FCS). FCS uses the fluctuations in the number of freely diffusing fluorescently-labeled molecules within a confocal volume to characterize their diffusion time [49], [50] (Figure 5A). Thus, the formation of protein-DNA complexes can be monitored by the increase in the apparent size (related to the diffusion time) of a fluorescently-labeled dsDNA fragment upon protein binding. In our case, we used the 58 bp DNAS fragment (harboring three CGATA motifs, 5′-Cy3B labeled, see Materials and Methods) as a fluorescent reporter to quantitatively monitor the formation of BEAF32/CP190/Chromator complexes (Figure 5D). Identical results were obtained with an atto655-DNAS probe (Supplementary Figure S3). Incubation of DNAS with saturating concentrations of BEAF32 (400 nM) led to an increase in its apparent diffusion time from 0.53±0.03 to 0.85±0.04 ms, obtained by fitting our measurements to a 3-D diffusion model with a triplet state (Eq. 3, Text S1). This shift is consistent with the binding of BEAF32 to DNAS leading to the production of a molecular complex (hereafter B32S complex) with an increased apparent size (Figure 5B). The addition of low concentrations of CP190 (50 nM) to DNAS produced a small increase in the diffusion time (from 0.53±0.03 to 0.65±0.09 ms), consistent with the sub-affinity concentrations used. In contrast, CP190-C did not change the diffusion time of DNAS (Figure 5B), in agreement with our previous results showing no DNA-binding activity for this domain of CP190 (Figure 2H).
Next, we investigated whether CP190 binds to B32S complexes. We observed that the incubation of pre-formed B32S complexes with low-concentrations of CP190 (50 nM) led to a considerable increase in the size of complexes (Figure 5C). This low CP190 concentration (below its affinity) was used to enhance the specificity of CP190/BEAF32 interactions and limit the direct binding of CP190 to DNAS. Conversely, the addition of CP190-C to B32S slightly decreased the apparent size of the complex (Figure 5C). To ensure that this small decrease in diffusion time was not due to the dissociation of BEAF32 from DNAS, we performed fluorescence anisotropy experiments. The anisotropy of pre-formed B32S complexes was independent of the concentration of CP190-C, but decreased to the anisotropy of free DNAS upon addition of high salt concentrations (Supplementary Figure S4). These results indicate that the decrease in diffusion time observed in B32S/CP190-C complexes is not due to the dissociation of BEAF32 from DNAS, but to the change in the shape of the complex upon CP190-C binding. Overall, these results are consistent with either CP190 binding a B32S complex or triggering long-range inter-segment interactions between two B32S complexes.
To discriminate between these two models, we turned to fluorescence cross-correlation spectroscopy (FCCS). FCCS measures the correlated fluorescence intensity fluctuations of two spectrally-distinct, fluorescently-labeled molecules to quantitatively determine whether they are in the same molecular complex [50], [51]. When two DNA fragments labeled with different colors are part of the same molecular complex, their fluorescence fluctuations will be correlated (LRI), whereas no cross-correlation will be observed if the diffusion of the two DNA fragments is independent (no LRI, Figure 5D). We used a 50/50 mixture of DNAS labeled with Cy3B and atto655. Since these two fluorophores can display a significant level of crosstalk between detection channels, introducing apparent cross-correlation in the absence of interaction, we used pulsed interleaved excitation (PIE-FCCS) [52], [53] a technique that eliminates this artifactual effect and allows quantitative fluorescence cross-correlation measurements. The cross-correlation signals were measured for DNAS, B32S, and solutions of pre-formed B32S complex incubated with either CP190 or CP190-C and fitted with Eq. 4 (Text S1). Neither DNAS, nor B32S showed cross-correlation (Figure 5E), demonstrating the inability of BEAF32 alone to mediate long-range intermolecular interactions between CGATA motifs. In agreement with our previous observations (lack of band III in lane 8, Figure 4A), addition of CP190-C to B32S did not trigger the formation of intermolecular complexes, suggesting that the E-rich domain of CP190 is not sufficient to generate LRIs in vitro. In contrast, these complexes were formed in the presence of full-length CP190, demonstrated by the appearance of a clear cross-correlation signal (18±6%, Figure 5E). From the DNA labeling efficiencies of Cy3B- and atto655-labeled oligonucleotides (∼57 and 97%, respectively), and the fact that a maximum of 50% of the bridged DNA can be observed in the cross-correlation amplitude (since Cy3B-Cy3B or atto655-atto655 complexes do not produce a cross-correlation signal), we can conclude that 65±22% of the B32S-atto655 complexes take part in LRIs mediated by CP190. Importantly, under these conditions CP190 alone was not able to generate LRIs (Figure 5E), and addition of neither full-length CP190 nor CP190-C affected the specific binding of BEAF32 to DNAS (Supplementary Figure S4).
Thus, while CP190-C interacts with BEAF32, the N-terminal domain of CP190 appears necessary for the formation of inter-segment LRIs mediated by BEAF32-bound DNA in vitro (as CP190-C is not sufficient to mediate these interactions). In agreement with this model, the competition of pre-formed B32S-CP190-B32S complexes with the purified, isolated CP190-BTB/POZ domain (Figure 1A) led to the disappearance of cross-correlation signal (Figure 5F), but not to the displacement of BEAF32 from DNAS (Supplementary Figure S1D).
Overall, the FCCS data strongly suggest that the BTB/POZ domain of CP190 is involved in the direct protein-protein interactions required for the establishment of long-range contacts. To directly test this hypothesis, we solved the crystal structure of the CP190-BTB/POZ. BTB/POZ motifs are widespread in eukaryotes (350 BTB/POZ-containing proteins in the human genome). Despite a low degree of primary sequence conservation (as low as 10%), the various structures reported in the literature are very similar (root mean square deviation, or RMSD ∼1–2 Å) with the overall architecture being composed of a cluster of five alpha helices capped on one end by three beta sheets. BTB/POZ motifs have been found to homodimerize, heterodimerize, and in rare cases to promote tetramerization. These different types of oligomerization states depend primarily on the surface residues involved in oligomerization and have been well documented elsewhere [54], [55].
The CP190-BTB/POZ domain crystallized as a stable and symmetric homodimer, in agreement with gel-filtration analysis. The overall structure is similar to classic BTB/POZ-ZF transcriptional factors where the N-terminal BTB/POZ domain is followed by several Zinc-Fingers domains (Figure 6A, Materials and Methods and Supplementary Table S1). The dimerization interface is stabilized by a swapped β-strand that forms a long groove where extended polypeptidic segments can bind in order to recruit other protein partners. The dimer interface (1902 Å2/monomer according to PISA 1.47 [56]) is composed of numerous hydrophobic interactions mainly from alpha helices α1 and α2 (i.e. W12, F15, F16, F23, L47…). The native homodimeric organization is also reinforced by the N-terminal strand (residues Glu2 to Asp10) being swapped: the β1 stand of a monomer interacts with the β5 strand of the other monomer. The sequence conservation among CP190 orthologs from insects (55 sequences analyzed using CONSURF [57]) show little conservation besides the domain core, the dimerization interface and the peptide binding groove (Figure 6A). Interestingly, this suggests that CP190-BTB/POZ does not form higher order macromolecular assemblies by itself while partner recruitment requires homo-dimerization. Importantly, we found that CP190-BTB/POZ forms strict homo-dimers (Figure 6A), consistent with the ability of CP190 to form LRIs.
Finally, we used FCS and PIE-FCCS to test whether Chromator was able to mediate LRIs between two B32S complexes in vitro. The addition of 100 nM Chromator to DNAS generated a small but noticeable change in the diffusion time (0.53±0.03 to 0.59±0.03 ms) (Figure 5G), consistent with our previous results (Figures 2F and 2G) indicating that Chromator interacts non-specifically with DNA. In contrast, incubation of DNAS with Chromator-C did not induce any change in the diffusion time of the probe (Figure 5G), in agreement with anisotropy experiments (Figure 2H). Interestingly, addition of Chromator (but not of Chromator-C) to B32S considerably changed the diffusion time of the complex (Figure 5H), suggesting an interaction between Chromator and the B32S complex. Similarly to the results obtained for CP190, addition of Chromator to pre-formed B32S complexes led to a cross-correlation amplitude of 13±4%, corresponding to a total of 47±14% of the B32S-atto655 complexes bridged by Chromator interactions (Figure 5I). The formation of these complexes was not observed when Chromator-C was added to B32S, nor when Chromator was added to DNAS in the absence of BEAF32 (Figure 5I). Overall, these results are consistent with interactions between the N-terminal domains of Chromator being required for the bridging function of Chromator, with Chromator-C providing the main direct interactions to BEAF32. We cannot discard, however, the possibility that Chromator-N may also partially interact with BEAF32.
Chromatin insulators promote higher-order nuclear organization through the establishment and maintenance of distinct transcriptional domains. Notably, this activity requires the formation of barriers between chromatin domains and the establishment of specific LRIs. In this paper, we investigated the molecular mechanism by which insulator proteins bind DNA, interact with each other and form long-range contacts.
Recently, genome-wide approaches have been used to investigate the roles of different insulator types in genome organization. Insulators enriched in both BEAF32 and CP190 are implicated in the segregation of differentially expressed genes and in delimiting the boundaries of silenced chromatin [25]. Notably, BEAF32 and CP190 are often found to bind jointly to the same genetic locus (>50% of CP190 binding sites contain BEAF32) [10], [25]. However, the molecular origin of this genome-wide co-localization was unknown as there was no direct proof of interaction between these proteins. Here, we showed, for the first time to our knowledge, that BEAF32 is able to interact specifically with CP190 in vitro and in vivo. In particular, we observed that this interaction is mediated by the C-terminal domain of CP190, with no implication of the C2H2 zinc-finger or the BTB/POZ domains, consistent with previous studies showing that the N-terminus of CP190 was not essential for its association with BEAF32 in vivo [58]. BEAF32 interacts specifically and cooperatively with DNA fragments containing CGATA motifs, consistent with previous observations [19]. In contrast, the binding of CP190 to DNA showed lower affinity and no specificity and required its N-terminal domain (containing four C2H2 zinc-fingers). Overall, these data suggest that one pathway for CP190 recruitment to DNA genome-wide requires specific interactions of its C-terminal domain with BEAF32. Other factors, such as GAF [38], are likely also involved in the recruitment of CP190 to chromatin, explaining why RNAi depletion of BEAF32 does not lead to the dissociation of CP190 from an insulator binding class containing high quantities of BEAF32 and CP190 [47]. We cannot discard that post-translational modifications in CP190 may also allow it to bind DNA directly and specifically, providing a second pathway for locus-specific localization.
In addition to acting as chromatin barriers, insulators have been typically characterized for their ability to block interactions between enhancers and promoters through the formation of long-range contacts [7], [16], [27], [28], [59]–[62]. Here, we developed a fluorescence cross correlation-based assay that allowed us, for the first time to our knowledge, to investigate the ability of BEAF32, CP190 and their complex to bridge specific DNA fragments, mimicking LRIs. We show that specific LRI can be stably formed between two DNA fragments containing BEAF32 binding sites, solely in the presence of both BEAF32 and CP190. Interestingly, LRI are displaced by competition in trans with the BTB/POZ domain of CP190, and LRIs are not observed in the presence of BEAF32 and CP190-C. Thus, both protein domains are required for the bridging activity of CP190. These data strongly suggest that the C-terminal domain is responsible for BEAF32-specific contacts whereas the N-terminal domain of CP190 is involved in the formation of LRI through CP190/CP190 contacts (Figure 6B). The role of the N-terminal domain of CP190 in protein-protein interactions is consistent with previous studies showing that N-terminal fragments of CP190 containing the BTB/POZ domains co-localize with full-length CP190 in polytene chromosomes [58].
BTB/POZ are a family of protein-protein interaction motifs conserved from Drosophila to mammals, and present in a variety of transcriptional regulators. BTB/POZ are found primarily at the N-terminus of proteins containing C2H2 zinc-finger motifs [63]–[65], and can be monomeric, dimeric, or multimeric [54], [55]. In fact, a recent study proposed that isolated CP190-BTB/POZ domains can exist as dimers or tetramers in solution [66]. The oligomerization behavior of CP190-BTB/POZ could have important implications for the role and mechanism by which CP190 bridges LRIs. Here, we showed that the BTB/POZ domains of CP190 forms homo-dimers with a large, conserved interaction surface (Figure 6A), consistent with these domains being responsible for the formation of the direct protein-protein interactions required for the establishment of long-range contacts. Interestingly, the oligomerization of CP190-BTB/POZ into homo-dimers implies a binary interaction between two distant DNA sequences, imposing important constraints for the mechanisms of DNA bridging by CP190.
In addition to interacting with BEAF32, CP190 is able to directly interact with other insulator binding proteins, such as dCTCF, Su(HW), and Mod(Mdg4) [21], [39], [66]–[68], or with the RNA interference machinery [69]. These interactions are usually mediated by the C-terminal domain of CP190, but a role for the C2H2 zinc-finger or the BTB/POZ domains in providing specific protein-protein contacts cannot be discarded [70]. In fact, an interesting feature of several homo-dimeric BTB/POZ domains is their ability to recruit a multitude of protein partners using a single protein-protein binding interface. For instance, several transcriptional co-repressors (BCOR, SMRT and NCor) are able to bind with micromolar affinity (2∶2 stoichiometry) to the BTB/POZ domain of BCL6, despite their low sequence homology [71], [72]. In this case, the mechanism of binding involves the formation of a third strand by the N-terminus of co-repressors folding onto the two strands exchanged by the BCL6-BTB/POZ monomers on their interface, with the rest of the minimal domain of interaction (10 residues) winding up along the lateral groove of the BCL6-BTB/POZ dimer (peptide binding groove in Figure 6A). In the case of CP190, the sequence and structural features of the conserved peptide binding groove within insect CP190-BTB/POZ domains suggest that the dimer interface of CP190 may act as a protein-protein interaction platform. Thus, the ability of BTB/POZ domains to form dimers and the promiscuous binding of CP190 to different insulator binding proteins (Su(HW), dCTCF [39], [67], and BEAF32) suggest not only that insulators share protein components [73], but also that CP190 may bridge long-range contacts involving distinct factors at each end of the DNA loop (Figure 6B). This model is consistent with previous proposals [73], and with the requirement of both C- and N-terminal domains of CP190 for fly viability [58]. Importantly, it provides a rationale for CP190 being a common factor between insulator binding proteins.
CP190 frequently binds with additional insulator binding proteins (∼85%), with BEAF32 and dCTCF being the most common partners (∼50% and ∼25%, respectively), and Su(Hw) amongst the least frequent partner (∼20%) [10], [25]. Importantly, BEAF32 does not show clustering with either dCTCF or Su(HW) in the absence of CP190 (<0.5% or ∼0.1%, respectively) [10], suggesting that the clustering of two insulator binding proteins requires CP190. The ability of CP190 to mediate LRIs between sites harboring different insulator binding proteins raises important questions: Are these LRIs specific? How is this specificity regulated? Are other factors or post-translational modifications involved in this selectivity? Future research will be needed to address these important questions.
Chromator localizes to inter-band regions of polytene chromosomes [42], [43] and binds to the barriers of physical domains genome-wide [13], however the mechanism leading to these localization patterns has been lacking. Previous studies showed that BEAF32 and Chromator co-localize at some genomic sites, and suggested that these proteins may participate in the formation of a single complex [@Gan:2011hy]. Here, we showed for the first time that BEAF32 directly and specifically interacts with Chromator in vivo and in vitro. This interaction is mediated by the C-terminal domain of Chromator, thus the ChD domain does not seem to be directly involved in interactions with BEAF32. Our results show that Chromator possesses a reduced affinity for DNA and binds with no sequence specificity to loci displaying strong Chromator binding peaks at the site tested (Tudor-SN locus, Figures 2G and 1C). Thus, we suggest that specific interactions between BEAF32 and Chromator may be responsible for its recruitment to polytene inter-band regions and domain barriers. Significantly, most BEAF32 binding sites genome-wide (>90%, Figure 6C and Supplementary Figure S5A) contain Chromator, suggesting an almost ubiquitous interaction between the two factors.
Interestingly, Chromator also co-localizes with the JIL-1 kinase at polytene inter-band regions and the two proteins directly interact by their C-terminal domains [41]. JIL-1 is an ubiquitous tandem kinase essential for Drosophila development and key in defining de-condensed domains of larval polytene chromosomes. Importantly, JIL-1 participates in a complex histone modification network that characterizes active, de-condensed chromatin, and is thought to reinforce the status of active chromatin through the phosphorylation of histone H3 at serine 10 (H3S10) [74]–[76]. Thus, BEAF32 could be responsible for the recruitment of the Chromator/JIL-1 complex to active chromatin domains to prevent heterochromatin spreading (Figure 6D) [@Gan:2011hy]. This mechanism would be consistent with the observation that BEAF32 localizes primarily to de-condensed chromatin regions in polytene chromosomes [15], is implicated in the regulation of active genes [10], [11], [25], [77] and delimits the boundaries of chromatin silencing [25].
CP190 is a common partner of BEAF32, dCTCF, and Su(HW), and has been thus proposed to play a role in the formation of long-range interactions at these insulators [10], [67]. On the other hand, both CP190 and Chromator have been recently shown to be massively overrepresented at barriers between transcriptional domains [12], [13]. In this paper, we show, for the first time, that only when CP190 or Chromator are present can long-range interactions between BEAF32-bound DNA molecules be generated. We provide strong evidence that the formation of in vitro LRI requires three ingredients: (1) binding of BEAF32 to its specific DNA binding sites; (2) specific interactions between the C-terminal domains of CP190/Chromator and BEAF32; and (3) homo- interactions between CP190/Chromator molecules mediated by their N-terminal ends.
To further investigate the roles of CP190 and Chromator in the formation of LRIs, we aggregated together statistically relevant contacts containing specific combinations of insulator factors from Hi-C data from embryos [13] (Figure 6E, and Materials and Methods). This analysis shows a relatively high correlation between the presence of BEAF32 and both CP190 and Chromator in sites displaying a high proportion of interacting bins between distant BEAF32 sites (Figure 6E), as compared with neighboring sites (16.9% of interacting bins for Chromator and CP190 sites; Wilcoxon test: p-value ∼1e-7). Thus, CP190 and Chromator may play a role at a subset of genetic loci by mediating and/or stabilizing interactions between BEAF32 and a distant locus bound by BEAF32 or a different insulator binding protein. Interestingly, the binding of BEAF32 to CGATA sites as multimers, and the existence of CP190-Chromator interactions suggest that long-range interactions at a single locus could involve hybrid/mixed complexes comprising at least these three factors.
These observations suggest a general model for insulator function in which BEAF32/dCTCF/Su(HW) provide DNA specificity (first layer proteins) whereas CP190/Chromator are responsible for the physical interactions required for long-range contacts (second layer). Direct or indirect interactions of first layer insulator proteins with additional factors (e.g. JIL-1, NELF, mediator) are very likely involved in directing alternative activities (e.g. histone modifications, regulation of RNAPII pausing) to specific chromatin loci. This model provides a rationale for the compositional complexity of insulator sequences [25] and for the multiplicity of functions often attributed to insulators (e.g. enhancer blocker, chromatin barrier, transcriptional regulator). Ultimately, a characterization of the locus-specific composition of insulator complexes and their locus-specific function may be required to obtain a general picture of insulator function.
In mammals, CTCF is the only insulator protein identified so far, but other factors, such as cohesin have been identified as necessary and essential for the formation of CTCF-mediated long-range interactions [28], [30], [32]. Mammalian CTCF contains eleven zinc-fingers, and it has been shown that different combinations of zinc-fingers could be used to bind different DNA sequences [78]. Thus, in mammals CTCF may play the role of first layer insulator protein, whereas other factors such as cohesin or mediator may play the role of second layer insulator proteins [31].
This model proposing different functional roles for insulator factors could also explain the mechanism by which insulators are able to help establish and reinforce the transcriptional state of chromatin domains throughout cell division. First layer proteins remain bound to chromatin at all stages of the cell cycle [15], [79]. In contrast, both CP190 and Chromator are chromatin-bound during interphase but display a drastic redistribution during mitosis: CP190 strongly binds to centrosomes while Chromator co-localizes to the spindle matrix [22], [43]. Thus, the dissociation and cellular redistribution of second layer insulator proteins during cell division would be responsible for the massive remodeling of chromosome architecture occurring during mitosis, and for the re-establishment of higher-order contacts at the onset of interphase. In contrast, first layer insulator proteins would act as anchor points for the re-establishment of higher-order interactions after mitosis, and for the maintenance of the transcriptional identity of physical domains. Thus, our model suggest distinct roles for insulator binding proteins and co-factors in actively shaping the organization of chromatin into physical domains during the cell cycle. This model is consistent with recent genome-wide data suggesting that, overall, first layer insulator proteins remain bound to their binding sites during mitosis, whereas second layer insulator proteins tend to show a large change in binding patterns [79], [80]. Further genome-wide and microscopy experiments will be needed to quantitatively test this model.
DNA plasmids were propagated in E. coli DH5a or in DB3.1 cells (depending on vector used). Proteins were expressed and purified from E. coli BL21 (DE3)-pLysS cells (Invitrogen) as described elsewhere [81]. Details on vectors, primers, protein constructs and protein purification procedures can be found in Text S1 and in Supplementary Tables S2, S3.
A 447 bp genomic region containing the Tudor-SN locus was subcloned into pTST101 to make pTST101-447pos (oligonucleotides are shown in Supplementary Table S4). pTST101-447pos was digested by NdeI, HindIII, and SalI resulting in three linear fragments, including DNAtudor (1627 bp long dsDNA fragment containing the 447 bp Tudor-SN locus) and two additional dsDNA fragments (750 and 4025 bp). Restricted pTST101 (1.7 nM) was incubated with increasing amounts of purified BEAF32, CP190 or Chromator in 150 mM NaCl, 30 mM Tris/HCl pH 7.4, 5 mM mercaptoethanol. A gel loading buffer (50% glycerol, 50 mM Tris/HCl pH 7.4) was added and the DNA-protein mixture was directly analyzed in a 1% TAE agarose gel. DNA was labeled using Sybersafe (Invitrogen) and visualized on a gel imaging system (Image Station 4000 MM Pro–Carestream Molecular Imaging). No difference in binding specificity was observed when DNA competitors (e.g. dIdC) were added to the protein-DNA mix. For super-shift assays, the 447 bp Tudor-SN locus (chromosome 3L: 264375–264822) was PCR amplified from S2 Drosophila genomic DNA. Purified proteins were added to the DNA in a reaction mixture in a total volume of 20 µl and incubated for 10 min on ice. A gel loading solution (50% glycerol, 50 mM Tris/HCl pH 7.4) was added and the DNA-protein mixture was directly analyzed on a 2% TAE agarose gel.
Fluorescence anisotropy experiments used short, 5′-Cy3B labeled DNA fragments (DNAS and DNANS, Eurogentec, oligonucleotide sequences are shown in Supplementary Table S5). Anisotropy measurements were carried out using a Tecan Safire II micro plate reader fluorimeter and a Corning 384 Low Flange Black Flat Bottom plate. All measurements were carried out in 30 mM Tris/HCl pH 7.5, 0.01 mg/ml BSA, 0,004% Tween20, 100 mM NaCl, 20 µM ZnSO4, 5 mM mercaptoethanol in a final volume of 60 µl. DNA binding studies were performed by adding increasing amounts (0–800 nM) of purified proteins to 2.5 nM of Cy3B or atto-655 5′-labeled 58-bp dsDNA. Dissociation measurements were performed by adding large amounts (up to 1000 nM) of unlabeled DNAS or NaCl (350 mM final). Further details can be found in Text S1.
Reaction buffers and DNA substrates (at a final DNA concentration of 2.5 nM) were the same as those used for fluorescence anisotropy (oligonucleotide sequences are shown in Supplementary Table S5). Fluorescence correlation and cross-correlation experiments were carried out on a custom-built setup allowing Pulse Interleaved Excitation (PIE) with Time Correlated Single Photon Counting (TCSPC) detection as described elsewhere [53]. It is important to note that our measurements allow us to detect only 50% of the complexes involved in bridging, as complexes containing two DNA molecules with the same color do not contribute to the cross-correlation amplitude. Further details on PIE-FCS and the models used to fit data can be found in Text S1.
Drosophila S2 cells (DGRC) were grown in Schneider cell medium supplemented with 10% calf serum. 3×106 cells were centrifuged for 10 min at 1000 g and 4°C. All subsequent steps were performed on ice. Cells were washed twice in PBS and resuspended in hypotonic lysis buffer (10 mM Tris/HCl pH 7.5,10 mM KCl, 1.5 mM MgCl2, complete EDTA-free protease inhibitors (Roche)), and washed again twice with hypotonic buffer. After 30 min on ice, lysed cells were pushed through a 25G needle. In addition, lysates were washed with hypotonic buffer and centrifuged at 1000 g. Nuclei were resuspended in nuclear lysis buffer (300 mM KCl, 50 mM Tris/HCl Ph 7.5,10% glycerol, 1% Triton ×100, and protease inhibitors) with benzonase (Novagen, 71206) and incubated for 30 min on a rotating wheel at 4°C. Next, nuclear lysates were centrifuged at 14000 g for 15 min at 4°C. The supernatant was transferred to a clean tube. This resulted in 200 µl of nuclear extract with a total protein concentration of ∼20 mg/ml. This protocol was adapted from Hart et al. [82].
Purified proteins/S2 nuclear extracts were separated on a 10–12% SDS-Polyacrylamide-gel and electro-blotted for 1 h at 100 mV onto a nitrocellulose membrane (Protran* Nitrocellulose Membrane Filters, Whatman*). Next, membranes were blocked (3% BSA in TBST) for 1 h and subsequently washed (1% BSA in TBST) before incubation for 1 h with polyclonal purified primary antibody (guinea-pig-anti-Chromator/rabbit-anti-CP190 or mouse-anti-BEAF32 from DSHB). Several washing steps (1% BSA in TBST) followed before the incubation with HRP-labeled secondary antibody (goat anti-guinea pig IgG-HRP Conjugate Thermo scientific, Goat anti-Mouse IgG (H+L)-HRP conjugate Pierce, goat anti-rabbit IgG (H+L)-HRP Conjugate Biorad) for 40 min. After further washing steps the membrane was developed using Pierce ECL Western Blotting Substrate and imaged (Image Station 4000 MM Pro – Carestream Molecular Imaging).
Purified polyclonal antibodies (anti-Chromator (60 µg), anti-CP190 (60 µg), control goat-IgG (90 µg) were immobilized (2 h, room temperature) on 100 µl Amino Link Plus Coupling agarose-bead-slurry (Pierce Co-Immunoprecipitation Co-IP Kit) following the manufacturer instructions. Different concentrations of heterologous purified proteins or 100 µl of S2 nuclear extract (20 mg/ml) including protease inhibitor (Roche, EDTA free) were added for control goat-IgG, guinea-pig-anti-Chromator, or rabbit-anti-CP190 immobilized agarose beads in IP-Lysis buffer (part of the Cp-IP Pierce kit, total volume 400 µl) and incubated on a rotary wheel for 1–3 h at 4°C in a final volume of 400 µl. Depending on the bait protein used, the bead-antibody-protein-complex was washed several times with 400 µl IP lysis-buffer, followed by PBS including 200–1000 mM NaCl until no protein could be detected in the washing step. Elution was carried out after incubating the protein-bead complex for 3 min in elution buffer at pH 2.8. Eluted proteins were analyzed by Western-blot-analysis.
Aggregation plots were obtained from genome-wide data from Sexton et al. [13], and were constructed by following the strategy developed by Jee et al. [83]. First, interacting Hi-C DpnII bins containing genomic features of interest (BEAF32, CP190 or Chromator) were identified. BEAF32 binding sites were considered as anchors and CP190, Chromator or both sites as targets [83]. Second, only LRI at distances between 15 and 60 kbp and containing BEAF32 in the anchor and CP190/Chromator in the target were further considered. The lower limit was set to 15 kbp, as significantly high background levels occur for bins at distances <15 kbp. The upper limit (60 kbp) was set to be smaller than the average size of topological domains [13]. Third, Hi-C interaction profiles were binned in 500 bp windows +/−5 kbp around the target site. Next, target sites were aligned, aggregated together, and normalized (blue solid lines, Figure 6E). Internal controls (grey lines, Figure 6E) were obtained by using the same procedure but for target sites that did not contain any of the features (CP190 or Chromator). This procedure generated background interaction levels reflecting the chromatin context of the anchor site. Frequencies of interactions were statistically tested by Wilcoxon tests.
For the analysis of ChIP-chip data (Venn diagrams), publicly available .gff3 files were downloaded from the modENCODE website (http://data.modencode.org/) corresponding to CP190, BEAF32 and Chromator/Chriz ChIP-chip experiments performed in BG3 and S2 cells [48], [84] (datasets 274, 275, 278, 279, 280, 921, 924). Overlaps between binding sites were calculated with the intersectBed function of the BEDTools software [85]. Venn diagrams were generated with the vennDiagram package in R.
Crystallization trials was carried out by the sitting-drop technique using the classic, PEG, PACT and AmSO4 suites (Quiagen, France) and low-profile microplates (Grenier, France) at room temperature. 0.5 µl protein solution was mixed with an equal volume of reservoir solution. Several conditions yielded crystals. Optimizations were done with the hanging-drop vapor diffusion technique. 1 µl protein solution was mixed with 1 µl of reservoir. We obtained well diffracting crystals (2.03 Å) using 0.8 M NaH2PO4, 0.8 M KH2PO4, 0.1M Hepes/pH 7.5. Crystals were soaked in 30% glycerol for cryoprotection and diffraction data were collected under cryogenic conditions on our laboratory anode and at the European Synchrotron Radiation Facility (ESRF, Grenoble). Image data were processed and scaled using the programs MOSFLM (Leslie, 1999) and SCALA of the CCP4 suite [86]. The crystal belonged to space group P3221 with unit cell parameters a = b = 84.98 Å, c = 40.87 Å, α = β = 90° and γ = 120°.
The structure of CP190-BTB/POZ was solved by molecular replacement with an in-house dataset at 2.3 Å resolution using the program PHENIX (phenix.autoMR) [87] and a combination of five partial models extracted from the server TOME [88] used to gather potential templates through fold-recognition. Structure refinement and rebuilding were performed with COOT [89], PHENIX (phenix.refine) [87] and REFMAC (Murshudov et al, 1997) from the CCP4 suite [86] using a dataset recorded at the ESRF at 2.0 Å resolution. Data collection and refinement statistics are summarized in Supplementary Table S1. The structure has been deposited with the Protein Data Bank (PDB 4U77).
|
10.1371/journal.pgen.1005873 | A Novel Epigenetic Silencing Pathway Involving the Highly Conserved 5’-3’ Exoribonuclease Dhp1/Rat1/Xrn2 in Schizosaccharomyces pombe | Epigenetic gene silencing plays a critical role in regulating gene expression and contributes to organismal development and cell fate acquisition in eukaryotes. In fission yeast, Schizosaccharomyces pombe, heterochromatin-associated gene silencing is known to be mediated by RNA processing pathways including RNA interference (RNAi) and a 3’-5’ exoribonuclease complex, the exosome. Here, we report a new RNA-processing pathway that contributes to epigenetic gene silencing and assembly of heterochromatin mediated by 5’-3’ exoribonuclease Dhp1/Rat1/Xrn2. Dhp1 mutation causes defective gene silencing both at peri-centromeric regions and at the silent mating type locus. Intriguingly, mutation in either of the two well-characterized Dhp1-interacting proteins, the Din1 pyrophosphohydrolase or the Rhn1 transcription termination factor, does not result in silencing defects at the main heterochromatic regions. We demonstrate that Dhp1 interacts with heterochromatic factors and is essential in the sequential steps of establishing silencing in a manner independent of both RNAi and the exosome. Genomic and genetic analyses suggest that Dhp1 is involved in post-transcriptional silencing of repetitive regions through its RNA processing activity. The results describe the unexpected role of Dhp1/Rat1/Xrn2 in chromatin-based silencing and elucidate how various RNA-processing pathways, acting together or independently, contribute to epigenetic regulation of the eukaryotic genome.
| Epigenetic mechanisms regulate when, where, and how an organism uses the genetic information stored in its genome. They are essential to many cellular processes, such as the regulation of gene expression, genome organization, and cell-fate determination. They also govern growth, development, and ultimately human health. Heterochromatin constitutes silenced chromatic domains, in which gene silencing occurs through epigenetic mechanisms. RNA processing pathways, such as RNA interference (RNAi) and the exosome, are known to mediate the silencing of genes via degradation of unwanted or aberrant transcripts. In this study, we describe a new RNA processing mechanism in epigenetic silencing using fission yeast, a premier model for studying these processes. With genetic, cell biology, and genomic approaches, we uncovered a previously unrecognized function of Dhp1, a highly conserved 5’-3’ exoribonuclease and ortholog of budding yeast Rat1 and metazoan Xrn2. We show that Dhp1 mediates a novel RNA processing mechanism in epigenetic silencing which occurs independently of both RNAi and the exosome. Our results clarify how multiple RNA processing pathways are involved in the regulation of eukaryotic gene expression and chromatin organization.
| In eukaryotic cells, DNA coils around histones to form nucleosomes, which are packaged into chromatin. Various post-translational modifications (PTMs) of histones, histone variants, and nucleosome remodeling factors confer distinct chromatin states on genes and facilitate the organization of large chromatin tracts into domains [1–3]. Euchromatic domains (euchromatin) contain actively transcribed genes and are enriched with hyperacetylated histones, while heterochromatic domains (heterochromatin) contain highly repetitive elements that are transcriptionally silenced and are associated with hypoacetylated histones [4–7]. In addition to its primary role in transcriptional gene silencing, heterochromatin is crucial for centromere-mediated chromosome segregation, cell fate determination, and the silencing of repetitive DNA elements [4]. In fission yeast, Schizosaccharomyces pombe (S. pombe), its formation requires both histone hypoacetylation and histone H3 methylation at lysine 9 (H3K9me), which provides a binding site for HP1 family proteins [8–10]. Heterochromatin is first nucleated at specific repetitive loci and subsequently spread for up to hundreds of kilobases (kb) into surrounding regions [4, 11, 12]. Once established, these silenced heterochromatic domains are heritable, and can be stably maintained through successive cell divisions [4, 13, 14].
Epigenetic silencing includes both transcriptional (TGS) and post-transcriptional gene silencing (PTGS). In general, heterochromatin limits the access of RNA polymerase II (RNAPII) machinery to the DNA template and can therefore mediate transcriptional gene silencing (TGS) by preventing unwanted transcription from a given genomic region [15, 16]. PTGS employs RNA processing machinery to rapidly degrade nascent RNAs to repress gene expression or to protect the genome from foreign genetic elements such as retroviral RNA or transposable DNA [17–20]. RNA processing machineries ensure the maturation and packaging of RNAs from longer precursors into mRNA/protein particles (mRNPs) before they are exported to the cytoplasm for translation [21]. Most of these processing events, such as addition of the 5’ cap, removal of introns, and polyadenylation at the 3’ end, occur while the RNA is still attached to RNAPII and chromatin, and are therefore referred to as transcription-coupled RNA processing events [22, 23]. For example, RNA endocleavage at a polyadenylation (polyA) site is commonly required for transcriptional termination because the 5’ to 3’ exoribonucleolysis of the exposed 3’ fragment downstream of the polyA site facilitates RNAPII release from chromatin [24, 25]. Besides their roles in RNA maturation and RNAPII termination, RNA-processing enzymes act as quality control systems, screening partly or fully transcribed products and degrading abnormal RNAs [22, 26]. RNA processing pathways play an active role in epigenetic silencing, especially PTGS [27], as many nuclear processes rely on the fine balance between RNA maturation and destruction to regulate gene expression [28].
The best understood RNA processing pathway in PTGS is the RNA interference (RNAi)-mediated formation of heterochromatin at centromeric regions in S. pombe [27, 29, 30]. In RNAi, RNAPII transcripts originating from repetitive DNA regions are converted to double-stranded RNA by the RNA-Dependent RNA Polymerase Complex (RDRC) [31]. They are then processed by Dicer into small interfering RNAs (siRNAs) [32, 33] and loaded onto the Argonaute-containing RNA Induced Transcriptional Silencing (RITS) complex, which targets the repeat regions through the homology of the siRNA sequence [34]. RITS associates with chromatin by direct interaction with H3K9me [35], then recruits the Clr4 complex to initiate chromatin remodeling [36–39]. Recently, several studies reported an RNAi-independent RNA-processing pathway in heterochromatin assembly at the centromeric region and some heterochromatic islands in euchromatic regions [40–43]. This new pathway is mediated by the exosome complex, which degrades unwanted RNAs via its 3’-5’ exoribonuclease activity [26]. Although both RNAi and exosome pathways are RNA-mediated and involved in processing long noncoding RNAs (ncRNAs) into small RNAs, how the exosome pathway contributes to heterochromatin assembly is not well understood. In addition, it is not known whether other RNAi-independent RNA processing pathways participate in epigenetic silencing.
Here we report a new epigenetic silencing pathway involving Dhp1, a conserved 5’→3’ exoribonuclease, the ortholog of budding yeast Rat1 and metazoan Xrn2 known to promote termination of RNA polymerase II (RNAPII) transcription [44–46]. We show that Dhp1-mediated heterochromatic silencing is independent of Din1, an ortholog of budding yeast Rai1 that has been shown to stabilize Dhp1/Rat1 exoribonuclease activity [47]. In addition to maintenance of gene silencing, Dhp1 contributes to de novo establishment of heterochromatin at the centromeres and the silent mating type region. It also plays a role in the transcriptional-dependent spreading of heterochromatin. Importantly, Dhp1 interacts with heterochromatic factors and its catalytic activity is required for its role in silencing. Further genetic analyses indicate that Dhp1 operates in a distinct pathway parallel to RNAi and the exosome to mediate heterochromatic gene silencing. Finally, RNAPII localization and transcriptome-wide maps of RNAs associated with RNAPII revealed that Dhp1 likely acts at the post-transcriptional level to affect gene silencing. We propose that, in addition to RNAi and exosomes, Dhp1 constitutes a distinct RNA-processing pathway that enforces post-transcriptional gene silencing across the fission yeast transcriptome.
Dhp1/Xrn2 is an essential gene required for transcriptional termination and RNA quality control [22, 45, 46]. A recent study reported that impairment of transcription termination is sufficient to induce the formation of heterochromatin at protein-coding genes by trans-acting siRNAs in S. pombe [48]. However, it is not clear whether impaired transcription termination would alter epigenetic silencing at the major heterochromatic regions such as centromeric regions and the mating type locus. We therefore tested whether silencing at these regions is affected in dhp1 mutant cells. Because its loss is lethal, we utilized a conditional temperature-sensitive (ts) allele, dhp1-1, which codes a truncated carboxyl-terminal form of Dhp1 that is partially replaced by a ura4+ transgene and is lethal at 37°C (dhp1-1>>ura4+) [47]. We generated an independent dhp1-2 mutant, which carries the same carboxylic terminal truncation but is replaced with a nourseothricin resistance gene (NatN2) (S1 Table). This allele has a less severe ts phenotype compared to dhp1-1, and is not lethal at 37°C (S1A Fig). dhp1-2 is not fused with ura4+, allowing us to investigate the silencing of the centromeric region by analyzing the expression of a ura4+ reporter gene inserted at the outer centromeric region (otr∷ ura4+) (Fig 1A, top panel). Wildtype cells carrying the reporter otr∷ura4+ grew well on counter-selective medium containing 5-Fluoroorotic Acid (5-FoA) indicating otr∷ura4+ was silenced (Fig 1A, bottom panel). The growth of dhp1-2 was greatly inhibited in the presence of 5-FoA indicating defective silencing of the reporter gene. Surprisingly, din1-null (din1Δ) cells do not have a severe growth or centromeric silencing defect, suggesting that Dhp1 is involved in a silencing pathway separate from its Din1-related activity (Fig 1A, bottom panel). To further examine the observed silencing defect, an ade6+ reporter gene was inserted into the silenced mating type region (Fig 1A, top panel). We can easily observe the silencing status of the ade6+ based on color; cells that cannot express the normal level of the reporter gene accumulate a red pigment due to blocked adenine biosynthesis. In wildtype cells, the reporter gene is silenced by heterochromatin, resulting in red/sectored colonies on low adenine media at 30°C. Cells lacking Clr4, the sole histone H3K9 methyltransferase in S. pombe [49], form white colonies due to loss of heterochromatin (Fig 1A). Similar to clr4Δ, dhp1-2 but not din1-null (din1Δ) cells form white colonies, indicating a silencing defect at the mating type locus unique to the dhp1 mutant. Since dhp1-1 has a more severe silencing defect than dhp1-2 as evaluated by silencing assay and quantitative Reverse Transcription Polymerase Chain Reaction (qRT-PCR) (S1 Fig), the rest of our studies focused on using the dhp1-1 allele. We next assessed whether loss of reporter gene silencing in the dhp1 mutant is correlated with the expression of heterochromatic repeats. Transcript analysis by qRT-PCR revealed substantially unregulated expression of repeats associated with pericentromeric heterochromatin and the silent mating type locus in dhp1 but not din1 mutant cells (Fig 1B). Further analysis by expression profiling using a tiling microarray on both DNA strands showed increased expression throughout both heterochromatic regions in dhp1-1, well above the increase observed in din1Δ cells (Fig 1C). These results indicate that Dhp1 plays a previously unrecognized, Din1-independent function in epigenetic silencing.
To avoid potential pleotropic effects caused by dhp1 mutation, we performed all our experiments at 30°C. This is a permissive temperature for dhp1-1, in which silencing defects but no obvious growth deficiency are observed (Fig 1A), suggesting that most of transcription-related functions of Dhp1 are retained. In addition, we carefully analyzed the transcriptional levels of all known heterochromatic factors in dhp1-1 cells using expression data. All coding transcripts of these proteins are affected less than 1.6-fold (a typical threshold difference for microarray data) compared to that of wildtype (S2 Table). Further, a plasmid-borne wildtype copy of dhp1+ rescues the ts phenotype of dhp1-1 (S2A Fig) and a diploid heterozygous strain carrying a wildtype and a dhp1-1 allele showed the same phenotype as a wildtype diploid strain (S2B and S2C Fig), demonstrating that dhp1-1 has no dominant negative effects. Altogether, these findings suggest that the loss of heterochromatic silencing in the dhp1 mutant is likely a direct consequence of impaired function of Dhp1 at heterochromatin rather than reduced transcription of heterochromatic factors.
The silencing defect in dhp1-1 is unexpected because impaired transcription termination would reduce RNAPII transcription and the subsequent release of the RNA from the site of transcription, which may enhance the assembly of heterochromatin through induction of RNA-mediated chromatin modification such as H3K9 methylation (H3K9me) [48, 50]. In addition, many reported Dhp1 functions are associated with Din1, which contributes to the generation of the proper substrates for Dhp1’s exoribonuclease activity [46, 51]. Because we did not observe a silencing defect in din1Δ cells, we wondered whether Din1, like Dhp1 is involved in transcription termination. According to the “Torpedo model” [24, 25], Dhp1-mediated exonucleolysis of the cleaved 3’ fragment downstream of the mRNA polyA site facilitates RNAPII release from chromatin. Deficiency in Dhp1/Din1 will cause RNAs to accumulate at the 3’ end of genes, due to an RNAPII transcription termination defect [24, 25]. To confirm this reported role of Dhp1/Din1, we analyzed the transcriptomes of dhp1-1 and din1Δ cells at euchromatic regions. We detected a genome-wide increase of RNA levels at the 3’ end of genes compared to wildtype in both mutants, with a larger fraction of genes exhibiting transcription termination defects in dhp1-1 (S3 Fig). While the role of Dhp1 is more dominant than that of Din1, these results support earlier studies indicating that both Dhp1 and Din1 participate in RNAPII transcription termination.
As an interacting protein of Rat1/Xrn2, Rtt103 also contributes to transcription termination in yeast and humans [52–54]. Rhn1, the S. pombe ortholog of Rtt103, has a reported role in the suppression of meiotic mRNAs during vegetative growth [54]. However, whether it plays a role in heterochromatic silencing has not been reported. To further examine whether defective transcription termination is crucial for Dhp1-mediated epigenetic silencing, we compared the expression of repeat elements in wildtype and rhn1Δ cells using qRT-PCR and found that, like Din1, loss of Rhn1 did not cause a silencing defect (S4 Fig). Taken together, our data argue that Dhp1 plays a novel role in epigenetic silencing, which cannot be explained by its established function in transcription termination.
Cells with mutations in factors that contribute to epigenetic silencing often exhibit defects in chromosome segregation as determined by their sensitivity to the microtubule-destabilizing drug thiabendazole (TBZ) [55]. Because heterochromatin formation has been linked to centromere function in various organisms including S. pombe [56–58], we tested whether the dhp1 mutants are sensitive to TBZ, which would indicate impaired function of centromeric heterochromatin, resulting in a chromosome segregation defect. As expected, deletion of clr4 abolishes heterochromatin and causes severe TBZ sensitivity (Fig 2A). Our assay clearly shows that dhp1, but not din1, mutant cells are sensitive to TBZ, suggesting a chromosome segregation defect specific to dhp1 mutants (Fig 2A). To further examine the role of Dhp1 in chromosome segregation, we sporulated wildtype and mutant h90 strains to follow the segregation of chromosomes in tetrads using a fluorescence-based analysis (S5 Fig). To sporulate, two haploid cells with opposite mating types conjugate to form a zygote which then enters meiosis. During meiosis, cells undergo two consecutive rounds of chromosome segregation. A normal meiosis results in an ascus in which each of four spores contain relatively equal amounts of DNA (DAPI dots). Abnormal meiotic segregation within a tetrad will show an uneven distribution of DAPI staining in each spore, resulting in less than or greater than four dots. We found that meiotic chromosome segregation is severely perturbed in the dhp1-1, but not in din1Δ cells, with nearly 50% of tetrads containing abnormal numbers of DAPI dots (≤ 3 or ≥ 5) (S5 Fig).
To determine whether the chromosome segregation defect seen in the dhp1 mutant is linked to its role in epigenetic silencing at major heterochromatic domains, we assessed the status of H3K9me-associated heterochromatin by a Chromatin-Immunoprecipitation (ChIP) assay. Although no reduction of H3K9me2 was seen at the endogenous repetitive regions (Fig 2D and 2E), the levels of H3K9me2 at the reporter genes embedded in these regions were substantially reduced at these loci in cells deficient in dhp1 (Fig 2B and 2C). Loss of din1 has no negative effect on the enrichment of the H3K9me mark at either the endogenous repetitive regions or the reporter genes (Fig 2B–2E). These results suggest that Dhp1’s role in chromosome segregation is linked to its requirement to maintain functional heterochromatin at the centromeres.
We next wondered whether Dhp1 interacts with heterochromatic proteins, which would support a direct role of Dhp1 in facilitating heterochromatin assembly. We purified Dhp1 and Din1 through two-step affinity purification (S6 Fig) and identified the co-purified proteins by mass spectrometry analysis (S3 Table). Strains used for purification carry a functional Dhp1 or Din1 fused with FTP, a modified TAP tag comprising a protein A motif and a FLAG tag separated by a TEV protease cleavage site. Since Dhp1 and Din1 are associated with transcribing RNAs and chromatin, we performed all purifications in the presence of Benzonase to avoid indirect protein-protein interactions mediated by nucleic acids. Din1 is the major Dhp1-interacting protein as recovered Din1 peptides were found to be approximately 50% as abundant as those of the bait protein, Dhp1. As expected, Dhp1 also co-purified with many RNAPII-related factors, consistent with its role in transcriptional termination. In particular, it is associated with several heterochromatic proteins, including Clr4 methyltransferase complex (ClrC) subunit Rik1 and exosome subunit Rrp6. These data are consistent with interactions recently identified in a parallel study [59]. Notably, these heterochromatic proteins were not present in those fractions when Din1 was used as the bait, supporting a distinct role of Dhp1 in heterochromatic formation. Our data indicates that Dhp1 interacts with heterochromatic proteins and is likely directly involved in heterochromatin assembly.
Multiple pathways are utilized to initiate epigenetic silencing including both RNA and DNA sequence-dependent mechanisms [9, 60, 61]. Several studies have shown that in S. pombe, both RNAi and the exosome contribute to the initiation of silencing at the centromere by processing RNAs transcribed from repetitive regions [13, 41, 61]. We next sought to determine whether Dhp1 also contributes to this process through examination of reporter gene expression following the reintroduction of functional clr4+ into dhp1-1 clr4Δ double mutant cells (Fig 3A). Deletion of clr4 results in the abolition of H3K9me and the loss of heterochromatin. However, reintroduction of functional clr4+ is sufficient for de novo heterochromatin formation as previously reported [50] (Fig 3B). One of the key members in RNAi machinery, Dcr1, is the sole Dicer-family endoribonuclease in S. pombe [62]. dcr1Δ cells lose the ability to initiate heterochromatin formation de novo at repeat regions [13]. Consistent with previous findings, silencing at the centromeric region cannot be efficiently established without Dcr1 [13]. Reintroduction of clr4+ into clr4Δ cells shows a complete alleviation of TBZ sensitivity, while clr4+ reintroduction into dcr1Δ clr4Δ double mutant cells has no effect (Fig 3B). Additionally, complementation of clr4+ has little effect on the relative expression of centromeric- and mating type locus-specific repeats in dcr1Δ clr4Δ cells, however silencing of these repeats is fully resumed in clr4Δ single mutants (Fig 3B). Reintroduction of clr4+ in dhp1-1 clr4Δ double mutant cells partially resumed the silencing at the centromeric region as indicated by qRT-PCR (Fig 3B). At the mating type locus, silencing is barely restored in dhp1-1 clr4Δ, having 10-fold more expression than dhp1-1 alone (Fig 3B). H3K9me2 ChIP analysis further demonstrated that without functional Dhp1, H3K9me2 is partially re-established at the centromeric region, but only a low level of H3K9me2 can be found at the mating type locus after complementation of clr4+ (Fig 3C and 3D). These results indicate that Dhp1 is essential for efficient de novo heterochromatin assembly at peri-centromeres and the silent mating type locus.
Spreading of heterochromatin from the nucleation sites enables the establishment of a heterochromatin domain spanning many kbps [12]. Although the mechanism is poorly understood, it depends on the oligomerization of chromatin modifiers such as HP1 and Tas3, a RITS component, and the actions of Swi6-recruited histone deacetylases (HDACs) on adjacent nucleosomes [63–66]. While the polymerization of chromatin modifiers indeed constitutes a major part of heterochromatin spreading, a role for RNAPII in transcription-mediated spreading is currently being explored. The effect of spreading on the silencing of reporter genes inserted into centromeric repeat regions has been shown to vary with position relative to the RNAPII promoter; reporter genes downstream of the promoter are more effectively silenced than those inserted upstream [67, 68]. While the molecular details remain unclear, transcription-mediated spreading appears to require transcription of the 3’ untranslated region as well as degradation of these transcripts by RNAi machinery [67].
Decreased enrichment of H3K9me at the reporter genes in dhp1-1 suggests that Dhp1 may be involved in the spreading of H3K9me (Fig 2B and 2C). Because of its role in transcriptional termination, we next examined whether Dhp1 is required for RNAi-dependent spreading of heterochromatin, which partially relies on RNAPII transcription [68]. To this end, we adopted a spreading assay [13, 69] (Fig 4A). Nucleation of heterochromatin at cenH is dependent on RNAi and the cenH sequence itself [13, 69]. Inserting the cenH sequence into a euchromatic locus causes ectopic establishment followed by spreading of heterochromatin and subsequent silencing of proximal genes [13]. By coupling ectopic cenH with an adjacent reporter gene (ade6+), we can directly observe the effects of spreading of silencing to the proximal reporter gene (Fig 4A). In the wildtype background, about 35% of cells form pink/sectoring colonies, indicating the spreading of heterochromatin assembled at ectopic cenH to the ade6+ reporter gene. Cells lacking ago1, a critical factor in RNAi form white colonies at 100% efficiency showing that the ade6+ reporter gene cannot be silenced. This is consistent with previous results indicating that RNAi machinery is required for the transcriptional-dependent spreading of heterochromatin assembled at cenH region [13, 68]. Similar to ago1Δ, no pink colonies were formed in dhp1-1 background (Fig 4B). H3K9me2 ChIP using multiple primers along ade6+ and its surrounding regions shows moderate reduction of this heterochromatic mark in dhp1-1 cells compared to wildtype cells (Fig 4C). Altogether, these results suggest that Dhp1 plays a role in the transcription-related spreading of the H3K9me mark.
While RNAi is critical to the establishment and spreading of heterochromatin, it is dispensable for the maintenance of a previously assembled chromatin state at the mating type locus and sub-telomeric regions [64, 70]. To investigate the role of Dhp1 in heterochromatin maintenance, we introduced the dhp1 mutation into cells that lack part of the K region in the mating type locus, but continue to repress a proximal ade6+ reporter gene (KΔ∷ade6+off) (Fig 5A). Deleting the K region results in loss of heterochromatin establishment within the mating type locus, but because heterochromatin is stably inherited through cell division, derepression is rarely seen without concomitant loss of the maintenance machinery [71]. Loss of maintenance machinery will result in derepression of ade6+ (KΔ∷ade6+off will switch to KΔ∷ade6+on). While the molecular mechanisms which mediate maintenance remain unclear, Clr4 and Swi6 have been implicated [13, 38, 72]. We detected partial loss of repression of ade6+ in dhp1-1, an intermediate phenotype between wildtype and swi6Δ (Fig 5B and 5C). In the wildtype background, nearly 80% of cells form dark red colonies and only about 20% of cells form pink colonies. Unlike swi6Δ, which form 100% white colonies, about 75% of dhp1-1 cells form pink colonies, although no dark red colonies were ever observed. We further analyzed the ade6+ RNA level by qRT-PCR (Fig 5D). Indeed, we observed a more than 20-fold increase the amount of ade6+ transcripts in dhp1-1 cells compared to that of wildtype cells. Consistent with previous studies, loss of Swi6 abolishes the enrichment of H3K9me2 at KΔ∷ade6+ (Fig 5E)[72]. Interestingly, mutation of dhp1 does not reduce this histone modification at the same region (Fig 5E), suggesting that Dhp1 plays a role in effective maintenance of epigenetic silencing downstream of H3K9me.
We consistently observed a stronger silencing defect in dhp1-1 at the mating type locus than the pericentromeric region (Figs 1–3). RNAi is known to play a major role in silencing centromeric repeats but only partially contributes to silencing at the mating type locus [64, 65]. These results suggest that Dhp1-mediated silencing might be distinct from that of RNAi. To test this, we combined dhp1-1 with a deletion of ago1, the sole Argonaute protein in S. pombe [62], and analyzed the silencing defect at the centromeric region and the mating type locus by qRT-PCR. Consistent with previous findings, loss of Ago1 caused an upregulation of centromeric repeat transcripts (Fig 6A) and did not show an obvious silencing defect at the mating type locus (Fig 6B). Whereas dhp1-1 exhibited a modest increase in transcription at the centromere and the mating type locus, a double mutant dhp1-1 ago1Δ showed a large increase beyond the cumulative effects of either single mutation (Fig 6B), indicating that Dhp1 contributes to heterochromatic silencing in a pathway parallel to RNAi. Our conclusion was also supported by the expression data shown in Fig 3B: we consistently observed more relative expression of repeats in dhp1-1 dcr1Δ double mutant cells compared to that of dcr1Δ or dhp1-1 single mutant cells from both the centromeric region and mating type locus, further indicating an RNAi-independent role of Dhp1.
Many transcripts degraded by RNAi are also targets of Rrp6 [42], the catalytic subunit of the nuclear exosome required for rapid elimination of cryptic unstable transcripts (CUTs) [73–75]. Its RNA degradation activities act in parallel with RNAi to promote heterochromatin assembly [43, 50]. Since Dhp1 is an exoribonuclease and plays an independent role from RNAi, we next wondered whether Dhp1 has overlapping function with Rrp6 in the silencing of repeat elements. Indeed, qRT-PCR showed that dhp1-1 rrp6Δ double mutant cells have stronger silencing defects at both the centromeric region and mating type locus (Fig 6C and 6D), although the effect is less than that observed in dhp1-1 ago1Δ.
We next investigated whether the accumulated silencing defects in double mutants of dhp1 with ago1Δ or rrp6Δ are resultant from additive deficiencies of H3K9me2. ChIP experiments show that, except for dhp1-1 ago1Δ at the centromeric repeats, none of the double mutants exhibit further reduction of H3K9me2 compared to single mutants (Fig 6C and 6D), suggesting the role of Dhp1 in epigenetic silencing does not rely on H3 K9 methylation at repeat regions. Notably, combining dhp1-1 and rrp6Δ mutations enhances H3K9me both at the centromeric region and the mating type locus (Fig 6C and 6D). A recent study reported that Rrp6 is required for RNAPII termination at specific targets [73]. Our observation of enhanced H3K9me occurring in dhp1-1 rrp6Δ double mutant cells suggests that transcription termination defects impair RNAPII transcription and favor the induction of RNA-mediated chromatin modification such as H3K9 methylation. In dhp1-1 rrp6Δ, the compounded silencing defect must be overcompensating for the increased silencing effect of the additively enhanced H3K9me (S7 Fig and discussion). Collectively, these results show that the Dhp1-mediated silencing mechanism is independent of both RNAi and the exosome, and is likely downstream of H3K9me.
Dhp1 is a conserved 5’-3’ exoribonuclease [44, 46]. Previous studies of Xrn1 in Kluyveromyces lactis (K. lactis) indicated that switching the acidic aspartate at position 35 or glutamate at position 178 to neutral residues, such as alanine (D35A) or glutamine (E178Q), completely abolished enzymatic activity [76]. In S. pombe, Dhp1D55 and E207 are conserved residues corresponding to K. lactis Xrn1D35 and E178 (Fig 7A). To test whether the RNA processing activity of Dhp1 is important for its role in epigenetic silencing, we generated plasmids carrying a copy of dhp1 with both D55 and E207 mutated (dhp1-D55A E207Q), which abolishes the catalytic activity of Dhp1. A plasmid carrying a wildtype allele of dhp1+ can rescue both the ts phenotype and the silencing defect of dhp1-1 analyzed by dilution assays (Fig 7B) and qRT-PCR (Fig 7C and 7D), suggesting that the wildtype allele can completely complement the C-terminal truncated form of dhp1, further supporting that there is no dominate negative effect of dhp1-1. However, the plasmid carrying the catalytic mutant could not rescue the ts phenotype (Fig 7B) or the silencing defect of dhp1-1 at the centromeric region and the mating type locus (Fig 7C and 7D), indicating that the catalytic activity of Dhp1 is essential for its role in epigenetic silencing.
In S. pombe, epigenetic silencing requires cooperation between the TGS and PTGS pathways [15, 62, 64]. As a classic example of PTGS, RNAi allows processing of RNAs transcribed from these regions to facilitate or reinforce heterochromatin assembly in a RNAPII-dependent manner [15, 64]. In this process, siRNAs maintain the feedback loop and propagate heterochromatin. RNAPII activity is required for generating precursors of siRNA and thereby is crucial for heterochromatin assembly [68, 77]. Additionally, RNAPII may have a more direct role in epigenetic silencing because mutation of RNAPII subunits, splicing factors, and RNA processing machineries impair heterochromatin [68, 77–79].
TGS relies on heterochromatin which is mediated by histone modifications that recruit silencing effectors [4]. In addition to the H3K9 methyltransferase Clr4 and HP1 family proteins, HDACs are critical mediators of all three phases of heterochromatin formation [80–82]. Especially, deletion of class II HDACs clr3 or sirtuin sir2 cause marked reduction of H3K9me across the centromeric regions and mating type locus [80, 82]. To gain further insight into the function of Dhp1 in TGS or PTGS, we compared the localization of Mit1-Myc, one of the core subunits of SHREC (Clr3 complex) at the centromeric regions and the mating type locus in wildtype, dhp1-1, din1Δ, and clr4Δ cells (Fig 8A and 8B). Mit1-Myc is a fully functional allele of Mit1, and has been employed in previous studies [80]. Unlike clr4Δ, which abolishes the localization of Mit1, dhp1-1 does not show any reduction of Mit1 localization at these regions (Fig 8A and 8B), indicating that the localization of SHREC is not reduced. Next, we combined dhp1-1 with either clr3Δ or sir2Δ, and examined the silencing of repeat regions in wildtype, single and double mutant cells (S8 Fig). RT-PCRs show that all double mutant cells have enhanced silencing defects compared to single mutants suggesting overlapping functions between Dhp1 and these HDACs (S8 Fig). These results suggest that Dhp1 likely has a major role in PTGS and acts in a distinct pathway parallel to SHREC-mediated TGS.
To further investigate the role of Dhp1 in TGS or PTGS, we analyzed the relationship between Dhp1 and RNAPII in heterochromatin formation. First, we attempted to combine dhp1-1 with rpb7-G150D, which carries a mutation on the RNAPII subunit Rpb7 and has a specific defect in centromeric pre-siRNA transcription [68]. Surprisingly, combined mutation of dhp1-1 with rbp7-G150D is lethal, suggesting the presence of a compensatory mechanism between Dhp1 and RNAPII to ensure proper regulation of the transcriptome. We next combined dhp1-1 with rpb2-m203, a mutant of the second largest subunit of RNAPII [77]. This mutation does not affect the global transcriptional activity of RNAPII [77]. Instead, it specifically influences the generation of siRNA [77]. Our data indicates that Dhp1 plays a role in a pathway parallel to RNAi in the silencing of repetitive regions (Fig 6). Therefore, the Dhp1-mediated silencing defect is unlikely to be linked through rpb2-m203. Indeed, rpb2-m203 dhp1-1 double mutant cells are viable, and have a stronger silencing defect at the centromere region than either single mutant (S9 Fig), indicating independent, parallel functions in epigenetic silencing.
Heterochromatic regions commonly exclude RNAPII as a mechanism of TGS, but PTGS mechanisms occur downstream of RNAPII recruitment. We wondered whether, like RNAi and the exosome, Dhp1 plays a major role in PTGS. Therefore, RNAPII inclusion or exclusion from chromosomal regions will serve as an indicator to elucidate the function of Dhp1 in transcriptional and/or post-transcriptional actions. We mapped RNAPII occupancy in wildtype and dhp1-1 by ChIP using clr4Δ as a control (Fig 8C and 8D). Loss of Clr4 completely abolishes heterochromatin, thereby shows a strong TGS defect as indicated by dramatically increased RNAPII occupancy at the repetitive regions. However, no difference of RNAPII occupancy was observed between dhp1-1 and the wildtype control at repetitive regions, suggesting the role of Dhp1 is not in TGS, but rather PTGS (Fig 8C and 8D).
Decreased transcription termination demonstrated in dhp1 mutants may reduce the level of available RNAPII complexes for initiation of transcription and could mask the true extent of silencing. Additionally, protracted RNAPII association at a given locus due to stalling might confound ChIP results. To ensure that the true activity of RNAPII was measured, we performed a genome-wide survey of RNAPII targets using Cross-linking and analyses of cDNA (CRAC) in wildtype and dhp1-1 cells (Fig 9). This assay mapped the genome-wide distribution of RNAPII and also monitored the RNAPII complexes actively synthesizing RNAs [83](Fig 9A). A genome-wide study is necessary in this case as Dhp1 may serve distinct roles in euchromatin and heterochromatin, as genome-wide expression profiling suggested (Fig 1 and S3 Fig). In euchromatic regions, defects in terminating RNAPII transcription caused by the dhp1 mutation led to an accumulation of unreleased RNAPII complexes at the 3’end of genes in dhp1-1 (Fig 9B). However, the same phenotype was not observed in clr4Δ cells, indicating that loss of clr4 causes no transcription termination defect. In heterochromatic regions, although clr4Δ dramatically enhanced RNAPII-RNA associations at the centromeric region and mating type locus compared to that of wildtype cells due to complete loss of TGS and partial loss of PTGS, such a difference was not detected upon mutation of dhp1 (Fig 9C). Given the fact that mutation of dhp1 leads to substantial upregulation of repeat transcripts (Fig 1C) without reduction of H3K9me at repetitive regions (Figs 2 and 6), and only marginally affects RNAPII occupancy and its association with repeat transcripts (Figs 8C, 8D and 9), the results support a primary role of Dhp1 in PTGS.
It is well appreciated that RNA processing pathways, including RNAi and the exosome, play crucial roles in the assembly of heterochromatin and elimination of unwanted transcripts [27]. Here we identify a novel RNA processing mechanism mediated by the essential 5’-3’ exoribonuclease Dhp1, which participates in epigenetic silencing in S. pombe independent of both RNAi and the exosome. Interestingly, defective gene silencing at two major heterochromatic regions, the centromeric region and the mating type locus, was observed selectively in dhp1-1 and not in din1Δ or rhn1Δ cells (Fig 1 and S4 Fig). This result suggests that Dhp1 engages in a silencing pathway that is beyond its Din1- or Rhn1-related activity. Our further genetic analyses support a role for Dhp1-mediated silencing in the sequential establishment of epigenetic silencing (Figs 3–5), parallel to RNAi and the exosome (Fig 6), and most likely involved in PTGS via its RNA processing activity (Figs 7–9).
In spite of the crucial role for RNAi in heterochromatin assembly, heterochromatin is not completely abolished in RNAi mutants indicating that other pathways are involved [50, 65, 84]. These pathways are mediated by DNA-binding factors, RNA or RNAi-independent RNA processing factors [50, 61]. In Arabidopsis, the flowering repressor gene FLC is thought to provide links between RNA processing activities and chromatin regulation in gene silencing [85]. In S. pombe, recent studies reveal the nuclear exosome, which governs RNA quality control and ensures the elimination of unwanted RNAs, exists as an RNAi-independent silencing mechanism [42, 43, 50, 61]. Co-activators of the exosome, including TRAMP and MTREC, which help to recognize and degrade its substrates, are also connected to epigenetic silencing without affecting H3K9me, thereby play a major role in PTGS [61, 86]. Additional studies on Triman, a 3’-5’ exonuclease in S. pombe, show that it generates Dicer-independent primal RNAs and is required for initiation of heterochromatin assembly via a mechanism requiring Ago1 [87]. In this study, we described a novel pathway involving Dhp1, a conserved RNA 5’ to 3’ processing enzyme that contributes to PTGS (Fig 10A). We propose that three RNA processing activities, RNAi, the exosome, and Dhp1/Xrn2 degrade repetitive transcripts to mediate the post-transcriptional gene silencing of repeat transcripts (Fig 10B).
Heterochromatin assembly is a dynamic process with distinct steps [4]. It is nucleated at genomic regions containing highly repetitive DNA elements and spread to surrounding regions [88]. Its structure is recaptured during DNA replication and maintained through cell division [14]. Silencing factors often participate at discrete step(s) rather than throughout the process. In particular, RNA-mediated silencing pathways are often required to nucleate heterochromatin formation [9]. Once silencing is established, these factors are dispensable [13]; the heterochromatic state persists in the absence of the initial stimulus. For example, at the mating type locus of S. pombe, RNAi machinery cooperates to nucleate heterochromatin assembly but is dispensable for its inheritance [13]. The re-establishment assay clearly indicates that Dhp1 is indispensable for efficient establishment of silencing at heterochromatic repeat regions (Fig 3). RNAi is well-known as the major nucleation pathway at centromeric regions but not at the mating type locus [64, 65]. Interestingly, dhp1-1 ago1Δ double mutant cells have cumulative defects at the mating type locus indicating separate functions of these two pathways (Fig 6). Unlike Triman, which requires Argonaute to be loaded on longer RNA precursors [87], Dhp1 has an Argonaute-independent role, although we cannot rule out the possibility that the slicer activity of Ago1 may also contribute to the generation of the substrates for Dhp1. Since RNAi itself can initiate heterochromatin formation, we observed re-establishment of heterochromatin at repetitive elements in dhp1-1 cells following clr4+ complementation, although the restoration was not complete (Fig 3). These results suggest that Dhp1 plays a unique but overlapping role in heterochromatin nucleation in concert with RNAi. It is possible that the Dhp1-mediated degradation of heterochromatic repeat transcripts is required for de novo assembly of heterochromatin through recruiting silencing effectors, similar to RITS [36, 38, 39]. It is also possible that the processing activity of Dhp1 is involved in generating the primary small RNAs that contribute to initiation of epigenetic silencing as suggested for the role of the exosome in heterochromatin assembly [61].
In addition to defective nucleation, the H3K9me mark in dhp1 mutants is reduced in the reporter genes embedded at the repetitive regions, suggesting a spreading defect (Fig 2B and 2C). The assay analyzing the spreading of H3K9me from an ectopic nucleation center to the surrounding regions indicates that Dhp1 facilitates the spreading of the heterochromatic mark (Fig 4). Although it is unclear how transcription-mediated spreading of heterochromatin occurs, it is possible that impaired transcription termination in the dhp1 mutant affects the rate of histone turnover during transcription and thereby impedes the spreading of H3K9me.
In addition to its role in initiation and spreading, we provide evidence to show that Dhp1 functions in the maintenance of pre-established silencing (Fig 5). How does Dhp1 function in the maintenance of silencing? It is known that heterochromatin maintenance relies on the binding of Swi6 and Clr4 to methylated H3K9, which facilitates recapitulation of the specific chromatin configuration following DNA replication [38, 70]. In addition, Swi6 and HP1 proteins work as binding platforms, recruiting other histone modifiers and with the factors that are involved in replication-coupled heterochromatin assembly, such as chromatin assembly factor 1(Caf1) [65, 80, 89, 90]. Although the levels of H3K9me2 at the repeat regions are not decreased upon mutation of dhp1, the dynamic binding of Swi6 could still be affected. In addition to H3K9me, Swi6 is also reported to bind “repellent” RNAs that antagonize the heterochromatic silencing [91]. Thus, Dhp1-mediated elimination of RNA may facilitate the dynamic binding of Swi6 to heterochromatin, and thereby ensure the maintenance of the silenced chromatic domains.
Although the centromeric regions and the mating type locus are assembled by heterochromatin, they occupy different chromosomal contexts and use distinct strategies to target heterochromatin [4]. Notably, the effects in double mutants of dhp1 with ago1Δ or rrp6Δ are different at centromeres and the mating type region (Fig 6). At the centromeric region, RNAi is the major pathway [64]. Therefore, as expected, the ago1Δ single mutant exhibits a severe silencing defect and decreased H3K9me mark (Fig 6 and S7 Fig). Compared to the already radically impaired silencing phenotype in the ago1Δ single mutant, the dhp1-1 ago1Δ double mutant shows even higher levels of repeat transcripts and lower levels of H3K9me2 (Fig 6), suggesting that transcripts produced from repeat regions in RNAi-deficient cells, are likely targets of Dhp1. In contrast, without Rrp6/exosome, RNAi machinery is still functional. Therefore, we only observe a moderate silencing defect at the centromeric region in the dhp1-1 rrp6Δ double mutant (Fig 6). At the mating typing locus, at least three pathways initiate heterochromatin assembly and target H3K9me [4, 27]. It is not surprising that the dhp1-1 ago1Δ double mutant maintains a high level of H3K9me2 (Fig 6); other pathways may compensate for loss of function for both Dhp1 and RNAi at the mating type locus [4]. In addition, transcription termination defects caused by rrp6 and dhp1 mutation may contribute to the increased level of H3K9me seen at both centromere and mating type locus (S7 Fig). Interestingly, cells containing the dhp1 mutation consistently show a stronger silencing defect at the mating type locus even in the presence of higher levels of H3K9me (Figs 1–3 and 6 and S7 Fig), suggesting that Dhp1-mediated silencing occurs primarily downstream of H3K9me, likely as a mechanism of PTGS.
In S. pombe, TGS and PTGS are intertwined. In TGS, heterochromatin greatly limits the access of RNAPII, allowing only a low level of transcription from highly repetitive DNA regions. RNAs transcribed from these regions are subject to PTGS by RNAi machinery, in which they are processed into siRNAs in order to feedback on chromatin to facilitate the assembly and propagation of heterochromatin [27, 88]. The silencing defect in dhp1-1 is unexpected considering that compromised transcription termination would weaken RNAPII transcription and delay the release of RNA from the site of transcription, which may then enhance the assembly of heterochromatin mediated by RNA as suggested by previous studies [48, 50]. To elucidate whether Dhp1 plays a major role in TGS, we used ChIP analysis to map H3K9me and SHREC (Figs 2, 8A and 8B), which have well-studied functions in TGS at repeat regions. If Dhp1 plays a role in TGS, we would expect to observe reduced enrichment of H3K9me and SHREC at the endogenous repetitive regions in dhp1-1. No reduction of enrichment occurred however for either H3K9me2 or SHREC at endogenous repetitive regions in dhp1 mutants, suggesting that the role of Dhp1 in gene silencing is primarily associated with PTGS rather than TGS (Figs 2, 8A and 8B). We further investigated RNAPII occupancy and the levels of actively transcribing RNAPII at repeat regions in wildtype and mutant cells using ChIP and CRAC respectively (Figs 8C, 8D and 9). A role in TGS for Dhp1 would be suggested by increased RNAPII occupancy occurring at repetitive regions in dhp1-1, as RNAPII in the context of impaired TGS would associate more frequently with heterochromatic transcripts. In contrast, no increase would implicate a role for Dhp1 in PTGS. Our RNAPII ChIP results clearly show no difference between dhp1-1 and wildtype, implicating a PTGS role for Dhp1 (Figs 8C and 8D). Additionally, we showed that the catalytic activity of Dhp1 is required for its role in epigenetic silencing, providing strong evidence to support that the RNA processing role of Dhp1 is associated with PTGS (Fig 7). Although our results pinpoint the primary role of Dhp1 in epigenetic silencing through PTGS, completely discounting a function of Dhp1 in TGS is a challenge as Dhp1/Rat1/Xrn2 has well-established activity that is linked to RNAPII.
RNAPII transcription and its associated activities are required for heterochromatin assembly. As a result, loss of silencing was reported to correlate with defective RNAPII transcription [68, 77, 92–94]. Is the RNAPII-linked function of Dhp1 related to epigenetic silencing? In agreement with reported termination defects upon mutation of Dhp1 and Din1, our expression profiling showed accumulation of 3’ untranslated transcripts at many genes in these mutants (S3 Fig). To execute its function in RNAPII transcription termination, Dhp1/Rat1 exonucleases target the downstream fragments produced by cleavage at the polyA site during 3’ end processing [44–46]. The processed mRNAs are packed into nuclear RNA transporting cargos and exported to the cytoplasm for translation. Since this action of Dhp1/Rat1 in transcription termination lies downstream of mRNA processing and packaging, defects of Dhp1/Rat1 are unlikely to dramatically influence the amount and the quality of coding mRNAs. Indeed, at least at the permissive temperature, we did not observe significant alterations of coding transcripts in the dhp1 mutant (S2 Table). Rather, the remarkable differences observed in the transcriptome were seen at non-coding regions (S3 Fig). Recently, Rat1 in budding yeast was reported to maintain the balance of RNAPII CTD phosphorylation, and therefore plays a role in transcription elongation [95]. This finding suggests that Rat1 may have more complex roles in transcription than previously thought. In addition, neither loss of Rnh1 nor Din1 causes growth defects or silencing defects as seen in the dhp1 mutant (Figs 1 and 2 and S4 Fig), raising the question about which role of Dhp1, transcription or RNA quality control, is essential for cell growth and silencing. In this study, we indeed observed higher enrichment of H3K9me2 at the endogenous repetitive regions in the dhp1 mutant (Figs 2 and 6). This observation is in agreement with a study showing that an impaired Paf1 complex is sufficient to induce RNAi-mediated epigenetic silencing in trans at euchromatic loci, likely through its termination defect [48]. While a parallel study reported significant reductions of H3K9me2 in dhp1 mutants at all major heterochromatic regions [59], we only observed reduced H3K9me at reporter genes but not at the repeat regions. It is likely that the discrepancies in the H3K9 methylation data are due to differences in culturing conditions. To minimize the pleiotropic impacts caused by dhp1 mutation and avoid the antagonistic effect of high temperature (37°C) for heterochromatin formation [96], we collected data at a permissive condition (30°C) without shifting cell cultures to 37°C, the restrictive condition applied in the parallel study [59]. Therefore, our results are more likely to accurately represent the true effect of Dhp1 in epigenetic silencing. In addition, we provided evidence to show that Dhp1-mediated silencing is independent of RNAi (Fig 6). Overall, the role of Dhp1 in epigenetic silencing at major heterochromatic regions cannot be explained by its known function in transcriptional termination. It is possible that RNAPII may couple repeat transcription with its degradation by Dhp1. A second possibility is that RNAPII may help “discriminate” noncoding pericentromeric repeat RNAs from general pre-mRNAs so that the former can be degraded by Dhp1. The basis for this selection may be the aberrant (double-stranded or abnormally capped) structure of the transcribed RNA. Alternatively, the chromatin structure of the transcribed repeat region may somehow determine the fate of the transcripts, feeding into RNAi-, exosome-, or Dhp1-mediated silencing.
The catalytic activity of Dhp1 is required for its role in epigenetic silencing (Fig 7). By what mechanism are the substrates for Dhp1-mediated silencing produced? Due to the strong conservation of the active site, it is likely that the mechanisms of Xrns are very similar [97]. The crystal structure of Drosophila XRN1 indicates that substrates are limited to 5’ monophosphate RNAs because larger structures, such as m7G Cap or triphosphorylated RNAs, do not fit into the pocket [98]. Hence, the RNA pyrophosphohydrolase activity of Din1 seems necessary for the generation of monophosphorylated RNA substrates for Dhp1, especially for decapping and RNA quality control. However, only Dhp1, not Din1, is essential for viability and epigenetic silencing. In addition, unlike Dhp1, Din1 and its orthologs are not widely conserved [47, 99]. Since Din1 is not essential and is not necessary for epigenetic silencing, an endoribonuclease or an extra RNA pyrophosphohydrolase likely produces the substrates for Dhp1 in silencing.
In yeast, abnormal pre-mRNAs are degraded rapidly from both 5’ and 3’ ends by Rat1/Xrn2 and the nuclear exosome, respectively, with the exosome playing a dominant role [100]. In human cells, XRN2 appears to be more crucial for degradation of abnormal pre-mRNAs than the exosome [101]. Given the fact that Xrn2 is conserved from yeast to humans, our results may yield insights broadly applicable to the gene silencing field, including mammals. Dhp1/Xrn2 may represent a more generalized mechanism of an RNA-based form of silencing. Future studies identifying additional Dhp1/Xrn2 interacting proteins may help to address these questions.
S. pombe strains used in this study are listed in S1 Table. Cells were cultured using standard procedures for growth and manipulation [102]. Epitope-tagged and deletion mutant strains were engineered using standard PCR methods as described previously [103]. Double mutants were constructed via genetic crossing followed by tetrad dissection. For dilution assay, liquid cultures were diluted in series (1:10) and plated using a pin transfer tool on YEA media (Rich, N/S), low adenine YE media, or YEA media containing either 20 μg/ml TBZ or 850 μg/ml 5-FoA. All cultures were grown at 30°C (or 37°C where indicated). The strains used for cross-linking and analyses of cDNA (CRAC) carry a carboxyl-terminal HTP-tagged subunit of RNAPII, Rpb2-HTP. An HTP tag contains a 6X- His epitope and a protein A epitope separated by a Tobacco Etch Virus (TEV) protease cleavage site [83]. Strains carrying either KΔ∷ade6+off or KΔ∷ade6+on were isolated and saved as previously described [71].
To generate pdhp1+, a PCR fragment amplified using oligos Dhp1-BamHI-Fw and Dhp1-Pst1-RV contains a wildtype dhp1+ gene including promoter, open reading frame, and 5’ and 3’ untranslated regions. The PCR fragment was digested by BamHI and PstI and ligated into a pREP41 digested with the same restriction enzymes (BamHI/PstI). After BamHI/PstI digestion, pREP41 lost its nmt promoter. The resulting pdhp1+ expresses the wildtype dhp1+ driven by its endogenous promoter. pdhp1D55A E207Q was generated using a QuickChange Site-Directed mutagenesis kit (Stratagene) based on pdhp1+. pclr4+ is a plasmid carrying a DNA fragment containing a wildtype clr4+ driven by its endogenous promoter as previously described [50].
Total RNA was prepared using the MasterPure Yeast RNA Purification Kit (Epicentre). First-strand cDNA was produced with M-MLV Reverse Transcriptase (Promega) using site-specific primers following manufacturer protocols. Real-time PCR was performed on a 7500 Fast Real-Time PCR System (Applied Biosystems) with SYBR Select Master Mix (Applied Biosystems). First-strand cDNA synthesis without reverse transcriptase was performed for negative controls. At least two biological repeats were performed for all experiments. Statistical analysis was performed using a student’s t test (two-tailed distribution). Error bars represent standard error of mean (s.e.m). Primers are listed in the S4 Table.
Mating-type switching-competent (h90) mid-log phase cells (wildtype, dhp1-1, or din1Δ) were plated on solid sporulation medium (SPA). Cells grew at 30°C for 6 hr, then switched to 37°C for 2 hr, and finally finished the sporulation at 30°C for 12hr. Cells were washed 3X with water. Ten microliter cells in water were spread on a glass slide, and fixed by heat at 70°C. The slides were then covered by 5μl of mounting buffer with DAPI (VECTOR, H1500) and 13mm coverslips. The stained cells were imaged by a confocal microscope.
Sample preparation for the expression array and array design were reported previously [104]. The expression profiling is performed as previously described [105]. The composite plot was generated using GenomicRanges R-package (version 1.20.5), from the high-resolution part of the microarray (2320 genes). The genes were aligned at the transcriptional termination site TTS (S. pombe 2007_April annotation) and the geometric means of the ratios (Mutant/wt) were plotted.
Flag-TEV-protein A (FTP)-tagged purification and mass spectrometry were performed as previously described [105].
ChIP experiments were performed as described previously using antibodies against histone H3 (di-methyl K9) (Abcam,Ab1220), RNAPII (Abcam, Hab5408), or Myc (Santa Cruz, A-14) [106]. Real-time PCR was performed on a 7500 Fast Real-time PCR System (Applied Biosystems) with SYBR Select Master Mix (Applied Biosystems). At least two biological repeats were performed for all ChIP experiments. Statistical analysis was performed using a Student’s t test (two-tailed distribution). Error bars represent s.e.m.
In vivo CRAC was performed as described with modifications [83]. Two-liter yeast cultures were grown to an OD600≈2 at 30°C. Cells were harvested by centrifugation and cell pellets were resuspended in 2.5L Phosphate-Buffered Saline (PBS) followed by UV-irradiation in a “Megatron” UV-cross-linker (254 nm) for 3 min before cells were pelleted and frozen in liquid nitrogen. The pellets were then lysed by grinding in liquid nitrogen (Resch, MM400) and resuspended in 10 ml of 1x TN150 lysis buffer (10x TN150: 0.5 M Tris-HCl (pH 7.8), 1.5 M NaCl, 1% NP-40). Extracts were clarified by centrifugation (10 min at 4000 rpm and 45 min at 15,000 rpm at 4°C) and incubated with 150 μl of equilibrated IgG Sepharose beads (GE Healthcare) for 1h at 4°C. After two washes with TN1000 buffer (100 mM Tris-HCl (pH 7.8), 2 M NaCl, 0.2% NP-40) and two washes with TN150 lysis buffer, the beads were incubated with GST-TEV protease for 2h at 16°C. The TEV eluates were collected by centrifugation and incubated with 10U of Turbo DNase (Ambion) for 8 min at 37°C followed by incubation with RNase Cocktail Enzyme Mix (Ambion; 0.005 U RnaseA, 0.2 U Rnase T1) for 2 min at 37°C. Guanidine-HCl (0.4g) was dissolved in 500 μl of TEV eluates. NaCl and Imidazole were added to final concentrations of 300 mM and 10 mM, respectively. Samples were incubated with 50 μl of nickel agarose beads (Macherey-Nagel) over night at 4°C. All washes, alkaline phosphatase treatment and 3’ linker ligation were carried out as described except that 40U T4 RNA ligase 2 truncated K227Q (NEB) was used instead of T4 RNA ligase. The beads were incubated in 80 μl phosphorylation mix (16 μl 5x PNK buffer (250 mM Tris-HCl (pH 7.8), 50 mM MgCl2, 50 mM β-mercaptoethanol), 200 mM ATP (Sigma, A6559), 20U T4 polynucleotide kinase (NEB), 80U RNase Inhibitor) for 40 min at 37°C. For the ligation of the 5’ linker the beads were resuspended in 80 μl of 5’ ligation mix (16 μl 5x PNK buffer, 80U RNasin, 40U T4 RNA ligase, 100 pmol 5’linker, and 80 mM ATP) and incubated at 16°C. After two washes with wash buffer II (50 mM Tris-HCl (pH 7.8), 50 mM NaCl, 10 mM Imidazole, 0.1% NP-40) the material was eluted with elution buffer (10 mM Tris (pH 7.8), 50 mM NaCl, 150 mM Imidazole and 0.1% NP-40). The final eluate was incubated with 2 M EDTA, 20 μl 20% SDS and 100 μg proteinase K (Ambion AM2548) for 2 hours at 50°C and the RNA was extracted using Phenol-Chloroform followed by ethanol precipitation. Reverse transcription with SuperScript III was performed following the manufacturer’s instructions (Invitrogen) followed by RNase H (NEB) digest (10U) for 30 min at 37°C. The cDNA was amplified and the PCR-product was purified with the Agencourt AMPure XP PCR purification beads (Beckman Coulter) following the manufacturer’s instructions. The quality of the library was verified with the Bioanalyzer 2100 (Agilent) and the Agilent High Sensitivity DNA Kit (Agilent). The amplified library was subject to high-throughput sequencing at BGI-Hong Kong Co. Ltd. The datasets were mapped to the mating type locus of S. pombe (41249 bp from chromosome 2) and to the full genome of S. pombe (ASM294v1.17) using Tophat2 software. (Tophat-2.0.14; [107]). Downstream data analysis was performed using R Bioconductor packages.
The mean coverage over mRNA loci was normalized to 20 in the datasets. The difference plot was generated from all protein coding ORFs (5115 genes), aligning them at the annotated transcriptional termination site (TTS) (S. pombe EF2 annotation). The plot is showing the ratio between the normalized Mutant/wt coverage. Two biological duplicates have been performed for genome-wide analysis for wildtype and dhp1-1 strains. All microarray and CRAC data sets are available at NCBI GSE77291, GSE77289 and GSE77290.
|
10.1371/journal.pntd.0006393 | A novel immunotherapy of Brucellosis in cows monitored non invasively through a specific biomarker | Brucellosis is an important zoonotic disease causing huge economic losses worldwide. Currently no effective immunotherapy for Brucellosis or any biomarker to monitor the efficacy of therapy is available. Treatment is ineffective and animals remain carrier lifelong. S19 and RB51 are live attenuated vaccine strains of Brucella abortus. However, S19 induces only antibody, ineffective for intracellular pathogen. RB51 induces cell mediated immunity (CMI) but it is Rifampicin resistant. Both organisms are secreted in milk and can infect humans and cause abortions in animals. Phage lysed bacteria (lysates) retain maximum immunogenicity as opposed to killing by heat or chemicals. We report here the successful immunotherapy of bovine Brucellosis by phage lysates of RB51 (RL) and S19 (SL). The SL induced strong antibody response and RL stimulated CMI. In vitro restimulation of leukocytes from RL immunized cattle induced interferon gamma production. A single subcutaneous dose of 2 ml of cocktail lysate (both RL and SL), eliminated live virulent Brucella from Brucellosis affected cattle with plasma level of Brucella specific 223 bp amplicon undetectable by RT-PCR and blood negative for live Brucella by culture in 3 months post-immunization. This is the first report on minimally invasive monitoring of the efficacy of antibacterial therapy employing plasma RNA specific for live bacteria as a biomarker as well as on the use of RB51 phage lysate for successful immunotherapy of Brucellosis in cattle.
| We report here a novel immunotherapy for cost effective and simple treatment of Brucellosis, a zoonotic disease of global importance. The two main novelties of this work which have important implications are the following: Plasma RNA (specific to Brucella abortus in this case) has been used as a biomarker for the first time to successfully monitor the efficacy of an antibacterial therapy (in this case, immunotherapy of brucellosis). There is no published report of such a biomarker for monitoring the therapy in Brucellosis or any other bacterial infection. Bacterial counts in spleen and other organs before and after phage therapy (after sacrificing) have been used in mice. However, this is not applicable in cattle and humans. Therefore, this novel application of a specific biomarker for periodic minimally invasive and non destructive monitoring of the efficacy of antibacterial therapy could be very useful in Medicine. The phage lysate of RB51 strain of Brucella abortus has been successfully used for immunotherapy of Brucellosis in cattle for the first time. Although S19 lysate has been used earlier in mice for prophylaxis, it induces only antibody and not cell mediated immune response which is essential for protection against intracellular pathogens like Brucella.
| Brucellosis is a major re-emerging bacterial zoonosis of global importance affecting a range of different animal species and man worldwide and is of importance from economic and public health points of view. It remains an uncontrolled problem in regions of high endemicity such as Africa, the Mediterranean, Middle East, parts of Asia and Latin America[1]. Outbreaks of bovine Brucellosis caused by Brucella abortus are associated with abortion in the last trimester of gestation, delayed conception, temporary or permanent infertility in the affected animals, and weak newborn calves. It causes infertility in both sexes and adult male cattle develop orchitis. Hygromas involving leg joints are a common manifestation of Brucellosis. Once infected, the animal may become carrier and continue to shed the organism for long period[2].
Brucella abortus S19 and RB51 strains have been used to control bovine Brucellosis worldwide. However, S19 vaccination in adult animals has been shown to result in abortions and arthopathy in some cases. Antibodies induced by vaccination also interfere in serological diagnosis of the disease[3]. RB51 does not induce antibodies against smooth lipopolysaccharide detectable by routine serological tests and is responsible for inducing cell mediated immunity[4]. However, the RB51 organisms are Rifampicin resistant[5]. Both these vaccines are secreted in milk and can infect humans on direct contact and cause abortions in pregnant animals. An appropriate antibiotic therapy for animals and human beings is still disputed and would be too expensive in most of the animal species. Despite the availability of the two vaccines for bovines, the search for improved vaccines has continued[6].
Lytic bacteriophages quickly reproduce and grow exponentially in the viable bacterial cells and lyse the bacteria specifically without damaging the normal flora[7]. Phage lysed bacteria induce effective immunization and exhibit protective effect stronger than the vaccines produced by heat or chemical inactivation of bacteria[8, 9, 10]. Phage lysates are non toxic and immunization causes no known adverse reactions.
We explored the immunotherapeutic potential of phage lysates of S19 and RB51 in Brucellosis in adult cattle and employed Brucella specific RNA in plasma as a biomarker to monitor the effect of therapy in the live animal in a non destructive and minimally invasive manner.
Live organisms from the attenuated vaccines Brucella abortus—strain 19 and strain RB51 (Indian Immunologicals, Hyderabad) were used for making phage lysates (immunotherapeutic agent). The identity of strain 19 and strain RB51 organisms was confirmed by microscopy and biochemical tests viz. catalase, oxidase, urease tests and H2S production as per the standard methods [11]. The bacterial cultures were maintained on Brucella agar plates, Trypticase Soy Agar (TSA) and Farrell’s medium and slants by serial sub-culturing in Brucella selective broth on every fortnight and storing the cultures at 4°C.
A broad acting phage lytic to Brucella organisms isolated in our laboratory [12, 13] was used for lysing Brucella abortus strains 19 and RB51 for making lysates. Phage as a crude, concentrated suspension prepared in SM diluent, was first revived by agar overlay technique [14]. The procedure reported earlier[13] was followed. The phage preparation obtained as mentioned above was amplified to 250 ml master lot using the liquid culture method as described earlier [15].
Serial 10 fold dilutions of the purified phage stock up to 10−12 were prepared in sterile SM buffer. Equal quantity of each phage dilution was mixed in 3 ml semisolid NZCYM agar at 47°C and seeded with the log phase cultures of Brucella abortus strain S19 and Brucella abortus strain RB51. After mixing properly, the soft agar in each tube was plated on hard Brucella agar plates and left for solidifying. All the plates were then incubated at 37°C for 48 hours. The plaques produced were counted and phage count was determined. The phage titer was expressed in pfu/ml after multiplying with the dilution factor.
The indicator strains Brucella abortus S19 and Brucella abortus RB51 were grown in about 100 ml NZCYM broth at 37°C (log phase). The bacterial cells were harvested in fresh sterile NZCYM broth aseptically after centrifugation at 5000 rpm for 20 min. The suspension was standardized to match with McFarland standard 3 (approx. 108 viable cells/ml) and distributed aseptically in 5ml aliquots in six sterilized test tubes in racks. Phage was added to make final phage: bacteria ratios of 1:104; 1:5x103; 1:103, 1:5x102; 1:102 and 1:50, respectively and were mixed thoroughly by vortexing and incubated at 37°C.100μl aliquots from each tube were taken aseptically at 20, 180, 240 and 300 min post incubation, mixed with melted NZCYM soft agar and spread on NZCYM+BSM agar plates after making appropriate dilution with 0.85% Normal Saline Solution (NSS). Plates were incubated at 37°C and colony count was conducted. The optimum phage-bacteria ratio that showed complete lysis of indicator strains Brucella abortus S19 and RB51 within shortest period of time was considered as Multiplicity of Infection(MOI) of the phage for all future uses.
The method for generation of phage lysate preparations against Brucella abortus using Brucellaphage reported earlier [15] was suitably modified and used for making lysates of both S19 and RB51 [16]. The total viable count (TVC) of 24–48 hr Brucella selective broth culture of the strains S19 and RB51 incubated at 37°C was adjusted to 2 x 108 cfu/ml. Opacity of broth was adjusted to Mac Farland’s tube no 3 to obtain growth sufficient for appropriate antigenic biomass. Phage was added as per optimized MOI and TVC of the respective indicator strains (with phage bacteria ratio of 1: 50) and the mixture was further incubated for 6–7 hours at 37°C for complete lysis and clearance of turbidity. 100 ml each of phage lysate was prepared against both the strains (Brucella abortus S 19 and RB 51).
The phage lysate cocktail was then passed through a 0.1 μm filter (Pall Life Science) to separate out the phage from the lysate and the filtrate was stored in sterilized vials at 4°C. Sterile aluminium hydroxide gel suspension in saline in ratio of 1:10 was added to the preparation. Total protein content of the phage lysate, as determined by Nanodrop spectrophotometer, was 0.58 mg/ml for S19 lysate and 0.64 mg/ml for RB 51 lysate. The phage lysate cocktail (combination of Brucella abortus S19 and RB51 lysates) was prepared by addition of equal quantity of S19 phage lysate and RB51 lysate. Sterilized 1% Aluminium hydroxide gel suspension in saline was mixed aseptically with the test preparations in ratio of 1:10 (final Aluminium concentration 0.1%) and incubated at 37°C for 24 hours and then stored at 4°C. The phage lysates were then subjected to sterility and safety tests.
A loopful of the lysate was suspended in 5ml BHI and BSM broth as well as streaked on BHI and BSM, Trypticase Soy Agar followed by incubation at 37°C. The broth and plates were examined up to 48 hours for any microbial growth.
All the experimental protocols performed on mice and cattle were approved by the Institutional Animal Ethics Committee (IAEC). Proposals GADVASU/2016/IAEC/33/09 and GADVASU/2016/IAEC/33/10 were approved by the Institutional Animal Ethics Committee (IAEC) in its XXXIV meeting held on 2nd July, 2016 communicated vide its proceedings number IAEC/2016/481-509 dated 8th July, 2016.
Animals were kept in approved facilities with adequate feed, water, bedding, pucca flooring, covered space and light and all the methods were performed in accordance with the guidelines and regulations of the Committee for the Purpose of Control and Supervision of Experiments on Animals (CPCSEA), Animal Welfare Division, Ministry of Environment, Forests and Climate Change.
Safety test of all the preparations of lysates was conducted in mice as recommended in Indian Pharmacopoeia [17] before commencement of immunization of cattle. Three groups of three adult healthy mice, each were injected with 0.2 ml volume of lysate (S19 or RB51 or cocktail of both) by subcutaneous route. One group of three adult healthy mice was left untreated as control. The mice were observed for any untoward reaction or mortality up to 7th day of inoculation.
Trials of phage lysate therapy were carried out on 21 naturally Brucellosis affected adult cows. The Brucellosis positive adult cows were divided into four groups:
Group I animals (n = 6) were immunized with Brucella abortus S19 lysate only, Group II animals (n = 5) were immunized with Brucella abortus RB51 lysate only, Group III animals (n = 5) were immunized with cocktail of S19 and RB51 lysates. A dose of 2 ml lysate was administered through subcutaneous route whereas Group IV (n = 5) served as a control, and received no immunization.
Blood samples from cattle were collected from the jugular vein at 0 day and at 30, 60, 75 and 90 days post treatment for studying the immune response of the animals and for detection of nucleic acid specific to the Brucella organisms. Sera and plasma were separated from blood and stored at -20°C and -80°C, respectively till further use.
The serum samples of cattle were tested using INGEZIM Brucella Bovina 2.0 serum ELISA kit. The kit is based on the indirect immunoenzymatic assay. The antigen used is a purified extract of the LPS of Brucella. The absence or presence of antibody in sera was determined by colorimetric reaction. The degree of the colour that develops (Optical Density measured at 450 nm) is directly proportional to the amount of antibody specific to B. abortus present in the sera of animals. The diagnostic relevance of the result is obtained by comparing the optical density (OD) that develops in wells containing samples with the O. D. from wells containing the positive control.
The test was considered valid when:
O.D.valueofthepositivecontrolserumwas>1.0O.D.valueofthenegativecontrolserumwas<0.2
Positive and Negative cut offs were calculated as follows:
Cutoff=Absorbanceat450nmofpositivecontrolx0.4=40percentofPositivity
The Total Leukocyte Count (TLC) of the blood samples collected from cattle at 0, 30, 60, 75 and 90 days post vaccination was determined by automatic Hematology Analyzer. Differential Leukocyte Count (DLC) was carried out by making smears on clean, grease—free glass slides, staining with Leishman’s stain and examining under the microscope.
Anticoagulated blood of lysate treated cattle was centrifuged at 2000–2500 rpm and leukocytes present in the buffy coat layer of the centrifuged blood were separated. Cells were counted on Neubauer’s Hemocytometer and adjusted to 2x105 cells/ml, and were stimulated in vitro by phage lysate preparations. Bovine IFN-gamma ELISpot Kit (Essence Life Sciences) was used for the assay. The assay was carried out as per instructions supplied with kit. The developed microplate was analysed by counting spots visually using a microscope. Quantitation of results was done by calculating the number of spot forming cells (SFC) per number of cells added to each well.
Isolation of total RNA from the plasma samples of cattle was done as per the method reported earlier [18] with suitable modifications using RNA isolation kit TRI reagent BD (Sigma Aldrich). The modification in procedure was that equal volume of RNase-free isopropanol was mixed with the aqueous phase and kept at -80°C overnight, to allow sufficient precipitation of RNA[19].
The optical density of nucleic acid preparations was measured in an ultraviolet ray Nanodrop spectrophotometer. For quantification of the amount of RNA, readings were taken at wavelengths of 260 nm and 280 nm. The readings at 260 nm allow the calculations of the concentration of nucleic acid in the sample.
Pure preparations of RNA with OD260/280 ratio ranging between 2.0–2.2 were selected. The concentration of total RNA varied between 500–2200 ng/μL in various samples. The amount of RNA used for cDNA synthesis was 2 μg for each sample.
The RNA isolated from the Brucella abortus strain 19 organism from the live attenuated vaccine Bruvax (Indian Immunologicals) was used as a positive control for the RT-PCR detection of Brucella abortus.
Total RNA extracted was reverse transcribed into cDNA using high capacity cDNA reverse transcription kit (Applied Biosystems) as per the method reported earlier [20].
The RT-PCR assay for the amplification of the genes encoding 31 kDa B. abortus antigen was carried out by using Brucella genus specific primers B4/B5 reported elsewhere [21]which were got synthesized from Promega.
Sequences of the primers used for detection of B. abortus are given below:
B4F—5’-TGG CTC GGT TGC CAA TAT CAA-3’
B5 R—5’-CGC GCT TGC CTT TCA GGT CTG-3’
The size of amplified product was 223 bp.
In the present study, cDNA was prepared from RNA extracted from plasma samples which were then subjected to PCR amplification of B4/B5 gene. The method described earlier [22] was followed.
Data pertaining to serum antibody titers by ELISA, TLC, DLC and plasma RNA concentration in peripheral blood were statistically analyzed by ANOVA and Tukey’s post-hoc test was performed.
The present study was undertaken to explore the potential of bacteriophage lysates of Brucella organisms as an immunotherapy for Brucellosis in cattle.
Brucella abortus live attenuated vaccine strains S19 and RB51 (Indian Immunologicals, Pune) were used for preparation of phage lysates. The Brucella organisms were cultured on Brucella selective agar. Identification of the S19 and RB51 organisms was done on the basis of cultural and staining characteristics and biochemical properties.
A phage lytic to Brucella organisms isolated earlier[13]was used in the present study. Brucella abortus strains S19 and RB51 were used for revival and propagation of phage. The plates having plaque formation were confirmed for the presence of phage by secondary streaking.
The brucellaphage lysed Brucella abortus strain S 19, and Brucella abortus strain RB 51 but did not lyse any heterologous species tested viz. Salmonella species, Escherichia coli and Pasturella multocida.
The optimum phage bacteria ratio to achieve maximum lysis of the indicator strains (Brucella abortus strain 19 and RB51) within the shortest period of incubation was found out to be 1:50. PFU/ml of the phage calculated by using several dilutions of the phage in SM buffer followed by plating (agar overlay technique) was 2.4 x 108 Pfu/ml. At higher dilutions the plaques were countable but the size of the plaques decreased appreciably. No reduction in counts was observed during the entire period of investigation.
The phage lysates were subjected to sterility tests and were found to be bacteriologically sterile when tested on BSM agar plates and free from any fungal contamination.
Safety tests of the preparations were conducted in mice as per the recommended protocol[17]. The mice were inoculated with lysates subcutaneously. All the mice survived and there was no untoward reaction or toxic side effects or any abnormal lesions during the course of observation period of seven days. The preparations were therefore considered to be safe in mice.
The present study was undertaken to explore the therapeutic potential of phage lysate of Brucella abortus vaccine strains S19 and RB51. Brucellaphage which was isolated earlier [12, 13] was used to prepare the phage lysates. The immunotherapeutic agents (phage lysates of Brucella abortus strain 19 and RB51) were injected only once in the Brucellosis affected cattle at a dose of 2 ml subcutaneously. The controls were left untreated. Pre (0 day) and post treatment (30, 60, 75 and 90 days) sera and plasma samples were collected and stored at -20°C and -80°C respectively till use. Titers of antibody induced by the phage lysates were estimated by indirect ELISA and cellular immune response was studied by evaluating the Total Leukocyte Count and Differential Leukocyte Count (TLC and DLC) in the immunized animals. ELISPOT assay was conducted to investigate γ- interferon production by the lymphocytes of the immunized animals.
Antibody titers were monitored by ELISA pre- immunization (0 day) and at various intervals post-immunization in Brucellosis affected cattle immunized with phage lysates and in untreated infected cattle(Fig 1).
Antibody titers at 0 day and various intervals post-immunization in Brucellosis affected cattle treated with S19 lysate, RB51 lysate, or cocktail (S19+RB51) lysate as well as affected untreated cattle are shown in Table 1.
ANOVA and Tukey’s post-hoc test revealed that in case of Brucella abortus S 19 lysate immunized cattle, there was a significant (P<0.01) increase in the mean titers from 0 d to 75 day, 0 to 90 day, 30 to 75 day, 30 and 90 day and between 60 and 90 day. In case of Brucella abortus RB51 lysate immunized cattle, the mean titers varied significantly (P<0.01) between 0 to 60 day, 0 to 75 day, and 0 to 90 day. In case of cocktail (S19+RB51) lysate immunized cattle, there was a significant (P<0.01) increase in the titers from 0D to 30D, 60D, 75D, and 90D, from 30D to 75D and 90D, from 60D to 75D and 90D and from 75D to 90 days. In case of untreated cattle affected with Brucellosis the difference in the mean titers was non-significant.
The results indicate that the antibody titers enhanced considerably by Brucella abortus S19 lysate. Interestingly, Brucella abortus RB51 lysate immunization contributed to an enhancement in titer at a later stage, whereas cocktail lysate generated a robust antibody response which consistently increased significantly throughout the 3 months period of observation.
The status of the cellular immunity in lysate immunized animals was investigated by assessing Total Leukocyte Count (TLC) and Differential Leukocyte Count (DLC) in peripheral blood of the infected animals at various intervals. The results are summarized in Fig 2.
The total leukocyte counts (x103/μL) at 0 day and various intervals post-immunization in Brucellosis affected cattle treated with S19 lysate, RB51 lysate, or cocktail (S19+RB51) lysate as well as affected untreated cattle are shown in Table 2.
ANOVA and Tukey’s post-hoc test revealed that in case of Brucellosis affected cattle immunized with Brucella abortus S19 lysate, the leukocyte counts increased significantly (P<0.01) from 0 day to 90 day; 30 day to 90 day and 60 day to 90 day, respectively and significantly (P<0.05) from 75 day to 90 day post immunization. In case of animals immunized with Brucella abortus RB51 lysate, it was observed that there was no significant difference in the total leukocyte counts between the various intervals. In animals immunized with cocktail lysate preparation (Brucella abortus S19 and RB51), there was a significant difference (P<0.05) in the total leukocyte count from 60 day to 90 day and 0 day to 90 day post-immunization. The differences among the mean values of 30D vs 90D and 60D vs 90D was significant (P<0.01). It was observed that the leukocyte counts in the Brucellosis infected untreated cattle did not vary significantly between 0 to 90 days.
The lymphocyte counts (percent) at 0 day and various intervals post-immunization in Brucellosis affected cattle treated with S19 lysate, RB51 lysate, or cocktail (S19+RB51) lysate as well as affected untreated cattle are shown in Table 3.
ANOVA and Tukey’s post-hoc test revealed that in case of untreated cattle the increase in lymphocyte counts (%) was significant (P<0.05) from 0 day to 90 day; 30 day to 90 day and 75 day to 90 day and the increase in the lymphocyte counts from 60 day to 90 day was significant (P<0.01). The variation in lymphocyte counts (%) in animals treated with Brucella abortus S19 lysate or cocktail (combination of Brucella abortus S19 and RB 51) lysate was non-significant at various intervals. However, in case of RB51 lysate immunized animals the increase in lymphocyte counts from 0 day to 30 day and 60 day to 90 day post immunization was significant (P<0.05). The increase in counts from 0 day to 60 day, 75 day and 90 day, 30 to 75 day and 90 day was significant (P<0.01).
The neutrophil counts (percent) at 0 day and various intervals post-immunization in Brucellosis affected cattle treated with S19 lysate, RB51 lysate, or cocktail (S19+RB51) lysate as well as affected untreated cattle are shown in Table 4.
ANOVA and Tukey’s post-hoc test revealed that in animals which were treated with Brucella abortus S19 lysate, increase in the neutrophil count from 30 to 90 day and 60 to 90 day was significant (P<0.05). The increase in neutrophil levels from 0 to 90 day was significant (P<0.01). In case of animals immunized with Brucella abortus RB51 lysate, it was observed that there was no significant difference in the neutrophil counts between the various intervals. In animals treated with cocktail lysate preparation the difference between the mean values at 0D vs 90D was significant (P<0.01). The difference between the mean values at 0D vs 75D was significant (P<0.05). In Brucellosis affected animals which received no treatment, the variations in neutrophil counts were non-significant at various intervals.
In order to assess the effect of RB51 lysate on cell mediated immune status of infected animals, γ-interferon production was taken as a criteria of induction of cellular immunity. Brucella spp. are facultative intracellular pathogens inducing a cell-mediated immunity which leads to the T-cell-dependent activation of macrophages through gamma interferon.
The effect of lysate immunization on cell mediated immunity against Brucella was assessed for detection of interferon-γ producing lymphocytes. The ELISPOT assay was carried out on peripheral blood lymphocytes from Brucellosis affected cattle treated with lysates of vaccine strains of Brucella. The number of spots indicating interferon-γ producing lymphocytes was taken as an index of stimulation of cell mediated immunity.
The production of interferon by peripheral blood lymphocytes of immunized animals was detected by ELISPOT assay (Fig 3). The number of spots per well was 1.916 ± 1.74 in S19 lysate treated, 94.583 ± 13.18 in RB51 lysate treated and 90.416 ± 17.05 in cocktail lysate treated animals. The difference between the mean values of spot forming cells (SFCs) of S19 and RB51 treated cattle was significant (P<0.01). The difference between the means of S19 vs cocktail treated animals was significant (P<0.01). The difference between the mean values of SFCs of RB51 vs combination of S19 and RB51 lysates immunized cattle was non-significant.
The Brucella abortus RB51 lysate induced a significantly (P<0.01) higher number of spots compared to S19 lysate. Similarly, the cocktail lysate induced significantly (P<0.01) higher amount of interferon-γ production in lymphocytes compared to Brucella aborus S19 lysate alone. This indicates that Brucella abortus RB51 lysate is very potent in stimulating the cell mediated immunity in cattle which could be protective against Brucellosis.
The interferon-γ assay, a rapid and convenient alternative to the DTH test, has been reported[23, 24], which appears to be an ideal method that is complementary to the serological diagnosis protocols. It has been reported[25] that IFN-gamma, which is an important mediator of acquired cell mediated immune response in murine model of Brucellosis, could be assayed for diagnostic purposes in the case of cattle also. Bovine IFN-γ has been used as a diagnostic test in the context of bovine Brucellosis, using Brucella proteins as an antigenic stimulus in vitro[26].
The total plasma RNA concentrations were determined in lysate treated and untreated cattle affected with Brucellosis (Fig 4).
The total RNA concentrations in plasma at 0 day and various intervals post-immunization in Brucellosis affected cattle treated with S19 lysate, RB51 lysate, or cocktail (S19+RB51) lysate as well as affected untreated cattle are shown in Table 5.
ANOVA and Tukey’s post-hoc test revealed that in case of S19 lysate treated cattle, the decrease in RNA concentrations between 0D and 30D, 60D, 75D and 90D, between 30D versus 75D and 90D, between 60D versus 90D and between 75D versus 90D were significant (P<0.01). In case of RB51 lysate treated cattle the mean values of RNA decreased significantly (P<0.01) from 0D vs 30D, 60D, 75D and 90D, 30D vs 90D and 60D vs 90D (P<0.01). The differences in mean values between 30D vs 75D, 60D vs 75D and 75D vs 90D were significant (P<0.05). The difference in mean values between 30D vs 60D was nonsignificant. In case of cattle treated with cocktail lysate (S19+RB51) the decrease in RNA concentrations between 0D and 30D, 60D, 75D and 90D, from 30D to 75D and 90D and 60D vs 90D were significant (P<0.01). The decrease in RNA concentrations between 60D vs 75D was significant (P<0.05). The differences in mean values between 30D vs 60D and 75D vs 90D were nonsignificant.
The effect of lysate immunization on survival of Brucella in Brucellosis affected cattle was studied by RT-PCR of plasma RNA employing Brucella specific primers.
To study the efficacy of the phage lysates on clearing of Brucella organisms, Reverse Transcriptase-PCR was employed. RNA was isolated from the plasma samples of cattle. Pre-preparations of RNA with OD260/OD280 ratio ranging between 2.0–2.2 were selected. The concentration of total RNA varied between 400–2000 ng/μL in different samples. The amount of total RNA used for cDNA synthesis was 2μg for each sample.
The complementary DNA (cDNA) synthesized was amplified for the gene encoding 31 kDa Brucella abortus antigen (bcsp 31) using B4/B5 primers. At 0 day, as expected, the 31 kDa bcsp 31 PCR assay resulted in an amplicon of 223 bp when applied to the plasma samples of experimental animals in the trial.
After the administration of the phage lysates, the RNA based monitoring was carried out at regular intervals of 0, 30, 60, 75 and 90 days post treatment. Plasma samples were collected and then subjected to RT-PCR (Figs 5 & 6).
Culture of blood, vaginal and uterine secretions on BSM agar showed no growth in case of lysate treated animals at 90D while in case of untreated animals it showed growth of Brucella organisms (Fig 7).
When supernatant—free lysate pellets obtained after high speed centrifugation of whole lysates were streaked on S19 lawn, they did not lyse S19 organisms after incubation (Fig 8) indicating that the lysates were free of phage after filtration and the immunotherapeutic effect observed was due to the immunogens present in the lysates and not due to the phage.
The Brucella abortus amplified gene detected in Brucellosis infected untreated animals could be due to the live Brucella abortus organisms which persisted even after eliciting the immune response in the animal. The infected animal may act as a carrier by shedding the live organism in its secretions or excretions. The Brucella specific 223 base pair amplicon was clearly evident at 0 day in case of lysate treated animals but started diminishing at 75 day post immunization and became apparently invisible by 90 day indicating that, live Brucella organisms were reduced to almost negligible numbers after 3 months of treatment with lysate.
The plasma RNA diminished very prominently in case of Brucella abortus RB51 lysate immunized animals and cocktail (S19+RB51) lysate immunized animals and very substantially in S19 immunized animals. Since Brucella abortus RB51 is known to induce cell mediated immunity in cattle, the pronounced antibacterial effect in RB51 lysate immunized animals as well as cocktail (S19+RB51) lysate immunized animals was expected. The lysate of Brucella abortus S19 which is known to induce antibody alone, also contributed significantly in reducing the bacterial load due to antibody mediated effect on extracellular Brucella organisms. On the other hand the Brucella specific band was consistently and very conspicuously present throughout the period from 0 to 90 day in untreated cattle suggesting the persistence of bacteria in the host in large proportions.
This study is probably the first of its kind to demonstrate unequivocally the effect of phage lysate in treatment of bacterial infections with actual quantification of bacterial load before and after treatment in a minimally invasive way employing a Brucella specific biomarker. Moreover the therapeutic benefit of Brucella abortus RB51 lysate in infected adult cattle has been documented for the first time. The comparative study has reconfirmed the relative importance of antibody and cell mediated immunity in the treatment of Brucellosis.
In principle, the presence of one or another type of nucleic acid (DNA, rRNA, mRNA or tRNA) in bacterial cells might be a useful indicator of viability[27]. PCR is a rapid and sensitive method for detecting and identifying the bacterial organisms. DNA targeted PCR is incapable of determining bacterial viability as DNA is having high stability and is also demonstrated to persist in the PCR detectable form in culture negative environmental and clinical samples[22, 28]. Due to the short half life of RNA, it has been considered as plausible indicator of bacterial viability[29]. We have also found plasma RNA to be a good biomarker in phage therapy of bovine Brucellosis[30].
In a study[31], PCR method targeting bcsp 31 was found to be most sensitive than that for outer memberane protein 2 (omp 2) and 16S rRNA. The 16S rRNA gene PCR used for detection of bovine blood samples has been reported[32] to be insensitive. B4/B5 amplification was reported[33]to be most sensitive as it could amplify DNA extracted as low as 25 and 100 cfu/ml in 1 ml water and blood, respectively. Primer set B4/B5 used in our study was able to detect DNA at 715cfu/ml. It has been reported[34] that B4/B5 primer pair was able to detect the smallest number of bacteria (700cfu/mL).
A single dose of 2 ml of phage lysate of Brucella abortus attenuated strain RB51 alone or in combination with lysate of Brucella abortus attenuated strain S19 administered subcutaneously in naturally Brucellosis infected adult cows successfully eliminated virulent Brucella abortus from the infected cattle in 3 months which were negative on blood culture. Brucella specific RNA in blood plasma was found to be a useful biomarker for minimally invasive and non destructive monitoring of the effect of antibacterial therapy in Brucellosis affected cattle.
|
10.1371/journal.ppat.1002294 | Biochemical and Structural Insights into the Mechanisms of SARS Coronavirus RNA Ribose 2′-O-Methylation by nsp16/nsp10 Protein Complex | The 5′-cap structure is a distinct feature of eukaryotic mRNAs, and eukaryotic viruses generally modify the 5′-end of viral RNAs to mimic cellular mRNA structure, which is important for RNA stability, protein translation and viral immune escape. SARS coronavirus (SARS-CoV) encodes two S-adenosyl-L-methionine (SAM)-dependent methyltransferases (MTase) which sequentially methylate the RNA cap at guanosine-N7 and ribose 2′-O positions, catalyzed by nsp14 N7-MTase and nsp16 2′-O-MTase, respectively. A unique feature for SARS-CoV is that nsp16 requires non-structural protein nsp10 as a stimulatory factor to execute its MTase activity. Here we report the biochemical characterization of SARS-CoV 2′-O-MTase and the crystal structure of nsp16/nsp10 complex bound with methyl donor SAM. We found that SARS-CoV nsp16 MTase methylated m7GpppA-RNA but not m7GpppG-RNA, which is in contrast with nsp14 MTase that functions in a sequence-independent manner. We demonstrated that nsp10 is required for nsp16 to bind both m7GpppA-RNA substrate and SAM cofactor. Structural analysis revealed that nsp16 possesses the canonical scaffold of MTase and associates with nsp10 at 1∶1 ratio. The structure of the nsp16/nsp10 interaction interface shows that nsp10 may stabilize the SAM-binding pocket and extend the substrate RNA-binding groove of nsp16, consistent with the findings in biochemical assays. These results suggest that nsp16/nsp10 interface may represent a better drug target than the viral MTase active site for developing highly specific anti-coronavirus drugs.
| The distinctive feature of eukaryotic mRNAs is the presence of methylated cap structure that is required for mRNA stability and protein translation. As all viruses employ cellular ribosomes for protein translation, most cytoplasmically replicating eukaryotic viruses including coronaviruses have evolved strategies to cap their RNAs. It was shown very recently that ribose 2′-O-methylation in the cap structure of viral RNAs plays an important role in viral escape from innate immune recognition. The 2′-O-methyltransferase (2′-O-MTase) encoded by SARS coronavirus is composed of two subunits, the catalytic subunit nsp16 and the stimulatory subunit nsp10, which is different from all other known 2′-O-MTases that are partner-independent. Here we show that the role of nsp10 is to promote nsp16 to bind capped RNA substrate and the methyl donor S-adenosyl-L-methionine (SAM). We solved the crystal structure of the nsp16/nsp10/SAM complex, and the structural analysis revealed that the details of the inter-molecular interactions and indicated that nsp10 may stabilize the SAM-binding pocket and extend the capped RNA-binding groove. The interaction interface of nsp16/nsp10 is unique for coronaviruses and thus may provide an attractive target for developing specific antiviral drugs for control of coronaviruses including the deadly SARS coronavirus.
| Coronaviruses are etiological agents of respiratory and enteric diseases in livestock, companion animals and humans, exemplified by severe acute respiratory syndrome coronavirus (SARS-CoV) which was responsible for a worldwide SARS outbreak in 2003 and caused over 8000 cases of infection with about 10% fatality rate. They are characterized by possessing the largest and most complex positive-stranded RNA genome (ranging from 27 to 31 kb) among RNA viruses. Fourteen open reading frames (ORFs) have been identified in the genome of SARS-CoV, of which 12 are located in the 3′-one third of the genome, encoding the structural and accessory proteins translated through a nested set of subgenomic RNAs [1], [2]. The 5′-proximal two thirds of the genome comprise 2 large overlapping ORFs (1a and 1b), which encode two large replicase polyproteins that are translated directly from the genome RNA, with 1b as the frameshifted extension of 1a. These two precursor polyproteins are cleaved into 16 mature replicase proteins, named as non-structural protein (nsp) 1–16, which form the replication-transcription complex (RTC) localized in endoplasmic reticulum-derived membranes [3], [4]. Strikingly, the coronavirus genome is predicted to encode several RNA processing enzymes that are not common to small RNA viruses [1], including nsp14 as an exoribonuclease and guanine N7-methyltransferase (N7-MTase) [5], [6], [7], [8] and nsp15 as a nidovirus-specific endonuclease [9], [10].
Eukaryotic and most viral mRNAs possess a 5′-terminal cap structure, in which an N7-methyl-guanine moiety is linked to the first transcribed nucleotide by a 5′-5′ triphosphate bridge [11], [12]. The cap structure is essential for efficient splicing, nuclear export, translation and stability of eukaryotic mRNA [13], [14], [15], [16]. All viruses use the translational machinery of host cells. With the exception of some viruses, such as picornaviruses and hepatitis C virus that circumvent the capping problem by using an internal ribosome entry site (IRES) for mRNA translation [17], [18], viruses of eukaryotes have evolved diversified strategies to cap their mRNAs that are thus translated by cap-dependent mechanisms in the manner of eukaryotic mRNAs. It has been suggested for three decades that coronavirus mRNA may carry a 5′-cap structure [19], [20], [21], [22], but the principal enzymes involved in coronavirus RNA capping and their biochemical mechanisms have not been characterized until recently. Cap formation of eukaryotic and viral mRNAs requires universally three sequential enzymatic reactions. First, an RNA triphosphatase (TPase) removes the γ-phosphate group from the 5′-triphosphate end (pppN) of the nascent mRNA chain to generate the diphosphate 5′ -ppN. Subsequently, a RNA guanylyltransferase (GTase) transfers a GMP to the 5′-diphosphate end to yield the cap core structure (GpppN). Then a N7-MTase methylates the capping guanylate at the N7 position to produce a cap-0 structure (m7GpppN) [13]. While lower eukaryotes, including yeast, employ a cap-0 structure, higher eukaryotes and viruses usually further methylate the cap-0 structure at the ribose 2′-O position of the first and second nucleotide of the mRNA by a ribose 2′-O MTase to form cap-1 and cap-2 structure, respectively [13]. Very recently, it was shown that ribose 2′-O-methylation of viral RNA cap provides a mechanism for viruses to escape host immune recognition [23], [24].
For coronaviruses, several nsps have been indicated to be involved in viral RNA capping. We have shown that SARS-CoV nsp14, a previously described exoribonuclease, acts as N7-MTase to generate cap-0 structure [8]. Recently, it was shown that SARS-CoV nsp16 acts as 2′-O-MTase in complex with nsp10 and selectively 2′-O-methylates the cap-0 structure to give rise to cap-1 structure [25]. Feline coronavirus (FCoV) nsp16 was also shown to bind N7-methyl guanosine cap (m7GpppAC3-6) and methylate the penultimate nucleotide at 2′-O position to yield a cap-1 structure in vitro [26]. Coronavirus nsp13 has been shown to exhibit RNA TPase activity in vitro and thus proposed to be functional in RNA capping reaction [27], [28], but a direct role for nsp13 in RNA capping still awaits experimental evidence. Currently, the RNA GTase that is essential for cap formation is completely unknown for coronaviruses.
SARS-CoV nsp16 requires nsp10 as a stimulatory factor to execute its 2′-O-MTase activity [25] and this also holds true for other coronaviruses (our unpublished results). This mechanism is unique for 2′-O-MTase of coronaviruses and has not been found in any other viruses or host cells. However, the molecular mechanisms underlying the enzymatic activity of nsp16 and the stimulatory effect of nsp10 are unknown. Here we report the crystal structure of the heterodimer of nsp16 and nsp10 (nsp16/nsp10) with bound methyl donor SAM and biochemical characterization of the stimulation mechanisms. We found that nsp10 is required for nsp16 to bind both SAM and RNA substrate, and the crystal structure shows that nsp10 may stabilize the SAM-binding pocket and extend the RNA-binding groove of nsp16. These results have implications for designing specific anti-coronavirus drugs to control infection.
In our previous work, we adopted a genetic screening system and biochemical assays to identify SARS-CoV nsp14 as N7-MTase [8]. However, we did not observe any 2′-O-MTase activity in various biochemical assays for SARS-CoV nsp16, which was previously predicted to be 2′-O-MTase [1], [29], although a low 2′-O-MTase activity was demonstrated for feline coronavirus nsp16 [26]. We and others showed previously that SARS-CoV nsp10 could interact with both nsp14 and nsp16 [30], [31], suggesting a role for nsp10 in the functions of nsp14 and nsp16. Therefore, we undertook to test the effects of nsp10 on the MTase activities of both nsp14 and nsp16. As shown in Figure 1A, by using radiolabeled and unmethylated G*pppA-capped RNA as substrate (where the * indicates that the following phosphate was 32P labeled, and the sequence is identical with viral genomic RNA except for the second nucleotide), nsp14 alone could efficiently N7-methylate GpppA-RNA to generate cap-0 structure (m7GpppA) (Figure 1A, lane 2) and nsp10 did not significantly alter nsp14 N7-MTase activity in our testing system (Figure 1A, lane 4). While either nsp16 or nsp10 alone did not show any MTase activity (Figure 1A, lanes 3 and 1), the mixture of nsp10/nsp14/nsp16 gave rise to cap-1 structure (m7GpppAm) (Figure 1A, lane 7), indicating that addition of nsp10 and nsp16 rendered the complex active in the 2′-O-MTase reaction. We then used radiolabeled m7G*pppA-capped RNA substrate and demonstrated that nsp14 could not 2′-O-methylate RNA cap-0 structure (Figure 1A, lane 12) but the mixture of nsp16 and nsp10 (nsp16/nsp10) did possess the 2′-O-MTase activity to convert cap-0 to cap-1 structure at pH 7.5 and 8.0 (Figure 1A, lanes 10 and 16). These results indicate that nsp10 may function as a stimulatory factor for nsp16 and is required for the 2′-O-MTase activity of nsp16. While this work was ongoing, similar observations were made by Bouvet et al. [25].
Our previous studies showed that SARS-CoV nsp14 N7-MTase could N7-methylate both GpppA- and GpppG- capped RNA in a sequence-independent manner [8]. As both the genomic and subgenomic RNAs of SARS-CoV all start with an adenine, we tested whether SARS-CoV nsp16 2′-O-MTase has sequence specificity by using m7GpppA- and m7GpppG-capped RNAs as substrates. As shown in Figures 1B and 1C, SARS-CoV nsp16/nsp10 complex gave a strong methylation signal on m7GpppA-capped RNA substrate but not on m7GpppG-capped RNA, suggesting that SARS-CoV nsp16/nsp10 functions in cap sequence-specific manner. The VP39 2′-O-MTase of vaccinia virus was used as a control in the methylation assay system (Figure 1C). Furthermore, we observed that nsp16/nsp10 could not methylate cap analogues and individual NTPs (Figure 1B), indicating that the RNA substrate for nsp16/nsp10 needs to contain a stretch of nucleotides linked to the cap. A recent study showed that the cap with a 5 nucleotide extension was sufficient as substrate for nsp16 [25]. In this short RNA substrate, only the first nucleotide is identical with the first nucleotide A in the SARS-CoV genome RNA and the remaining 4 nucleotides are different from the genomic sequence. Taken together, these results indicate that the first nucleotide A is the sequence determinant for nsp16 methylation specificity. These results also suggest that nsp16 or nsp16/nsp10 complex need to form an RNA-binding groove that has enough space and affinity to accommodate m7GpppA-capped RNA with an extension of a few of nucleotides. Binding of the cofactor SAM and substrate RNA to MTase is the prerequisite for its enzymatic activity, and therefore we next tested whether SARS-CoV nsp10 might promote the SAM- and RNA-binding capability of nsp16.
SAM is the methyl donor for both N7- and 2′-O-MTase in RNA cap methylation, and high affinity binding and correct positioning of the cofactor in the SAM-binding site of the MTase provides the basis for the methyl transfer into the substrate RNA. Therefore, we first tested whether nsp10 could influence the binding affinity of nsp16 toward SAM. As shown in Figure 2A, nsp16 alone or the complex of nsp10 and control proteins (nsp12N and nsp3-SUD) were not able to bind SAM at different pH values (Figure 2A, lanes 1, 2, 6, and 7). However, nsp16 could bind SAM when complexed with nsp10 (Figure 2A, lanes 3 and 4). In the mixture nsp10/nsp14/nsp16, both nsp14 and nsp16 could bind to SAM (Figure 2A, lane 5). There were no signals at the position of nsp10 in SAM binding assays (data not shown). These results demonstrate that nsp10 could boost nsp16 to bind the methyl donor SAM.
To further study the different SAM binding affinity of SARS-CoV nsp16/nsp10 complex and nsp16 or nsp10 alone, isothermal titration calorimetry (ITC) was used to measure the thermodynamic changes during SAM-binding. ITC profiles for the binding of SAM to nsp16, nsp10 and nsp16/nsp10 complex (Figures 2B, 2C and 2D, respectively) of SARS-CoV showed that the complex of SARS-CoV nsp16 and nsp10 could bind SAM specifically but either nsp16 or nsp10 alone could not. The top panels in Figures 2B-D show raw ITC curves resulting from the injections of SAM into a solution of nsp16 (Figure 2B), nsp10 (Figure 2C), and nsp16/nsp10 complex of SARS-CoV (Figure 2D). The titration curves show that SAM binding to nsp16/nsp10 complex is exothermic, resulting in negative peaks in the plots of power versus time (Figure 2D). However, adding SAM to either nsp16 or nsp10 alone is apyretic, resulting in random peaks around 0-baseline in the plots of power versus time (Figures 2B and 2C). The bottom panels in Figures 2B-D plot the heat evolved per mole of SAM added, corrected for the heat of SAM dilution, against the molar ratio of SAM to nsp16 (Figure 2B), nsp10 (Figure 2C), and nsp16/nsp10 complex of SARS-CoV (Figure 2D). The thermodynamic parameters for the binding between SAM and nsp16/nsp10 complex (N = 0.873±0.0817 sites, = −18.12±11.07 kcal mol−1, Kd = 5.59±1.15 µM, and = −36.7 cal mol−1 K−1) were obtained by fitting the data to a single set of identical sites model, indicating that SAM bound to nsp16/nsp10 complex with a moderate affinity in the absence of DTT. The N value indicated the binding stoichiometry and it suggested that the molecular ratio of SAM to nsp16/nsp10 complex is 1∶1. As the observed N value is less than 1, it may indicate that not all nsp16 bound with nsp10 to form nsp16/nsp10 complex in the assays. As shown by the value of dissociation constant (Kd), the binding affinity between SAM and nsp16/nsp10 complex was moderate. The mutual binding affinity of SARS-CoV nsp10 and nsp16 was also measured by ITC, and the Kd value was 2.11±0.97 µM, which indicated similar binding affinity to that of SAM and nsp16/nsp10 complex. Taken together, these results showed that nsp10 is required for nsp16 to bind the methyl donor SAM and the binding affinity of SAM and nsp16/nsp10 complex is moderate.
We then tested whether nsp10 is required for specific binding of m7GpppA-capped RNA. In gel shift assays in the absence of zinc ions, neither nsp10 nor nsp16 could shift the RNA bands (Figure 3A, lanes 1 and 2) but nsp14 and the complex of nsp16 and nsp10 could (Figure 3A, lanes 6 and 3). In the presence of zinc ions, nsp10 as zinc finger protein could bind RNA non-specifically (Figure 3A, lane 5). The complex of nsp16 and nsp10 did not associate with m7GpppG-capped RNA (Figure 3B) and cap analogues (m7GpppA and m7GpppG) (Figures 3C and 3D). These results indicate that SARS-CoV nsp10 promotes nsp16 to specifically bind m7GpppA-capped RNA. To confirm these results, we adopted pull-down assays by using hexahistidine-tagged proteins and radiolabeled RNAs or cap analogues. As shown in Figure 3E, pull-down of nsp16 and nsp10 mixture by nickel-nitrilotriacetic acid (Ni-NTA) resin gave rise to high level of radioactive signal for *m7GpppA-capped RNA (where * indicates that the methyl group was 3H labeled) while nsp10, nsp16, or nsp5 alone could not pull down the labeled RNA. Also in this testing system, nsp16/nsp10 complex had a low binding affinity to *m7GpppG-capped RNA and cap analogues (*m7GpppA and *m7GpppG) (Figure 3E). The pulled-down proteins were further checked by Western blotting to confirm the presence of the indicated proteins (Figure 3F). These results collectively showed that the nsp16 could bind specifically to m7GpppA-capped RNA only in the presence of nsp10.
Furthermore, we also analyzed the binding specificity and thermodynamics between nsp16/nsp10 and capped RNA in ITC assays. The thermodynamic changes are shown for nsp16 (Figure 3G), nsp10 (Figure 3H) and nsp16/nsp10 complex of SARS-CoV (Figure 3I) when added to the solution of m7GpppA-capped RNA, respectively. The top panels in Figures 3G-I show raw ITC curves and the bottom panels plot the heat evolved per mole of the injected protein, corrected for the heat of the corresponding proteins dilution, against the molar ratio to m7GpppA-capped RNA. The titration curves show that m7GpppA-capped RNA binding to nsp16/nsp10 complex is exothermic, resulting in negative peaks in the plots of power versus time (Figure 3I). However, all others were apyretic, resulting in random peaks around 0-baseline in the plots of power versus time (Figures 3G and 3H). The binding affinity of nsp16/nsp10 complex with m7GpppA-capped RNA is higher than that with SAM as shown by the thermodynamic parameters for the binding between nsp16/nsp10 complex and m7GpppA-capped RNA (N = 0.840±0.0711 sites, = −18.58±2.171 kcal mol−1, Kd = 1.21±0.41 µM, and = −35.2 cal mol−1 K−1). We further analysed the thermodynamic changes when cap analogues (m7GpppA and m7GpppG) were added to nsp16/nsp10 complex in ITC, but neither of the two cap analogues showed exothermic binding (data not shown), which suggested that the cap analogues are not the substrate of nsp16/nsp10 complex.
All together, these data showed that nsp10 plays an essential role in the specific binding of m7GpppA-capped RNA by nsp16. However, m7GpppG-capped RNA and cap analogues can not bind to the nsp16/nsp10 complex, and thus can not be used as substrate by 2′-O-MTase of SARS-CoV.
Based on the results presented above, it is suggested that SARS-CoV nsp10 may either stabilize the SAM-binding pocket of nsp16 or change the conformation of nsp16 so as to efficiently take in and hold the SAM molecule. Furthermore, the association of nsp10 with nsp16 may provide a proper groove for binding and holding of m7GpppA-capped RNA substrate. We thus expected to reveal the details of SAM- and RNA-binding and stimulation mechanisms from the crystal structure of nsp16/nsp10.
To obtain crystals of nsp16/nsp10 protein complex combined with its MTase co-substrate SAM, 6×histidine-tagged nsp16 and 6×histidine-tagged nsp10 with nsp11 extension were expressed individually in Escherichia coli cells and co-purified with Ni-NTA resin. The protein mixture of purified nsp16 and nsp10 was supplemented with methyl donor SAM to obtain protein complex and then applied to crystallization by the hanging-drop vapor diffusion method. Crystals appeared readily in hanging drops and were diffracted to high resolution at 2.0 Å under X-rays from synchrotron radiation source. The structure was solved subsequently by multi-wavelength anomalous diffraction method taking advantage of selenomethionine substituted crystals (see Materials and Methods). Within one asymmetric unit in the crystals, one nsp10, one nsp16 and one SAM molecule were identified unambiguously. Residues in both nsp10 and nsp16 were clearly traced except the C-terminus after Ser129 in nsp10 and the nsp11 extension. The atomic coordinates of the structure have been deposited in the Protein Data Bank (PDB) as entry 3R24.
The structure of nsp16 (Figures 4A and 4B) exhibits the characteristic fold of the class I MTase family, comprising a seven-stranded β-sheet surrounded by α-helices and loops [32]. Search of the PDB using DALI [33] identified high structural similarity of nsp16 with FtsJ (PDB entry 1EIZ), a partner-independent MTase from E. coli [34], with a 2.6 Å root-mean square deviation (RMSD) for 179 aligned Cα (from Y30 to A209 of nsp16 with 16% sequence identity and Dali Z-score 26.6). The residues 30-209 of nsp16 form the core MTase domain. Nevertheless, some differences between the two structures are evident. The most significant difference lies in the αD helix, which is visible on the surface of FtsJ (Figure S1E) but invisible for nsp16 (Figure S1B). The αD helix is important for both SAM-binding and RNA cap-binding, and it is very short in partner-dependent nsp16 2′-O-MTase but relatively long in all other known viral 2′-O-MTases, including vaccinia virus VP39 (PDB entry 1AV6) (Figure S1C), Dengue virus NS5 MTase (PDB entry 1L9K) (Figure S1D), and Bluetongue virus VP4 2′-O-MTase (PDB entry 2JHP) (Figure S1F), which are partner-independent MTases. The differences between nsp16 and other 2′-O-MTases are shown in the structure-based alignment of 2′-O-MTases (vaccinia virus VP39, Flavivirus NS5 MTase, and FtsJ) (Figure S2).
The structure of the stimulatory protein nsp10 in nsp16/nsp10 complex (Figures 4A and 4B) is consistent with the structure of nsp10 reported previously [35], [36],indicating that the structure of nsp10 is not impacted by the interaction between nsp16 and nsp10. Nsp10 can be roughly segregated into three regions: a helical domain at the N terminus followed by an irregular β-sheet region, and a C-terminal loop region. Two zinc ions were clearly identified in nsp10, which formed the center of two zinc fingers, one coordinated by Cys-74, Cys-77, His-83 and Cys-90, and the other coordinated by Cys-117, Cys-120, Cys-128 and Cys-130. The two zinc fingers render nsp10 the ability to bind polynucleotide chains in a nonselective manner in the presence of zinc ions [35], [36], [37].
During the review process of this work, a structure of nsp16/nsp10 complex of SARS-CoV was reported by Decroly and colleagues [38], which is generally the same as the structure described in this work. The major difference lies in that the structure we solved contains the methyl donor SAM that was purposely supplemented in the protein mixture, and the one by Decroly and colleagues contains S-adenosylhomocysteine (SAH) which is the product of SAM after methyltransfer and may be captured from the medium by nsp16 that was co-expressed with nsp10 in bacterial cells [38]. The atomic coordinates of nsp16/nsp10 structure solved by Decroly et al. have not been released until now, and thus detailed comparison of the two structures is not available.
Nsp10 and nsp16 formed a protein complex through an interaction surface covering approximately 1767 Å2 in total, indicating a very stable interaction. The interaction surface on nsp10 was dominated by hydrophobic interactions in center with surrounding hydrophilic interactions. By using the online software Interfaces and Assemblies of EMBL-EBI (http://www.ebi.ac.uk/msd-srv/prot_int/pistart.html), it was found that significant contacts between nsp10 and nsp16 involve residues 40–47/69–84/93–96 of nsp10, and residues 37–48/76–91/102–110/244–248 of nsp16 (Figures 4). Close inspection revealed a cluster of important residues involved in nsp16-nsp10 interaction, including Val-42, Met-44, Gly-70, Ser-72, Arg-78, His-80, Lys-93, Gly-94, Lys-95 and Tyr-96 in nsp10, in agreement with hotspots identified in biochemical assays [39]. Intermolecular hydrogen bonds exist between Gly-70/Ala-71/Gly-94 of nsp10 and Asp-106 of nsp16, Lys-93 of nsp10 and Ser-105 of nsp16, Leu-45 of nsp10 and Gln-87 of nsp16, Tyr-96 of nsp10 and Ala-83/Gln-87 of nsp16. There are two salt bridges between His-80 (nitrogen atoms ND1 and NE2) of nsp10 and Asp-102 (oxygen atom OD2) of nsp16 (Figures 4 D and 4E). Hydrophobic interactions were involved between Leu-45 in nsp10 and a hydrophobic pocket composed of Ile-40/Thr-48/Leu-244/Met-247 in nsp16, and Tyr-96 in nsp10 and a hydrophobic pocket consisting of Val-84, main chain of Gln-87 and Arg-86 from nsp16 (Figures 4C and 4E).
Biochemical assays showed that only nsp16/nsp10 complex could bind the substrates m7GpppA-RNA and SAM, and execute the 2′-O MTase activity. Therefore, mutations on the interaction surface of nsp16/nsp10 complex which can block this interaction should influence the substrates binding, and consequently the MTase activity of nsp16. A double mutant (H83A/P84A) and a triple mutant (Y76A/C77A/R78A) at the interaction surface of nsp10 were generated. These mutants almost completely abolished the SAM (Figure 5C) and m7GpppA-RNA (Figure 5D) binding of nsp16, and also abrogated MTase activity (Figures 5A and 5B). The main chain N atom of Gly-70 in nsp10 formed a hydrogen bond with Asp-106 of nsp16. The single mutation G70A only slightly influences this main chain interaction, and accordingly, it slightly impaired the SAM- and RNA-binding activity of nsp16 (Figures 5C, lane 3 and 5D) and attenuated by 30% the 2′-O-MTase activity of nsp16 (Figures 5A, lane 5 and 5B). Similar results were obtained by Lugari et al., except for an Y96F mutation, which increased both the nsp10-nsp16 affinity and the MTase activity of nsp16 [39]. In the interaction surface of nsp16/nsp10, the side chain of Tyr-96 in nsp10 stacks to the hydrophobic pocket consisting of Val-84, main chain of Gln-87 and Arg-86 in nsp16. Compared with tryptophan, the side chain of phenylalanine exhibits a stronger hydrophobicity. Therefore, Y96F mutation may strengthen this hydrophobic interaction of nsp10 and nsp16, and thus enhance nsp16/nsp10 MTase activity. This observation suggests that the hydrophobic interaction by the aromatic nucleus of Tyr-96 is more important than the hydrogen bond made by its hydroxyl group for maintaining the interaction surface with nsp16. Taken together, biochemical analysis of the critical residues involved in the interaction interface was consistent with the structural observations.
Methyl donor SAM was added to protein mixture used for crystal screening, and SAM molecule is visible in the crystal structure of nsp16/nsp10 complex. The ligand SAM lies at the C-terminus of strands β1 and β2, similar to most SAM-dependent MTases [32] (Figures 4A and 4B). As shown in Figure 6A, the adenosine moiety of SAM is stacked among Phe-149, Met-131 and Cys-115 in nsp16 through Van der Waals force, and polar contacts also exist, such as those of adenine N6 with side chain of Asp-114, N1 with main chain of Cys-115, and N3 with main chain of Leu-100. The 2′- and 3′-OH of the adenosine ribose of SAM are stabilized by the side chain of Asp-99 and the side chain of Asn-101 via hydrogen bonds. The hydrophobic patch of the ribose packs against Met-131 and the hydrophobic main chain of Tyr-132. The amino group of the methionine moiety of SAM is maintained by three polar contacts: N atom forms a hydrogen bond with main chain carbonyl group of Gly-71, side chain of Tyr-47 and Asp-130; hydroxyl O atom interacts with main chain of Gly-81 and main chain of Ala-72 and Gly-73; and carbonyl O atom interacts with side chain Asn-43 (Figure 6A).
The crystal structure of nsp16/nsp10 complex shows that the SAM binding cleft of nsp16 is composed of three loops: loop 71–79 (residues 71–79) together with loop 100–108 (residues 100–108) forming a wall on one side of the cleft and loop 130–148 (residues 130–148), which is followed by αD helix, forming a wall on the other side. Several ‘hot spot’ residues were located at the loop regions (Leu-100, Asn-101, Asp-130, Met-131, Tyr-132) (Figure S1A). Structural analysis revealed that the loop followed by αD helix of partner-independent 2′-O-MTases is shorter than that of nsp16 (loop 130–148) and is sustained by rigid αD helix such as to form a stable wall to keep the SAM inside the cleft. In contrast, in the SAM-binding pocket of nsp16, the corresponding loop (loop 130–148) is long and flexible while the supporting rigid αD helix is short (Figure S2). Moreover, the loop 100–108 of nsp16 is also longer and more flexible than that in other 2′-O-MTases. These structural features may make the SAM-binding cleft more flexible and thus in need of extra support from the stimulatory factor nsp10. In Flavivirus NS5 MTase, the loop corresponding to residues 100–108 was much longer for unknown reasons, but due to the strong sustaining effect of the long αD helix, the SAM-binding cleft of NS5 MTase is postulated to be more stable than that of nsp16. Therefore, it appears that at least one stable wall is essential for the SAM binding cleft to maintain the SAM binding activity. In the crystal structure of nsp16/nsp10 complex, one hydrogen bond forms between Lys-93 of nsp10 and Ser-105 of nsp16, and two salt bridges exist between His-80 ND1 and NE2 from nsp10 with Asp-102 OD2 from nsp16 (Figure 6B). Both Ser-105 and Asp-102 of nsp16 are located in the flexible loop region 100–108, which stabilizes one wall of the SAM binding cleft (Figures 6C and S1A), and consequently promotes the SAM binding activity of nsp16/nsp10 complex (Figure 2). This phenomenon of enhancing the SAM-binding activity by stabilization of the binding cleft via protein-protein interaction is observed for the first time, revealing a unique mode of SAM binding among the 2′-O-MTases.
Previous studies have demonstrated the existence of a conserved motif for methyl-transfer: K-D-K-E residues among various 2′-O MTases which catalyze an SN2-reaction-mediated 2′-O methyl transfer [40], [41], [42]. Structure-based alignment of nsp16 with VP39, NS5 MTase and FtsJ highlights these four strictly conserved residues (Lys-46, Asp-130, Lys-170 and Glu-203) in nsp16 (Figure S2). In the crystal structure, the SAM methyl group stretches out to the surface provided by the K-D-K-E motif (Figures 7B and S3), which are located at the bottom of the central groove, and might bind the first adenine nucleotide, conserved in SARS genomic and subgenomic RNAs, as the acceptor of methyl group during the methylation. This was also demonstrated by the crystal structures of other 2′-O MTases [40], [43], [44], [45]. The central groove of nsp16 is positively charged (Figures 7C and 7D), and the phosphate backbone of the cap-containing RNA is highly negatively charged. These observations indicate that the central groove in nsp16 is most likely the cap binding site. However, the cap binding groove of nsp16 is built by two flexible loops (residues 26-38 and residues 130–148) in nsp16 (Figures 7A and 7B), which replace the highly stable α-helices (A1, A2 and half of the αD) in flavivirus NS5 MTase along the cap-binding groove [46] (Figure S1D). Therefore, it is obvious that the RNA binding groove of nsp16 is too flexible and unstable to hold the substrate cap-0 RNA (Figures 7A and 7B), which explains why nsp16 alone shows an extraordinary low affinity for both m7GpppA-RNA and m7GpppA cap analogue in biochemical assays (Figures 3A, lane 2, 3C and 3E).
To further characterize the substrate RNA binding site of nsp16/nsp10 complex, we performed molecular modeling of nsp16/nsp10 complex with m7GpppAAAAAA (m7GpppA-RNA) and m7GpppGAAAAA (m7GpppG-RNA), respectively (Figure 7C, 7D and 7E). m7GpppG-RNA was derived from the structure of vaccinia virus 2′-O-MTase VP39 (PDB entry: 1AV6), and m7GpppA-RNA was mutated manually from m7GpppG-RNA. As shown in Figures 7C and 7D, the first three transcriptional nucleotides are contacted with nsp16 and the following transcriptional nucleotides are contacted with nsp10. This docking model shows that nsp10 is involved in substrate RNA binding in nsp16/nsp10 complex, as the existence of nsp10 extends the positively charged area (Figures 7C and 7D), consequently elongating the RNA-binding groove of nsp16, which may increase the RNA-binding affinity.
In biochemical assays, nsp16/nsp10 complex indeed showed increased binding affinity for m7GpppA-capped RNA (Figures 3A and 3E) but still maintained a very low affinity for m7GpppA cap analogue (Figures 3C and 3E). These results indicate that m7GpppA alone is not long enough to reach the extended positively charged area provided by nsp10 and additional nucleotides following the m7GpppA cap are needed for binding to the nsp16/nsp10 complex. As shown previously, m7GpppAC5 acted as an effective substrate of nsp16/nsp10 complex [25], suggesting that as few as 5 extra nucleotides are sufficient. Nsp10 itself possesses two zinc-finger motifs and has the ability to bind polynucleotides nonspecifically in the presence of zinc ions (Figure 3A, lane 5) [37]; therefore the RNA binding by the nsp10 portion is not sequence-specific. In conclusion, the RNA-binding groove extension provided by nsp10 may contribute to hold the extended RNA chain following the m7GpppA cap and stabilize the interaction between m7GpppA-RNA and nsp16 cap binding site.
By analysis of RNA binding site of nsp16/nsp10 complex, an unexpected promontory composed of residues 74–77 at the catalysis activity surface of nsp16 could be readily identified, which might have steric hindrance for binding GpppG-capped RNA. The residue Asp-75 (D75) stretches out to the C2 atom of first transcribed nucleotide (Figure 7E) and may thus functions as the specificity determinant. We also performed structural alignment analysis between nsp16/nsp10 complex and VP39 (PDB entry: 1AV6) at the methyltransferase activity site (Figure 7F and 7G). Compared with VP39, the binding pocket for substrate RNA of nsp16/nsp10 complex appears more limited, due to the existence of residues 74–77, especially D75. The amino group connected to C2 atom of first transcribed guanylic acid in m7GpppG-RNA seems too close to the side chain of D75 compared with m7GpppA-RNA, which has no amino group at C2 atom. This structural difference might explain why m7GpppG-RNA exhibits a relatively low affinity to the nsp16/nsp10 complex (Figure 3E).
SARS-CoV nsp16 is the only 2′-O-MTase currently known that needs a stimulatory subunit for exerting its methyltransferase activity. In this work, we showed that nsp10 could stimulate nsp16 to bind the methyl donor SAM and the capped RNA substrate. These mechanisms could be explained based on the crystal structure of nsp16/nsp10 complex, and confirmed by mutational analysis. This is reminiscent of the activation mechanism of the N7-MTase involved in vaccinia virus mRNA capping [47]. The vaccinia N7-MTase consists of two subunits, the catalytic subunit located in the C-terminal domain of D1 protein and the stimulatory subunit D12. The activation of D1 MTase activity by D12 is achieved through increase of the substrate and co-substrate affinity as well as enhancement of the stability of D1 protein [47]. Structural analysis revealed that D12 is structurally homologous to cap 2′-O-MTase with a truncation of the SAM-binding domain [48]. In contrast, SARS-CoV nsp10 does not possess an MTase fold and is structurally not similar to any other proteins deposited in the PDB database [35], [36]. Vaccinia virus VP39 protein acts as the viral 2′-O-MTase but it does not require a stimulatory factor for its enzymatic activity [40], [49].
Based on the crystal structure of nsp16/nsp10 complex and biochemical analysis, the mechanisms of nsp16 binding to substrates m7GpppA-RNA and SAM with the assistance of nsp10 were revealed. Our data showed that nsp10 acts as a buttress supporting the seemingly flexible loop 100–108 critically involved in SAM-binding (Figures 6B and 6C) and thus enhancing the SAM-binding affinity. For binding of nsp16 to capped RNA, it appears in the structure that the RNA binding groove in nsp16 has only sufficient space for binding the 5′-cap of the RNA, but for stable interaction of nsp16 and capped RNA substrate, nsp10 is needed to extend the groove and accommodate extra nucleotides following the cap (Figures 3 and 7). The residues Lys-93 and His-80 in nsp10 are involved in the interaction with the loop 100–108 of nsp16, and our results showed that the K93A mutation reduced the SAM-binding activity significantly but not the RNA-binding affinity (Figures 5C and 5D), indicating that this site is essential for SAM-binding but not for the overall interactions of nsp16 and nsp10 [39]. Similar results were obtained for H80R mutation (data not shown). The K93A mutation caused an overall decrease by 60% in the 2′-O MTase activity of nsp16/nsp10 complex (Figures 5A and 5B). The mutational and biochemical analysis of this and previous studies [39] further proved the structural model for the SAM and RNA binding mechanisms. It was reported that alanine replacements of nsp10 in murine hepatitis virus (MHV) resulting in lethal phenotypes mapped to a central core of nsp10 that is resistant to mutation, and the rescued viruses with mutations in nsp10 reduced viral RNA synthesis [50]. As most of these mutations of nsp10 were located at the interaction interface of nsp16/nsp10 complex, they might influence the activities of viral 2′-O-MTase.
The crystal structure of nsp10 in nsp16/nsp10 complex is the same as the nsp10 monomer [35]. Compared the crystal structure of nsp16/nsp10 complex with dodecamer of nsp10 [36], we found that the interaction surface of nsp16-nsp10 has partial overlap with the contact surface of nsp10 within dodecamer. Also the surface of nsp10 which may associate with nsp16 faces to the inner space of spherical nsp10 dodecamer, which does not leave sufficient space for nsp16 binding. This indicates that nsp10 dodecamer structure may not be involved in the stimulation of nsp16 2′-O-MTase activity. However, the existence of nsp10 dodecamer structure during viral infection could not be excluded because nsp10 is translated up to several times more than nsp16 [51], [52] and the surplus nsp10 may form structures other than the nsp16/nsp10 complex. In addition, nsp10 is found in the nsp4-nsp10 precursor [53] and abolishment of the nsp9-nsp10 cleavage site resulted in viable virus [54]. This observation suggests that fusion of nsp10 with nsp9 should not disrupt the essential activities of nsp10 involved in virus replication. In the nsp16/nsp10 complex, the N-terminal part of nsp10, which connects with nsp9 in the nsp9-nsp10 fusion, is exposed at the surface of the complex. The first 9 amino acids at the very N-terminus of nsp10 are invisible in the crystal structure, indicating that they might form a flexible loop. In addition, the C-terminus of nsp9 is located at the surface of this non-specific single-stranded RNA binding protein [51], [55]. Taken together, this suggests that the nsp9-nsp10 is structurally capable of forming a complex with nsp16 and may thus stimulate the 2′-O-MTase activity of nsp16. The nsp9-nsp10 fusion has less propensity to form a dodecamer, which may further weaken the biological relevance of the nsp10 dodecamer structure observed previously [36].
Nsp10 represents a multi-functional protein involved in viral RNA synthesis, polyprotein processing and RTC assembly [50], [53], [54]. It was shown that nsp10 could interact with both 2′-O-MTase nsp16 and N7-MTase nsp14 [30], [31], suggesting that a single nsp10 molecule or its dimer could associate with both nsp16 and nsp14 at the same time in the RTC. Thus, one model can be proposed to explain the RNA cap methylation during coronavirus replication: After translation and processing of polyproteins 1a and 1ab, the mature nsp14, nsp16 and nsp10 form the RNA methylation apparatus, where SAM is bound by nsp14 and nsp16 in the presence of nsp10. The newly transcribed viral RNA is capped by the unknown capping enzyme (GTase) and bound to nsp10. The 5′-end of the RNA is first associated with nsp14 and methylated at the N7 position of the cap guanine. Next, the conformation of the RNA-protein complex is altered, and the 5′-end of viral RNA is transferred from the RNA-binding groove of nsp14 to that of nsp16, resulting in second methylation at the 2′-O-site in the ribose of the first nucleotide following the cap. Further experiments are needed to confirm this model.
Cellular and DNA virus capping enzymes generally are not sequence-specific as they accommodate a large number of different mRNA species. However, the genomes of RNA viruses are very small in comparison with DNA genomes and usually encode just a few genomic and subgenomic mRNAs with conserved 5′-ends. It has been shown that the flavivirus MTases are sequence-dependent [42], [56]. In our previous work, we showed that the SARS-CoV nsp14 N7-MTase is sequence-unspecific as it could methylate the RNA cap of different RNAs both in vitro and in yeast cells [8]. However, in the current study, we showed that the nsp16 2′-O-MTase is sequence-dependent as it could only methylate m7GpppA-capped RNA, where the first nucleotide is absolutely restricted to adenosine. Structural modeling analysis suggests that the amino residues at positions 74–77 of nsp16 may be the determinant for such sequence specificity (Figures 7C, 7D and 7E). In addition, the unknown coronavirus GTase may also contribute to the specificity of coronavirus capping enzymes. In coronavirus life cycle, genomic RNA replication and subgenomic RNA transcription take place in association with double-membrane vesicle [3], and this physical restriction may make the capping apparatus accessible only to viral RNAs.
It is well known that mRNA capping and methylation play important roles in mRNA stability, processing, transport and protein translation. Very recently, it was found that 2′-O-methylation of the viral RNA cap is essential for RNA viruses to avoid innate immune recognition by the host immune system [23], [24]. Thus, inhibition of viral MTase activity should be able to suppress viral replication and attenuate viral virulence in infection and pathogenesis. The MTase active site has been suggested as a drug target for developing antiviral drugs [57], [58], [59]. However, the MTase fold is structurally conserved between viral and cellular MTases, and it is thus difficult to obtain antiviral compounds with high specificity targeting MTase active sites. For this reason, it looks more promising to target the interface of nsp16 and nsp10, which is unique to coronaviruses.
In summary, we have characterized the SARS-CoV 2′-O-MTase and the activation mechanism of nsp16 by nsp10 biochemically and structurally. We found that nsp10 promoted the substrate and co-substrate binding of nsp16 by increasing the stability of the SAM-binding pocket and by extending the RNA-binding groove of nsp16. The current findings not only provide insights into the mechanism of SARS-CoV 2′-O-methylation but also facilitate design and development of highly specific antiviral drugs targeting the nsp16/nsp10 interface.
The coding sequences of SARS-CoV nsp10-nsp11 fusion protein and nsp16 were PCR amplified from the cDNA sequence of SARS-CoV strain WHU (gi: 40795428) [2], [31] using the primers as listed in Table S1. The nsp10-nsp11 fusion protein and nsp16 genes (encoding residues Asn4240-Val4382 [35] and Ala6776-Asn7073 of replicase pp1ab) were cloned into pET30a (Novagen) (pET30a-His6-nsp10-nsp11, pET30a-His6-thrombin-nsp16) to produce recombinant proteins carrying an N-terminal His6-tag. The mutants of nsp10 (G70A, K93A, Y76A/C77A/R78A, and H83A/P84A) were generated by overlap PCR with mutagenic primers (Table S1) and cloned into pET30a as described for wild-type nsp10-nsp11 fusion protein. All constructs were verified by DNA sequencing. Both pET30a-His6-nsp10-nsp11 and pET30a-His6-thrombin-nsp16 transformed E. coli BL21 (DE3) cells were grown at 37°C in Luria-Bertani (LB) medium with 50 µg/mL kanamycin and induced with 0.4 mM isopropyl β-D- thiogalactopyranoside (IPTG) at 16°C for 12–16 hours. The SARS-CoV unique domain (SUD) of nsp3 (nsp3-SUD), nsp5, N-terminal domain of nsp12 (nsp12N), and nsp14 protein expression and purification were described previously [8]. The sequences of the cDNA and proteins have been deposited in GenBank database with accession numbers listed at the end of “Materials and Methods” section.
To obtain nsp10-nsp11 and nsp16 protein complex, 1 L of pET30a-His6-nsp10-nsp11 cells and 2 L of pET30a-His6-thrombin-nsp16 cells were mixed together and resuspended in buffer A [50 mM Tris-HCl, pH 7.5, 150 mM NaCl, 5 mM MgSO4, 5% glycerol] supplemented with 10 mM imidazole. After sonication and centrifugation, cleared lysates were applied to nickel-nitrilotriacetic acid (Ni-NTA) resin and washed with buffer A supplemented with a stepwise imidazole gradient of 20 mM, 50 mM, and 80 mM. Proteins were eluted with buffer A supplemented with 250 mM imidazole, and 0.5 mM SAM. After shaking at 4°C for 10 hours and centrifugation, the protein sample was further purified on a Superdex 200 column (GE) equilibrated with buffer A. Fractions containing the nsp16/nsp10 complex (nsp11 region was degraded during the process, data not shown) were concentrated to 10 mg/ml by ultrafiltration and frozen at −80°C for further use. The expression and purification conditions of selenomethionine (SeMet)-labeled nsp16 and nsp10 (unlabeled) protein complex were the same as for native nsp16/nsp10 complex except that modified M9 medium was used instead of LB medium during the expression of nsp16.
Crystals were grown by the hanging-drop vapor diffusion method. The drops contained 1 µl each of nsp16/nsp10 protein complex [10 mg/ml in 50 mM Tris-HCl (pH 7.5), 150 mM NaCl, 5 mM MgSO4, 5% glycerol, 5 mM SAM] and 1 µL mother liquor. Protein crystals were obtained at 25°C after 24 h in 0.1 M MES, pH 5.0, 2 M NaCl, 0.1 M NaH2PO4, 0.1 M KH2PO4. The crystals of SeMet labeled nsp16/nsp10 complex were obtained in the same conditions. For data collection, the crystals were cryocooled (by nitrogen gas stream, 100 K) in the original mother liquor containing 20% (vol/vol) glycerol and diffraction data sets were collected on beamline BL-17U1 at Shanghai Synchrotron Radiation Facility (SSRF) for native protein complex crystal and beamline 1W2B at Beijing Synchrotron Radiation Facility (BSRF) for SeMet labeled nsp16/nsp10 complex crystal. The diffraction data were processed and scaled with the HKL2000 package. Data collection statistics are listed in Table S2.
The structure was solved by the multi-wavelength anomalous diffraction (MAD) method based on two sets of derivative data and one set of native data. A preliminary model was readily built up by Phenix package [60] from derivative data, which covered around 60% of whole structure. Most of the secondary structures were obvious in the initial model, especially the 7-β-strand core. The model was then used for refining and manual building against high resolution native diffraction data. To speed up the refining process, the sole nsp10 structure was introduced into the complex by molecular replacement using Phaser [61] in CCP4 package [62]. The CNS suite [63] and Phenix refinement program (phenix.refine) [64] were used iteratively in refinement. Simulated annealing, position refining and B-factor refining were used in multiple rounds. The density for loop regions became visible gradually as refinement proceeded. A SAM molecule and coordinated zinc ion in nsp10 were included based on 2Fo-Fc electron density. Structure validation was performed periodically during refinement by Procheck [65], [66]. Eventually most of the protein sequences, except the disordered C-terminal tail in nsp10 and artificial tags generated from vectors, were involved in the final structure, and ordered water molecules were added.
The methyltransferase domain in the structure of VP39 was superimposed to nsp16 in the nsp16/nsp10 complex by LSQKAB [67] in CCP4 package [62], and the RNA molecule in VP39 was used to model the interaction between RNA and SARS nsp16/nsp10 complex. The interface between RNA and nsp16/nsp10 was optimized by energy minimization using PHENIX [64] for 3 cycles.
The ATP-initiated RNA substrates representing the 5′-terminal 259 nucleotides of the SARS-CoV genome and nonviral RNA substrate comprising 52 nucleotides (with G as the first nucleotide) were in vitro transcribed, 32P-labeled at cap structures (m7G*pppA-RNA, G*pppA-RNA, or m7G*pppG-RNA, where the * indicates that the following phosphate was radio-labeled.), and purified as previously described [8]. RNAs containing 32P-labeled cap-1 structure (m7G*pppAm-RNA) as positive control were converted from cap-0 structure m7G*pppA-RNA by a vaccinia virus 2′-O-methyltransferase VP39 following the manufacturer's protocol (Epicentre). RNAs containing unlabeled cap structures (m7GpppA-RNA or m7GpppG-RNA) were prepared by a vaccinia virus capping enzyme following the manufacturer's protocol (Epicentre) as well as 3H-labeled cap structures (*m7GpppA-RNA or *m7GpppG-RNA), except that 10 µCi of S-adenosyl [methyl-3H] methionine (67.3 Ci/mmol, 0.5 µCi/ul) was used as the methyl donor instead of cold SAM. The 32P-labeled cap analogue (m7G*pppA or m7G*pppG) and 3H-labeled cap analogue (*m7GpppA or *m7GpppG) were digested from m7G*pppA-RNA/m7G*pppG-RNA and *m7GpppA-RNA/*m7GpppG-RNA by nuclease P1 (Sigma) in 10 mM Tris-HCl, pH 7.5, 1 mM ZnCl2 at 50°C for 30 min. All the RNA substrates were extracted with phenol-chloroform and precipitated with ethanol. The unlabeled cap analogues (m7GpppA, GpppA, and m7GpppG) were purchased from New England BioLabs.
Purified recombinant or mutant proteins (0.5 µg) and 2×103 cpm of 32P-labeled m7G*pppA-RNA or G*pppA-RNA substrates were added to 8.5 µL reaction mixture [40 mM Tris-HCl (pH 7.5 or 8.0), 2 mM MgCl2, 2 mM DTT, 10 units RNase inhibitor, 0.2 mM SAM] and incubated at 37°C for 1.5 h. RNA cap structures were liberated with 5 µg of nuclease P1 (Sigma), then spotted onto polyethyleneimine cellulose-F plates (Merck) for TLC, and developed in 0.4 M ammonium sulfate. The extent of 32P-labeled cap was determined by scanning the chromatogram with a PhosphorImager [8].
MTase activity assays were carried out in 30 µL reaction mixture [40 mM Tris-HCl (pH 7.5), 2 mM MgCl2, 2 mM DTT, 40 units RNase inhibitor, 0.01 mM SAM], with 1 µCi of S-adenosyl [methyl-3H] methionine (67.3 Ci/mmol, 0.5 µCi/µl), 1 µg of purified proteins or mutant proteins, and 3 µg of m7GpppA/m7GpppG-RNA substrates or other RNA substrates (2 mM m7GpppA/GpppA/m7GpppG cap analogue or 15 mM NTPs) at 37°C for 1.5 h. 3H-labeled product was isolated in small DEAE-Sephadex columns and quantitated by liquid scintillation [68].
25 µL reaction mixtures [40 mM Tris-HCl (pH 7.5), 2 mM MgCl2, 2 mM DTT] containing 1 µg of purified proteins and 1 µCi of S-adenosyl [methyl-3H] methionine (67.3 Ci/mmol, 0.5 µCi/µl) were pipetted into wells of a microtiter plate. The reaction mixtures were incubated on ice and irradiated with 254-nm UV light in a Hoefer UVC500 cross-linking oven for 30 min. The distance of samples from the UV tubes was 4 cm. The samples were then analyzed by 12% sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). The gels were soaked in Enlightning Solution (PerkinElmer) and used for fluorography [68].
The gel shift assay provides a simple and rapid method for detecting RNA-binding proteins. This method has been widely used in the study of sequence-specific RNA-binding proteins. In this assay, 1 µg of purified proteins and 4×103 cpm of 32P-labeled RNA substrates (m7G*pppA-RNA, m7G*pppG-RNA, m7G*pppA cap analogue, and m7G*pppG cap analogue) were added to 20 µL reaction mixtures [40 mM Tris-HCl (pH 7.5), 2 mM MgCl2, 2 mM DTT]. The reactions were incubated at room temperature for 25 min, and separated by nondenaturing polyacrylamide gel (N-PAGE). The RNA substrate bands were quantitated by scanning the gels with a PhosphorImager.
In each set of RNA binding assay, 3 µg of freshly prepared 3H-labeled RNA substrates (*m7GpppA-RNA, *m7GpppG-RNA, *m7GpppA cap analogue, and *m7GpppG cap analogue) and 4 µg of purified His6-proteins were mixed in 100 µL binding buffer [40 mM Tris-HCl (pH 7.5), 2 mM MgCl2]. The binding reactions were shaken at 4°C over night. 20 µL high affinity Ni-NTA resin (50% slurry, GenScript), equilibrated with binding buffer, were added to binding reactions and mixed gently for 30 min at 4°C. The complex of 3H-labeled-RNA-His6 and Ni-NTA-bound protein was pelleted by centrifugation for 20 s at 1000 g, and washed twice with binding buffer to remove free 3H-labeled RNA substrates. The complex was finally resuspended in 100 µL of binding buffer, and a 30 µL aliquot was analyzed by Western Blotting analysis with anti-His-tag antibody. The remaining 70 µL was quantitated by liquid scintillation.
ITC experiments on the interaction of SAM with nsp16, nsp10, and nsp16/nsp10 complex of SARS-CoV in the absence of DTT were carried out at 25°C using a VP-ITC titration calorimeter (MicroCal, Northampton, MA). Freshly purified nsp16 and nsp10 proteins were mixed, and their final concentrations were 10 µM and 80 µM, respectively. Then nsp16, nsp10, and nsp16/nsp10 complex were dialyzed against 40 mM Tris-HCl buffer (pH 8.0) containing 50 mM NaCl, over night at 4°C, extensively to remove glycerol and DTT. A solution of nsp16, nsp10, or nsp16/nsp10 complex was loaded into the sample cell (1.43 mL), and a solution of 150 µM SAM was placed in the injection syringe (300 µL). The first injection (5 µL) was followed by 29 injections of 10 µL. Dilution heats of SAM were measured by injecting SAM solution into buffer alone and were subtracted from the experimental curves prior to data analysis.
The interaction of nsp16, nsp10, and nsp16/nsp10 complex of SARS-CoV with m7GpppA-capped RNA in the absence of DTT were carried out at 25°C using an ITC200 titration calorimeter (MicroCal, Northampton, MA), which has higher sensitivity than the equipment adopted for analyzing protein-SAM binding as described above. A solution of m7GpppA-capped RNA (7 µM) was loaded into the sample cell (500 µL), and a solution of 100 µM purified proteins were placed in the injection syringe (40 µL). Dilution heats of purified proteins were measured by injecting purified proteins solution into buffer alone and were subtracted from the experimental curves prior to data analysis.
The resulting data were fitted to a single set of identical sites model using MicroCal ORIGIN software supplied with the instrument, and the binding stoichiometry, N, the standard molar enthalpy change for the binding, , and the dissociation constant, Kd, were thus obtained. The standard molar free energy change, , and the standard molar entropy change, , for the binding reaction were calculated by the fundamental equations of thermodynamics: = ; = (-)/T.
The GenBank accession numbers for genes and proteins mentioned in the text are as follow: SARS coronavirus nsp3 unique domain (SUD), JN247391; SARS coronavirus nonstructural protein nsp5, JN247392; SARS coronavirus nonstructural protein nsp10, JN247393; SARS coronavirus nsp10-nsp11 fusion protein, JN247394; SARS coronavirus nsp12 N-terminal domain (nsp12N), JN247395; SARS coronavirus nonstructural protein nsp14, JN247396; SARS coronavirus nonstructural protein nsp16, JN247397.
|
10.1371/journal.pgen.1008296 | SweC and SweD are essential co-factors of the FtsEX-CwlO cell wall hydrolase complex in Bacillus subtilis | The peptidoglycan (PG) sacculus is composed of long glycan strands cross-linked together by short peptides forming a covalently closed meshwork that protects the bacterial cell from osmotic lysis and specifies its shape. PG hydrolases play essential roles in remodeling this three-dimensional network during growth and division but how these autolytic enzymes are regulated remains poorly understood. The FtsEX ABC transporter-like complex has emerged as a broadly conserved regulatory module in controlling cell wall hydrolases in diverse bacterial species. In most characterized examples, this complex regulates distinct PG hydrolases involved in cell division and is intimately associated with the cytokinetic machinery called the divisome. However, in the gram-positive bacterium Bacillus subtilis the FtsEX complex is required for cell wall elongation where it regulates the PG hydrolase CwlO that acts along the lateral cell wall. To investigate whether additional factors are required for FtsEX function outside the divisome, we performed a synthetic lethal screen taking advantage of the conditional essentiality of CwlO. This screen identified two uncharacterized factors (SweD and SweC) that are required for CwlO activity. We demonstrate that these proteins reside in a membrane complex with FtsX and that amino acid substitutions in residues adjacent to the ATPase domain of FtsE partially bypass the requirement for them. Collectively our data indicate that SweD and SweC function as essential co-factors of FtsEX in controlling CwlO during cell wall elongation. We propose that factors analogous to SweDC function to support FtsEX activity outside the divisome in other bacteria.
| Bacterial growth and division require the synthesis and remodeling of the cell wall exoskeleton. To prevent lethal breaches in this protective layer, peptidoglycan (PG) hydrolases that remodel the cell wall must be carefully regulated but the mechanisms underlying this control remain poorly understood. The noncanonical ABC transporter FtsEX has emerged as a broadly conserved regulator of PG hydrolases. In most characterized examples, FtsEX is integrated into the division machinery where it controls cell wall cleavage during cytokinesis. By contrast, in Bacillus subtilis the FtsEX complex functions in cell wall elongation. Here, we report the identification of two previously uncharacterized proteins (SweD and SweC) that function as essential co-factors of FtsEX in controlling PG hydrolase activity along the lateral cell wall. Homologs of SweD and SweC are found in a subset of firmicutes. We propose that these and analogous factors enable FtsEX to function outside the divisome to control cell wall elongation hydrolases.
| Most bacteria are encased within a cell wall exoskeleton composed of the heteropolymer peptidoglycan (PG). This macromolecule is assembled from long glycan strands cross-linked together by attached peptides, generating a continuous three-dimensional meshwork that encapsulates the cytoplasmic membrane, specifies cell shape, and protects the cell from its internal turgor pressure [1, 2]. Bacterial growth and division are intimately linked to hydrolysis of this covalently closed exoskeleton. To enlarge this meshwork during growth, bonds connecting the glycan strands must be broken to allow expansion of the meshwork and/or to incorporate new strands between the existing ones [3–5]. Similarly, during cell division, bonds must be broken in the nascent septal PG to allow invagination of the outer membrane in gram-negative bacteria and to promote cell separation in gram-positive bacteria. PG hydrolases play a central role in these processes, but their activities must be tightly regulated to prevent excessive degradation of the cell wall and the generation of lethal breaches in this protective layer. Many of the enzymes responsible for cell growth and division have been identified in a growing number of bacteria [6–20]. However, the mechanisms by which they are regulated remain incompletely understood. A deeper understanding of these regulatory systems has the potential to reveal new ways to subvert PG biogenesis for therapeutic intervention [4, 21].
Progress in our understanding of the control of PG hydrolysis comes from studies of the broadly conserved FtsEX complex. This substrate-less ABC transporter has been found to regulate distinct PG hydrolases in diverse bacterial species including E. coli, Streptococcus pneumoniae, B. subtilis, Bacillus anthracis, Mycobacterium tuberculosis, and Caulobacter crescentus [21–27]. FtsEX is a member of the Type VII ABC transporter superfamily that is thought to function in mechanotransmission rather than as transporters [28]. FtsE is the ATPase and FtsX is the transmembrane domain subunit of the complex. The large extracellular loops of FtsX interact with species-specific PG hydrolases that contain regulatory coiled-coil domains or, in the case of E. coli, an activator of PG hydrolases with a regulatory coiled-coil domain. Structural studies of the S. pneumoniae PG hydrolase PcsB that is controlled by FtsEX [26, 29] suggest that the coiled-coil domain resembles long molecular tweezers that hold the globular PG hydrolase domain in an inactive state [30]. Based on structural analysis of MacB, another member of the Type VII ABC transporter superfamily [28], FtsEX is thought to function by a mechanotransmission mechanism in which the ATPase cycle of FtsE controls conformational changes in the extracellular loop domains of FtsX triggering PG hydrolase activity perhaps by releasing the catalytic domain from the clutches of its regulatory coiled-coil domain.
In most bacteria in which this conserved regulatory module has been examined, FtsEX functions in the context of cell division. In the case of E. coli, FtsEX controls the activity of two PG amidases (AmiA and AmiB) via the coiled-coil domain-containing regulator EnvC [27, 31]. PG hydrolysis by these enzymes allows invagination of the outer membrane through the nascent septal PG layer during cytokinesis. FtsEX is intimately linked to divisome function. The complex is recruited to the cytokinetic ring at an early step in its assembly and recent work indicates that it interacts with the actin-like division protein FtsA to promote recruitment of downstream division factors [32, 33]. Furthermore, ATP hydrolysis by the FtsEX complex is not only required for cell wall hydrolysis by the two amidases it is also necessary for septal PG synthesis [32, 34] suggesting that FtsEX functions as a central coordinator of these two processes. FtsEX in C. crescentus has similarly been implicated in linking PG synthesis and its remodeling during cytokinesis [24]. In S. pneumoniae, FtsEX controls the PG hydrolase PcsB during cell division [26]. The mechanism by which FtsEX is recruited to the divisome is not known but, by analogy to E. coli, is thought to involve interactions with divisome components like FtsA or FtsZ [26]. The signal(s) that stimulate cycles of ATP hydrolysis in any of these systems are currently unknown but are likely intimately linked to the onset of PG synthesis during cytokinesis.
Interestingly, in B. subtilis, FtsEX is not involved in cell division but instead controls the PG hydrolase CwlO that plays a central role in cell wall elongation during growth [22, 25]. CwlO contains an amino-terminal coiled-coil domain followed by a NlpC/P60 D,L-endopeptidase domain that cleaves the bond between γ-D-glutamate and meso-diaminopimelic acid within the stem peptide [35]. CwlO and a second D,L-endopeptidase (LytE) [36] that is not regulated by FtsEX, are functionally redundant [6, 13]. Cells lacking either of these enzymes are viable but depletion of one in the absence of the other results in a lethal block to cell wall elongation. Consistent with the idea that CwlO is controlled by FtsEX, cells lacking either FtsE or FtsX or harboring point mutations in the putative ATPase domain of FtsE are blocked for cell wall elongation upon depletion of LytE [22, 25]. As in the case of the division-associated FtsEX complexes, the signals that stimulate FtsEX activity during growth are currently unknown. Here, we investigate whether other factors are required for FtsEX to function outside the multi-protein divisome complex. Taking advantage of the functional redundancy of CwlO and LytE we performed a synthetic lethal screen for mutants that require lytE for growth and identified two uncharacterized genes yqzD (sweD) and yqzC (sweC) that function in the same genetic pathway as ftsEX and cwlO. We demonstrate that SweD and SweC reside in a multimeric membrane complex with FtsX and function as essential co-factors of FtsEX in the control of CwlO activity. Orthologs of SweD and SweC are present in a subset of Bacilliaceae, Lactobacillales, and Listeriaceae suggesting these factors similarly enable FtsEX to control elongation hydrolases. Interestingly, Bernhardt and co-workers have identified two unrelated membrane proteins in Corynebacterium glutamicum that function as co-factors of the FtsEX-RipC cell wall hydrolase complex to promote cell separation in this organism. Thus, FtsEX complexes employ distinct co-factors to regulate PG hydrolases during cell division, separation, and elongation.
To identify additional factors required for FtsEX-CwlO function, we took advantage of the synthetic lethal relationship between cwlO and lytE. Using transposon-sequencing (Tn-Seq) [37] we screened for genes that could tolerate transposon insertions in wild-type (LytE+) B. subtilis but not in cells lacking lytE. Such genes are predicted to function in a genetic pathway with ftsEX and cwlO. Mariner transposon libraries with >100,000 unique insertions were generated in wild-type B. subtilis PY79 and an isogenic ΔlytE variant. The two libraries were separately pooled and the transposon-chromosome junctions were mapped by massively parallel DNA sequencing. As anticipated and in validation of our screen, transposon insertions in cwlO, ftsE, and ftsX were readily detected in the wild-type library but were virtually undetectable in the library generated in the ΔlytE mutant (Fig 1A). In addition to these positive controls we identified two genes (yqzD and yqzC) of unknown function that were statistically (P<0.05 Mann-Whiney U test) underrepresented in the ΔlytE library compared to wild-type (Fig 1A). Based on the experiments presented below we have renamed these genes sweD and sweC for synthetic lethal with LytE.
The conditional essentiality of sweD and sweC was confirmed by depleting LytE in cells lacking either of these new factors. Fig 1B shows that depletion of LytE in ΔsweD, ΔsweC, or ΔcwlO mutants does not support colony formation. Similarly, cells lacking LytE that were depleted of SweD and SweC were inviable (Fig 1B). Consistent with the idea that sweD and sweC specifically function in a genetic pathway with ftsEX and cwlO, there was no plating defect in the ΔcwlO and ΔftsEX mutants upon depletion of SweD and/or SweC (Fig 1B and S1A Fig.).
The sweD and sweC genes reside in an operon and are predicted to encode proteins of 13 kDa and 16 kDa, respectively (Fig 1C). Both proteins are predicted to contain N-terminal transmembrane (TM) segments. SweD contains a predicted coiled-coil (CC) domain followed by a C-terminal helix-turn-helix (HTH) motif (Fig 1C). The C-terminal region of SweC has remote homology to LysM domains that bind polysaccharides including chitin and peptidoglycan. Homologs of SweD and SweC are present in a subset of Bacilliaceae, Listeriaceae, and Lactobacillales family members (S2 Fig). PSI-BLAST also identified potential SweD and SweC homologs in the gram-negative bacterium Helicobacter pylori and a small group of unculturable bacteria that are similarly not firmicutes.
To investigate whether sweD and sweC are in the same genetic pathway as cwlO and ftsEX, we analyzed the cytological phenotypes of the mutants by fluorescence microscopy. Cells lacking CwlO or FtsEX (or both) are shorter and fatter than wild-type and often slightly curved or bent [6, 13, 25, 38]. Accordingly, we directly compared the morphologies of the ΔsweDC mutant to cells lacking CwlO or FtsEX. The cells were grown in defined rich (CH) medium and analyzed by fluorescence microscopy using the fluorescent membrane dye TMA-DPH. The majority of the ΔsweDC mutant cells were shorter and fatter than wild-type (Fig 2A). However, cells lacking sweDC appeared even more bent and curved than the ΔcwlO and ΔftsEX mutants. Importantly, these phenotypes were lost in cells that lacked both SweDC and CwlO, suggesting that these morphological defects require CwlO (Fig 2A). Similar phenotypes were observed in the ΔsweC and ΔsweD single mutants (S1B Fig), but as shown below the two proteins depend upon each other for stability, making it difficult to assign specific functions to either factor. Finally, as reported previously [25], cells lacking the second elongation PG hydrolase LytE resembled wild-type (Fig 2A).
A ΔlytE mutant depleted of CwlO or FtsEX is impaired in elongation but remains capable of cell division, generating short fat cells that ultimately lyse. To further explore the role of SweDC function in cell elongation, we examined the morphological defects in ΔlytE cells upon depletion of SweDC (Fig 2B). Cells lacking LytE and harboring an IPTG-regulated sweDC allele were grown in the presence of inducer in rich medium. Cytoplasmic mCherry and the membrane dye TMA-DPH were visualized by fluorescence microscopy before and at 30-minute intervals after removal of IPTG. As can be seen in Fig 2B, 150 minutes after IPTG removal, cell length was slightly reduced and this reduction continued over the next 60 minutes followed by the onset of lysis. Quantitative image analysis revealed a ~30% reduction in cell length over the depletion time course, similar to what was observed in a ΔlytE mutant depleted of CwlO (S3 Fig) [25]. Consistent with the curved and bent phenotypes observed in the LytE+ ΔsweDC mutant (Fig 2A), depletion of SweD and SweC in the absence of LytE resulted in even more dramatic curved morphologies prior to lysis (Fig 2B). A comparison of the terminal phenotypes of cells depleted for SweDC, CwlO and FtsEX in a ΔlytE mutant (Fig 2C) revealed that all three depletions generate shorter cells prior to lysis, however the dramatic curved morphologies were only observed when SweDC was depleted. Importantly, this phenotype was suppressed when LytE was depleted in cells in which both CwlO and SweDC were absent (S4 Fig). Altogether, these results argue that cwlO, ftsEX and sweDC reside in the same genetic pathway, and support the idea that SweD and SweC are required for PG hydrolase activity of CwlO.
To verify that SweD and SweC are integral membrane proteins and investigate their topologies, we raised antibodies against the soluble domains of both factors and analyzed the proteins by subcellular fractionation. Protoplasts were generated from exponentially growing wild-type cells and then lysed by addition of hypotonic buffer. The lysate was subjected to ultracentrifugation and the membrane fraction was homogenized with buffer in the presence and absence of the nonionic detergent TritonX-100 followed by a second round of centrifugation. As expected for integral membrane proteins, SweC and SweD fractionated with the membranes and could be solubilized with TritonX-100 (Fig 3A). The integral membrane protein EzrA [39] and cytoplasmic protein ScpB [40] served as positive and negative controls for this analysis.
Next, we investigated whether the soluble domains of SweC and SweD reside in the cytosol or on the extracellular face of the membrane. Membrane topology prediction programs [41, 42] yielded equivocal results for both proteins. We performed a protease accessibility assay using a B. subtilis strain engineered to express the sporulation membrane protein SpoIVFA (FA) [43]. FA is an integral membrane protein with a large extracellular domain that is accessible to protease cleavage and served as our positive control. Protoplasts were generated from an exponentially growing culture and then treated with buffer, Proteinase K, or Proteinase K and the detergent N-lauroylsarcosine. As anticipated, FA was accessible to Proteinase K resulting in the loss of its extracellular C-terminal domain, which is recognized by our anti-FA antibody (Fig 3B). Consistent with the idea that SweC and SweD are Type II integral membrane proteins with C-terminal intracellular domains, both proteins were inaccessible to protease degradation in protoplasts (Fig 3B). The Type II integral membrane protein EzrA and the cytoplasmic protein FtsE served as protease inaccessible controls and were both resistant to Proteinase K in protoplasts. Importantly all four proteins were efficiently degraded by Proteinase K when detergent was included in the reaction. Taken together, these data indicate that the putative coiled-coil and helix-turn-helix domains of SweD and the LysM-like domain on SweC are located in the cytoplasm.
Proteins that reside in complexes sometimes depend upon each other for stability. Accordingly, to explore whether SweD and SweC interact with each other or with FtsEX, we analyzed whether SweD, SweC, FtsE, FtsX, or CwlO depend on each other for stability. To this end, we compared the levels of all five proteins in wild-type and in strains lacking SweD, SweC, FtsEX, or CwlO. Cells lacking SweD had almost undetectable levels of SweC and reciprocally cells lacking SweC had barely detectable levels of SweD (Fig 4A). Expression of either gene in trans restored the levels of both proteins, indicating that the effects were not due to polarity or changes in mRNA stability (Fig 4A). By contrast, the levels of CwlO and FtsEX were unaffected by the absence of SweD/SweC and SweD/SweC levels were unchanged in the absence of either CwlO or FtsEX. These data are consistent with the idea that SweD and SweC interact with each other and in the absence of either one, the other is susceptible to degradation.
We previously reported that cell association of CwlO does not depend on FtsEX leading us to speculate that an additional factor holds CwlO at the cell surface [25]. This observation, in part, motivated the Tn-Seq screen that identified SweD and SweC. However, we have since discovered that CwlO non-specifically binds to plastic microfuge tubes confounding our original analysis. Using a modified protocol that accounts for this (see Methods) we could detect a decrease in cell-associated (C) CwlO and an increase of the protein in the cultured medium (M) in cells lacking FtsEX compared to wild-type (Fig 4B). These data and complementary analysis by Errington and co-workers [22] support the idea that FtsX maintains CwlO at the cell surface as has been observed in other bacteria [26, 27, 29]. Consistent with the topologies of SweD and SweC and the dependencies described above, the amount of surface-associated CwlO was similar in the presence and absence of SweD and SweC (Fig 4B).
To investigate the contribution of the intracellular domains on SweC and SweD to CwlO activity, a series of domain deletions were generated and tested for their ability to support growth in a LytE depletion strain. Strains lacking either the putative coiled-coil (CC) or HTH domain of SweD largely phenocopied the ΔsweD null (Fig 5A and 5B) suggesting both domains are critical for function. A strain lacking the LysM homology domain on SweC was impaired for growth upon LytE depletion with a >100-fold plating defect on LB agar (Fig 5A and 5B). On defined rich (CH) medium the SweC(ΔLysM) truncation was viable upon LytE depletion albeit with a small colony phenotype. To establish whether these deletion variants were stably produced in vivo, we monitored their levels by immunoblot (Fig 5C). Since our antibodies were raised against the soluble intracellular domains of SweD and SweC the immunoblots likely underestimate the levels of these truncations. However, because SweC and SweD require each other for stability we also used the level of the unmodified protein as a proxy for their stability. Based on this analysis, the variants appeared to be produced at levels similar to wild-type with the exception of SweD(ΔCC), which was somewhat reduced (Fig 5C). These data indicate that the HTH and CC domains of SweD and the LysM-like domain of SweC are important for function with SweD domains playing more critical roles. Furthermore, these data suggest that if SweD and SweC stabilize each other through interaction then they likely interact via their transmembrane segments.
Finally, we analyzed a SweD variant with mutations in the second helix of the HTH domain that in other HTH DNA binding proteins is involved in interaction with DNA [44]. The mutant was stably produced and maintained SweC levels but was impaired for function (Fig 5). ChIP-seq analysis to determine whether SweD binds DNA failed to identify specific binding sites. Specifically, the ratio of sequencing reads between wild-type and ΔsweDC mutant samples was ~1 across the entire genome (S5 Fig). These data indicate the HTH is important for function but probably not for DNA binding.
To gain additional insight into the role of SweDC in cell wall elongation, we sought to identify suppressors of the ΔsweDC ΔlytE double mutant. To this end, we used the LytE depletion strain to screen for conditions in which the double mutant was viable. Mutants with impair cell wall biogenesis are often partially suppressed by growth in hypertonic medium (0.25 M sucrose) supplemented with Mg2+ and these conditions were similarly suppressive for cells lacking SweDC and depleted of LytE (Fig 6A). The restoration of viability under these conditions is likely due to residual activity of FtsEX-CwlO in the absence of SweDC, because these conditions did not support growth of the ΔcwlO ΔlytE or ΔftsEX ΔlytE double mutants (S6 Fig). Next, we generated the ΔsweDC ΔlytE double mutant in the absence of the IPTG-regulated lytE allele using the permissive growth condition. We then grew the cells under permissive conditions and selected for suppressors on LB agar medium. Thirteen independently isolated suppressors were mapped by whole genome re-sequencing (S7A Fig). Eight of the suppressors had loss-of-function mutations in walH (yycH) encoding a negative regulator of the WalR-WalK two-component signaling pathway [45]. WalR~P is a positive regulator of several PG hydrolases genes including lytE and cwlO and represses the expression of genes [6, 46, 47] that encode negative regulators of PG hydrolase activity [48, 49]. Among the eight walH suppressors, five had second-site missense mutations in either ftsE or ftsX (S7A Fig). The two mutations in ftsX mapped to the first (S26Y) and second (S188F) TM segments of the protein (S7B Fig). Two mutations in ftsE (V176F and T188A) were located in positions predicted to lie adjacent to the nucleotide-binding domain (S7B Fig). The third mutation (L90F) was close to the predicted interface between FtsE and FtsX. Three suppressors had missense or small in-frame deletions in walK encoding the WalK sensor kinase. One of these mutations has been previously reported to generate a constitutively active allele of an unrelated sensor kinase [50] and we suspect that all three suppressors are hypermorphic alleles. Finally, we identified two independent mutations in rny encoding RNase Y, a major endonuclease involved in mRNA decay [51, 52].
The isolations of suppressor mutations in ftsE and ftsX provided a potential link between SweDC and FtsEX. Accordingly, we focused on these suppressors. To test whether these mutations contributed to the suppression of the ΔsweDC ΔlytE double mutant, we reconstructed ftsE(V176F) and ftsX(S26Y) alleles and tested them alone or in combination with a walH deletion mutant. As can be seen in Fig 6A, both the walH deletion and the ftsEX point mutations could suppress the lethality of ΔsweDC ΔlytE on defined (CH) rich medium but neither were able to support growth on LB. However, combining ΔwalH with either ftsEX allele provided suppression on LB.
Next, we analyzed the morphologies of the suppressors. Depletion of LytE in the absence of SweDC leads to cell lysis in CH medium (Fig 6B and S8B Fig). Under permissive growth conditions in the presence of sucrose and Mg2+ the cells are fat and curved (Fig 6B and S8A Fig). These phenotypes are likely due to residual FtsEX-CwlO activity in the absence of SweDC as we observe similar morphologies in a ΔlytE mutant expressing low levels of CwlO when grown under the same permissive condition (S9 Fig). We suspect that low-level PG hydrolysis by CwlO results in uneven cell wall elongation leading to the curved cell shape. Importantly, the ftsE and the ftsX suppressor mutations restored rod-shaped morphologies to the ΔsweDC ΔlytE double mutant on CH medium supplemented with sucrose and Mg2+ and partially suppressed the curved morphologies on CH medium without supplementation (Fig 6B). The ΔwalH mutant also partially restored rod-shape morphology to the ΔsweDC ΔlytE double mutant on CH medium supplemented with sucrose and Mg2+ (S8B Fig). Combining the ftsEX mutations with ΔwalH in the ΔsweDC ΔlytE background further restored wild-type-like morphologies on CH medium with and without supplementation (S8 Fig). Collectively, these data support the idea that SweD and SweC function as co-factors of FtsEX in the control of CwlO. The molecular basis for the partial suppression of the ΔsweDC ΔlytE double mutant by ΔwalH remains unclear but could be due to the increase in CwlO levels and/or the global change in cell wall hydrolase activity resulting from high WalR~P activity.
The genetic evidence presented thus far place SweD and SweC in the FtsEX-CwlO pathway. To investigate whether SweD or SweC reside in a complex with FtsX, we performed co-immunoprecipitation assays. Crude membrane preparations from wild-type and ΔftsX mutant cells were solubilized with the non-ionic detergent Digitonin. The solubilized membrane proteins were incubated with anti-FtsX polyclonal antiserum and precipitated by Protein A sepharose. The immunoprecipitated material was eluted with SDS sample buffer and analyzed by immunoblot. As can be seen in Fig 7A, the anti-FtsX antibodies efficiently immunoprecipitated FtsX and co-precipitated SweD and SweC but not an unrelated membrane protein WalI (YycI). Importantly, none of these proteins were precipitated from solubilized membrane preparations that lacked FtsX. Thus, these data indicate that SweD and SweC reside in a membrane complex with FtsX.
To investigate whether FtsE or FtsX interact directly with either of these factors we used the Bacterial Adenylate Cyclase Two Hybrid (BACTH) system. We generated fusions with complementary fragments (T18 and T25) of Bordetella pertussis adenylate cyclase to FtsE, FtsX, SweD and SweC. As anticipated, we detected positive interactions between FtsE and FtsX with several but not all fusions and with the positive control TolB and Pal (Fig 7B). A weak interaction was also observed between TolB and FtsE. This false positive highlights the limitations of the two-hybrid assay and the importance of interpreting positive interactions cautiously. The assay also revealed positive interactions between FtsX and SweD with two distinct fusion pairs and an interaction between SweD and itself. Furthermore, a SweD variant lacking its coiled-coil (CC) domain retained the ability to interact with FtsX but not with full-length SweD (Fig 7C), suggesting that the CC domain functions in SweD dimerization. Finally, a SweD variant lacking its entire intracellular domain [SweD(TM)] was unaffected in its ability to interact with FtsX (Fig 7C), suggesting that SweD interacts with FtsX via its TM segment. Despite our in vivo data that SweC and SweD depend on each other for stability and our co-immunoprecipitation assay that identified SweC in a complex with FtsX, we were unable to detect an interaction between SweC and either of these two proteins in the two-hybrid assay.
Altogether our data indicate that SweD and SweC reside in a multimeric membrane complex with FtsX and function as essential co-factors of FtsEX in its control of CwlO activity during cell wall elongation (Fig 8). Ribosome profiling from exponentially growing B. subtilis cells suggests that the levels of SweC are similar to FtsE and FtsX while SweD levels are ~2-fold higher [53]. Based on these data and our two-hybrid analysis, we hypothesize that the SweD-SweC-FtsE-FtsX complex has a stoichiometry of 4-2-2-2. Finally, our data indicate that both SweD and SweC have conserved intracellular domains that are important for function and could therefore play roles in maintaining or regulating the ATPase activity of FtsE; directly sensing a signal to modulate cleavage activity; and/or linking the PG hydrolase complex to PG synthesis machinery. The identification of suppressor mutations adjacent to the ATPase domain of FtsE that partially bypass the requirement for SweDC suggests that these co-factors could indeed help maintain and/or control cycles of ATP hydrolysis. However, attempts to reconstitute ATPase activity of FtsE in vitro have thus far been unsuccessful making it difficult to directly test this model. It is also noteworthy that our data indicate that the LysM-like domain on SweC resides in the cytosol rather than facing the cell wall. Since these domains bind polysaccharides, an attractive model is that this domain functions to coordinate PG hydrolase activity with cell wall synthetic capacity by monitoring cytosolic PG precursors. In vitro reconstitution of the complex in the future will allow us to explore this and related models for SweDC function.
A connection between the FtsEX-CwlO PG hydrolase complex and the cell wall elongation machinery (called the Rod complex) has been proposed previously [22] but the evidence remains incomplete. Cells lacking LytE but not FtsEX or CwlO were found to have a synthetic growth defect when the actin-like protein Mbl was depleted. Since Mbl functions as a scaffold for the Rod complex, these data suggest that CwlO, FtsEX, Mbl and by extension the rest of the cell wall elongation machinery are in the same genetic pathway [22]. We detect similar growth defects when Mbl is depleted in a ΔlytE mutant and the absence of significant growth defects when Mbl is depleted in strains lacking FtsEX, CwlO, or SweDC (S10 Fig). However, we found that a functional SweD-GFP fusion (S11 Fig) lacked the dynamic behavior of the Rod complex, which was shown to move in a directed and circumferential manner around the long axis of cell [54–56]. These data suggest that the SweDC-FtsEX-CwlO complex is unlikely to be a core component of the cell wall synthetic machinery but does not exclude the possibility that the two complexes transiently interact and influence each other's activity. Consistent with this idea, formaldehyde crosslinking of B. subtilis cells followed by affinity purification of Mbl and Mass Spectrometry identified FtsX and SweC in crosslinked complexes with Mbl and separately FtsE and FtsX in crosslinked complexes with MreB [57]. However, >65 proteins were identified in each of the crosslinked complexes making it difficult to draw strong conclusions from these findings. Finally, we note that the integral membrane protein RodZ, a core member of the Rod complex, has an intracellular HTH motif like SweD [58–60]. In the case of RodZ, this motif functions to position an adjacent helix such that it can interact with MreB [61]. Although we have shown that the HTH motif on SweD is important for function this domain lacks the analogous interaction helix found in RodZ. Furthermore, we did not observe an interaction between Mbl and SweD in the bacterial two-hybrid assay. Thus, it remains an open question whether the PG hydrolase activity of the SweDC-FtsEX-CwlO complex is coordinated with cell wall synthesis mediated by the Rod complex.
We note that gram-positive bacteria like Bacillus subtilis are thought to synthesize their cell wall via an inside to outside mechanism in which newly synthesized layers adjacent to the cell membrane are not initially load bearing [3]. Only as these layers migrate outward during cell growth do they experience stress. It is these more distal load-bearing layers that are likely to be the target of CwlO and LytE. Thus, the activity of these D,L-endopeptidases would not necessarily need to be directly coordinated with PG synthesis as has been proposed for gram-negative bacteria [62].
Finally, we return to FtsEX. This noncanonical ABC transporter is the most broadly conserved regulator of PG hydrolases. In many cases this complex is intimately associated with the divisome where it controls cell wall cleavage during cytokinesis and potentially coordinates PG hydrolysis with septal PG synthesis (Fig 8). Here, we establish that two co-factors SweD and SweC are critical for FtsEX function outside the divisome during cell wall elongation. Homologs of SweD and SweC are present in Bacilliaceae, Lactobacillales, and Listeriaceae species, where we propose that, like the B. subtilis co-factors, they enable FtsEX to control elongation PG hydrolases. Interestingly, work from the Bernhardt lab has uncovered a distinct set of factors that function with FtsEX in Corynebacterium glutamicum, an actinobacterium that undergoes cell separation through a mechanically driven process called V snapping [88]. These two membrane proteins (SteA and SteB) are required for the FtsEX-RipC PG hydrolase complex to promote for V snapping. Thus, the identification of SweDC and SteAB reveal that the broadly conserved FtsEX regulatory module works with distinct co-factors to control PG hydrolases required for cell division, cell separation, and cell wall elongation.
All B. subtilis strains were derived from the prototrophic strain PY79 [63]. Unless otherwise indicated, cells were grown in LB or defined rich (casein hydrolysate, CH) medium at 37°C. Insertion-deletion mutations were generated by isothermal assembly [64] of PCR products followed by direct transformation into B. subtilis. Tables of strains, plasmids and oligonucleotide primers and a description of strain and plasmid construction can be found online as supplementary data (S1, S2 and S3 Tables, and S1 Text).
Transposon insertion sequencing (Tn-seq) was performed as described previously [65–67]. Libraries of >100,000 independent transposants were separately generated in wild-type and a ΔlytE mutant. Genomic DNA was extracted from each and digested with MmeI, followed by adapter ligation. Transposon-chromosome junctions were amplified by PCR (17 amplification cycles). PCR products were pooled, gel-purified, and sequenced on the Illumina HiSeq platform using TruSeq reagents (Tufts University TUCF Genomics facility). Reads were mapped to the B. subtilis 168 genome (NCBI NC_000964.3), tallied at each TA site, and genes in which reads were statistically underrepresented were identified using the Mann Whitney U test and by visual inspection using Sanger Artemis Genome Browser and Annotation tool [68].
Immunoblot analysis was performed as described previously [69]. Briefly, 1ml of culture was collected and resuspended in lysis buffer [20 mM Tris pH 7.0, 10mM MgCl2 and 1mM EDTA, 1 mg/ml lysozyme, 10 μg/ml DNase I, 100 μg/ml RNase A, 1 mM PMSF, 1 μg/ml leupeptin, 1 μg/ml pepstatin] to a final OD600 of 10 for equivalent loading. The cells were incubated at 37°C for 10 min followed by addition of an equal volume of sodium dodecyl sulfate (SDS) sample buffer [0.25 M Tris pH 6.8, 4% SDS, 20% glycerol, 10 mM EDTA] containing 10% 2-Mercaptoethanol. Samples were heated for 15 min at 65°C prior to loading. Proteins were separated by SDS-PAGE on 10% (for SMC), 12.5% (for FtsE, FtsX, CwlO, SigA, EzrA, ScpB, SpoIVFA, and WalI) or 20% (for SweD and SweC) polyacrylamide gels, electroblotted onto Immobilon-P membranes (Millipore) and blocked in 5% nonfat milk in phosphate-buffered saline (PBS) with 0.5% Tween-20. The blocked membranes were probed with anti-SweD (1:10,000), anti-SweC (1:10,000), anti-EzrA (1:10,000) [39], anti-SpoIVFA (1:10,000) [70], anti-SMC (1:10,000) [71], anti-SigA (1:10,000) [72], anti-ScpB (1:10,000) [73], anti-FtsE (1:20,000) [25], anti-FtsX (1:10,000) [25], anti-CwlO (1:10,000) [25], anti-WalI (YycI) [74] diluted into 3% BSA in 1x PBS with 0.05% Tween-20. Primary antibodies were detected using horseradish peroxidase-conjugated goat anti-rabbit IgG (BioRad) and the Super Signal chemiluminescence reagent as described by the manufacturer (Pierce). Signal was detected using a Bio-Techne FluorChem R System.
Fluorescence microscopy was performed on a Nikon Ti microscope equipped with Plan Apo 100x/1.4NA phase contrast oil objective and a CoolSnapHQ2 camera. Cells were immobilized using 1.5% agarose pads containing CH medium. Membranes were stained with TMA-DPH (50μM) (Molecular Probes). Exposure times were 400 ms and 800 ms for TMA-DPH and mCherry, respectively. Quantitative image analysis was performed using Oufti [75]. Meshes were created using the cytoplasmic mCherry images. The length of the long axis of >350 cells was determined and the mean cell length was calculated. Images were cropped and adjusted using MetaMorph software (Molecular Devices). Final figures were prepared in Microsoft PowerPoint.
Recombinant proteins were expressed in E. coli strain BL21 (DE3). Strains were grown in 500 mL of auto-induction Overnight Express Instant TB Medium (Novagen) supplemented with 100 μg/ml ampicillin at 22°C. After 16 h, cultures were subjected to centrifugation at 10,000 × g for 10 min. Cell pellets were resuspended in 15mL Lysis Buffer (20 mM Tris pH 7.5, 300 mM NaCl, 5 mM imidazole, 10% glycerol, 0.1 μM Dithiothreitol) and Complete EDTA-free protease inhibitors (Roche) and lysed via passage through a French press. Cell lysates were clarified by centrifugation at 10,000 X g for 10 minutes at 4°C. Clarified lysates were mixed with 0.5 mL of Ni-NTA agarose resin (Qiagen) and incubated for 2 hours at 4°C. The mixture was loaded onto a BioRad column, washed with 10 mL Buffer A (20 mM Tris pH 7.5, 300 mM NaCl, 5 mM imidazole, 10% glycerol, 0.1 μM Dithiothreitol). His6-SUMO fusion proteins were eluted with Buffer B (20 mM Tris pH 7.5, 300 mM NaCl, 200 mM imidazole, 0.1 μM Dithiothreitol). Eluates were pooled and dialyzed into 20 mM Tris pH 7.5, 300 mM NaCl, 10% glycerol, 0.1 μM Dithiothreitol at 4°C. The dialysates were incubated with His6-Ulp1 protease overnight on ice. Reactions were mixed with 0.5 mL Ni-NTA agarose and loaded onto BioRad columns. Flow through fractions containing cleaved (untagged) proteins were collected and used to generate rabbit polyclonal antibodies (Covance).
25 mL of exponentially growing cells were collected, washed, and resuspended in 5mL 1X SMM buffer (0.5 M sucrose, 20 mM MgCl2, 20 mM maleic acid pH 6.5) [76] supplemented with Lysozyme (4 mg/ml). Cells were incubated for 30 minutes at RT with gentle agitation. Protoplast formation was monitored by microscopic observation on 2% 1X SMM-agarose pads. When >95% of the cells were protoplasted they were collected by centrifugation (5 Krpm) and resuspended in 1 mL of 1X SMM buffer and distributed into three microfuge tubes. Protoplasts were incubated with: 1X SMM buffer alone or with Proteinase K (NEB, 50 ug/ml final), or Proteinase K and Sodium-lauroyl-sarcosinate (1%) for 15 min at room temperature. Proteinase K was inactivated by the addition of 2X sample buffer [0.25 M Tris pH 6.8, 4% SDS, 20% glycerol, 10 mM EDTA 10% 2-Mercaptoethanol] supplemented with 2 mM PMSF and immediately incubated at 100°C for 10 minutes. Reactions were analyzed by immunoblot.
CwlO binds non-specifically to plastic microfuge tubes but can be recovered from them with SDS sample buffer. To monitor the amount of cell-associated and released CwlO, 1.5 mL of a mid-exponential phase culture (grown in LB) was collected in a plastic microfuge tube and centrifuged for 10 min at 3,000 × g. The culture supernatant was transferred to a fresh microfuge tube and the secreted proteins were precipitated by the addition of 15% trichloroacetic acid. The cell pellet was gently resuspended in 50 μL lysis buffer [20 mM Tris pH 7.0, 10mM MgCl2 and 1mM EDTA, 10 μg/ml DNase I, 100 μg/ml RNase A, 1 mM PMSF, 1 μg/ml leupeptin, 1 μg/ml pepstatin] and transferred to a fresh microfuge tube followed by the addition of lysozyme (1 mg/ml). A whole cell lysate was then prepared as described in the immunoblot protocol above. The microfuge tube used to collect cells and culture medium was washed twice with 1 mL LB to remove any remaining cells. The proteins bound non-specifically to the tube were then released with SDS sample buffer. The sample buffer with released proteins was then used to resuspend the TCA-precipitated proteins from the culture supernatant. Equivalent amounts of cell lysate (C) and culture medium (M) were resolved by SDS-PAGE and CwlO and SMC were analyzed by immunoblot.
The ΔsweDC::kan ΔlytE::cat double mutant was constructed under permissive growth conditions (CH agar supplemented with 20 mM MgCl2 and 0.25 M sucrose). Three independent clones (BYB366, BYB367, BYB368) were grown to early stationary phase under permissive conditions at 37°C. Cells were washed twice in fresh LB and plated under permissive and restrictive conditions (LB agar) at 37°C. The suppressor frequency was ~10−7. Genomic DNA from 13 suppressors was prepared for whole genome sequencing using a modified Nextera library preparation protocol [77]. DNA concentrations were determined using the Qubit dsDNA HS Assay Kit and fragment sizes were determined using a High Sensitivity D1000 screen tape run on an Agilent 4200 TapeStation system. Sequencing was performed using a MiSeq Kit v6, with the Miseq System (Illumina). Reads were mapped using CLC Genomics Workbench software (Qiagen).
Co-IPs were performed as described previously [78]. Briefly, 150 mL cultures of wild-type and the ΔftsX mutant were harvested at an OD600 of 0.5 washed twice with 1X SMM (0.5 M sucrose, 20 mM MgCl2, 20 mM maleic acid pH 6.5) at room temperature. Cells were resuspended in 1:10 volume 1XSMM and protoplasted with lysozyme (0.5 mg/mL final) for 30 minutes. Protoplasts were collected by centrifugation and disrupted by osmotic lysis with 3 ml hypotonic buffer (Buffer H) (20 mM Hepes pH 8, 200 mM NaCl, 1 mM Dithiothreitol) with protease inhibitors: 1 mM phenylmethylsulfonyl fluoride and EDTA-free protease inhibitor cocktail complete (Roche). MgCl2 and CaCl2 were added to 1 mM and lysates were treated with DNAse I (10 μg/ml) (Sigma-Aldrich) and RNAse A (20 μg/ml) (USB) for 1 h on ice. The membrane fraction was separated by ultracentrifugation at 100,000 × g for 1 h at 4°C. The supernatant was carefully removed, and the membrane pellet was dispersed in 400 μL of Buffer G (Buffer H with 10% glycerol). Crude membranes were aliquoted and flash-frozen in N2(l). 200 μL crude membranes were diluted 5-fold with Buffer S (Buffer H with 20% glycerol and 100 μg/ml lysozyme), and membrane proteins were solubilized by the addition of the nonionic detergent digitonin (Sigma) to a final concentration of 0.5%. The mixture was rotated at 4°C for 1 h. Soluble and insoluble fractions were separated by ultracentrifugation at 100,000 × g for 1 h at 4°C. The soluble fraction from the digitonin-treated membrane preparation (the load) was mixed with 4 μl of crude anti-FtsX antiserum [25] and rotated for 3 h at 4°C. The mixture was added to 25 μL of Protein A Sepharose resin (GE Healthcare) and rotated for 1 h at 4°C. The resin was pelleted at 5 Krpm, and the supernatant (the flow-through) was collected. The resin was washed four times with 0.4 mL of Buffer S + 0.5% digitonin. Immunoprecipitated proteins were eluted by the addition of 50 μl of sodium dodecyl sulfate (SDS) sample buffer (0.25 M Tris, pH 6.8, 6% SDS, 10 mM EDTA, 20% glycerol) and heated for 15 min at 50°C. The resin was pelleted, and the supernatant (the IP) was transferred to a fresh tube and 2-mercaptoethanol was added to a final concentration of 10%. The load, flow-through, and immunoprecipitate were analyzed by immunoblot.
The Bacterial Adenylate Cyclase-based Two Hybrid (BACTH) system was used as previously described ([79, 80]. Briefly, pairs of proteins were fused to the complementary fragments (T18 and T25) of the Bordetella pertusis adenylate cyclase. After co-transformation into BTH101, independent transformants were inoculated in LB medium supplemented with ampicillin (100 μg/mL), kanamycin (30 μg/mL) and 0.5 mM isopropyl-β-thio-galactoside (IPTG). Cells were grown at 30°C overnight and spotted on LB agar plates supplemented with ampicillin (100 μg/mL), kanamycin (30 μg/mL), IPTG (0.25 mM) and 5-bromo-4-chloro-3-indolyl-β-D-galactopyrannoside (X-Gal) (20 mg/mL). Plates were photographed after incubation at 30°C for 16 hours.
Chromatin immunoprecipitation (ChIP) was performed as described previously [69]. Briefly, cells were crosslinked using 3% formaldehyde for 30 min at room temperature and then quenched, washed, and lysed. Chromosomal DNA was sheared to an average size of 250 bp by sonication using a Qsonica Q800 water bath sonicator. The lysate was then incubated overnight at 4°C with anti-SweD antisera, and was subsequently incubated with Protein A-Sepharose resin (GE HealthCare) for 1 hr at 4°C. After washes and elution, the immunoprecipitate was incubated at 65°C overnight to reverse the crosslinks. The DNA was further treated with RNase A, Proteinase K, extracted using Phenol-Chloroform, resuspended in 50 μl EB and used for library preparation with the NEBNext Ultra kit (E7370S) and sequenced using the Illumina MiSeq platform.
The sequencing reads from wild-type and ΔsweDC ChIP samples were mapped to the B. subtilis PY79 genome (NCBI Reference Sequence: NC_022898.1) using CLC Genomics Workbench (CLC bio, QIAGEN). Subsequent normalization, plotting, and analyses were done using R plots as follows. Samples were first normalized to the total number of reads. Then the ratio of ChIP signal in wild-type relative to ΔsweDC was calculated and plotted in S4A Fig. The data were plotted in 1 kb windows. For S4B Fig, ChIP signals of wild-type and ΔsweDC between 2390kb and 2400kb were plotted.
|
10.1371/journal.pcbi.1005153 | Structural Identifiability of Dynamic Systems Biology Models | A powerful way of gaining insight into biological systems is by creating a nonlinear differential equation model, which usually contains many unknown parameters. Such a model is called structurally identifiable if it is possible to determine the values of its parameters from measurements of the model outputs. Structural identifiability is a prerequisite for parameter estimation, and should be assessed before exploiting a model. However, this analysis is seldom performed due to the high computational cost involved in the necessary symbolic calculations, which quickly becomes prohibitive as the problem size increases. In this paper we show how to analyse the structural identifiability of a very general class of nonlinear models by extending methods originally developed for studying observability. We present results about models whose identifiability had not been previously determined, report unidentifiabilities that had not been found before, and show how to modify those unidentifiable models to make them identifiable. This method helps prevent problems caused by lack of identifiability analysis, which can compromise the success of tasks such as experiment design, parameter estimation, and model-based optimization. The procedure is called STRIKE-GOLDD (STRuctural Identifiability taKen as Extended-Generalized Observability with Lie Derivatives and Decomposition), and it is implemented in a MATLAB toolbox which is available as open source software. The broad applicability of this approach facilitates the analysis of the increasingly complex models used in systems biology and other areas.
| Advances in computing power have facilitated the development of increasingly larger dynamic models of biological processes, which usually have many unknown parameters. Often times, such models contain parameters that are structurally unidentifiable, i.e., they cannot be uniquely determined from experiments. Any parameter estimation algorithm will fail when trying to estimate unidentifiable parameters, leading to waste of resources and possibly wrong model predictions. Hence, it is essential to assess structural identifiability before exploiting a model. However, performing such analysis can be hard, especially as models become increasingly complicated. To address this challenge, we developed a methodology for structural identifiability analysis that aims at generality—it can handle any analytic model written as a set of ordinary differential equations—and computational efficiency—it includes features that facilitate the analysis of large systems. We provide an implementation of the methodology as a MATLAB toolbox called STRIKE-GOLDD. We illustrate its applicability to systems biology models of genetic, signalling, metabolic, and pharmacokinetic networks, showing which of them are unidentifiable and how they can be made identifiable.
| Mathematical modelling has become a fundamental tool in present day biology [1], and system identification is one of the key tasks of this process [2]. Building a dynamic model usually requires establishing the values of some unknown parameters, which raises the issue of parameter identifiability [3]. A model is structurally identifiable if it is possible to determine the values of its parameters from observations of its outputs and knowledge of its dynamic equations [4]. While the related concept of practical identifiability refers to quantifying the uncertainty in parameter values when estimated from noisy measurements, structural identifiability does not take into account limitations caused by the quality or availability of experimental data. It is, however, a necessary (a priori) condition for practical identifiability, which, in turn, is a prerequisite for model calibration, also known as parameter estimation [5]. Any identification efforts aimed at estimating unidentifiable parameters will fail, leading to wrong estimates, waste of resources, and possibly misleading model predictions [6]. Furthermore, if structural unidentifiability is mistaken for practical unidentifiability, it may lead to trying to overcome it by investing additional efforts in designing and performing new experiments [7], which will nevertheless be sterile. Hence it is essential to assess the structural identifiability of any unknown parameters in a model before attempting to calibrate it. As stressed in the conclusions of a recent parameter estimation challenge [8], “modelers must avoid creating structurally unidentifiable parameters that can never be estimated”. However, in real applications structural identifiability is seldom checked before performing parameter estimation [9]. This is at least partly due to the computational complexity of the problem: structural identifiability methods generally require symbolic manipulations, which can quickly give rise to long expressions as the system size increases [10].
This is a major challenge in systems biology, as the models constructed are increasingly complex, large [11], and more difficult to identify [8]. However, the development of structural identifiability tools has been lagging behind, and, despite the wide variety of methods developed for this task (some of which have publicly available implementations [12–16]), the analysis of some models remains elusive. For example, although recent improvements in efficiency [16, 17] have enabled the analysis of increasingly large rational models (those that can be expressed as fractions of polynomial functions), non-rational systems such as those including trigonometric expressions or Hill-type kinetics (which are common in mechanical and biochemical models, respectively) can currently be analysed only for small sizes. While in certain cases non-rational models can be rewritten in rational form, by introducing additional variables and equations, it is not always possible or convenient to do so. Furthermore, the results obtained for the rational counterpart are not necessarily valid for the original non-rational model in the case of unidentifiability [18]. Recent studies [9, 10, 19–21] show that, in general, the choice of a structural identifiability method involves trade-offs between generality of the application, computational cost, and level of detail of the results. In conclusion, there is currently a lack of structural identifiability methods of the sufficient generality and robustness to be applied to nonlinear models of general form and realistic size [21, 22].
To address this issue, we propose a methodology applicable in principle to any analytic system and geared towards computational efficiency. This method approaches local structural identifiability as a generalized version of observability, a classic concept in systems and control theory [23]. A system is observable if it is possible to determine its internal state from output measurements in finite time. If the model parameters are considered as state variables with zero dynamics, structural identifiability analysis can be recast as a generalization of observability analysis [17, 24, 25]. In this way it is possible to assess the structural identifiability of nonlinear systems using results from differential geometry [21]. Essentially, identifiability is determined by calculating the rank of a generalized observability-identifiability matrix, which is constructed using Lie derivatives. When this rank test classifies a model as unidentifiable, the procedure determines the subset of identifiable parameters. In some cases it is also possible to find identifiable combinations of the remaining parameters. This approach is directly applicable to many models of small and medium size; larger systems can be analysed using additional features of the method. One of them is decomposition into more tractable submodels, which is performed with a combinatorial optimization metaheuristic as in [26]. Another possibility is to build identifiability matrices with a reduced number of Lie derivatives. In some cases these additional procedures allow to determine the identifiability of every parameter in the model (complete case analysis); when such result cannot be achieved, at least partial results—i.e. identifiability of a subset of parameters—can be obtained.
We illustrate the applicability of this method to systems biology models of different types, including genetic, signalling, metabolic, and pharmacokinetic networks. Some of them are non-rational systems exhibiting Hill kinetics, that is, with expressions containing terms of the form k1xn/(k2 + xn), such as the Goodwin model of transcriptional repression [27], the mitogen-activated protein kinase (MAPK) signalling cascade [28], and the genetic network that controls the circadian clock in Arabidopsis thaliana [29]. Other models analysed here include drug uptake into hepatocytes [19], NF-κB [30] and JAK/STAT [31] signalling pathways, and the central carbon metabolism of Chinese hamster ovary cells [32]. These case studies include models whose identifiability had not been previously determined, and for some of them we found unidentifiabilities that had not been reported before. In those cases, we obtained identifiable reparameterizations by removing redundant parameters and fixing the values of other parameters a priori.
We consider dynamic models described by ordinary differential equations of the following general form:
M : x ˙ ( t ) = f [ x ( t ) , u ( t ) , p ] , y ( t ) = g [ x ( t ) , p ] , x 0 = x ( t 0 , p ) (1)
where f and g are analytic (and therefore infinitely differentiable) vector functions, p ∈ R q is a real-valued vector of parameters, u ( t ) ∈ R r is the input vector, x ( t ) ∈ R n the state variable vector, and y ( t ) ∈ R m the measurable output, also called the observables vector. In Eq (1) the dependence on the parameters p is made explicit, but it will be usually dropped for ease of notation. Parameter pi is structurally globally identifiable (s.g.i.) if it can be uniquely determined from the system output, that is, if for almost any p * ∈ R q (i.e., for any p except those belonging to a set of measure zero) the following property holds [5, 33]:
y ( t , p ^ ) = y ( t , p * ) ⇒ p i ^ = p i * (2)
A parameter pi is structurally locally identifiable (s.l.i.) if for almost any p* there is a neighbourhood V(p*) in which Eq (2) holds. A model M is said to be s.g.i. if all its parameters are s.g.i., and s.l.i. if all its parameters are s.l.i. If Eq (2) does not hold in any neighbourhood of p*, parameter pi is structurally unidentifiable (s.u.), and a model M is s.u. if at least one of its parameters is s.u.
In this work we consider identifiability as an augmented observability property. We begin the description of the approach by defining observability and showing how it can be assessed. A system is (locally) observable at a state x0 if there exists a neighbourhood N of x0 such that every other state x1 ∈ N is distinguishable from x0. Two states x0 ≠ x1 are said to be distinguishable when there exists some input u(t) such that y(t, x0, u(t)) ≠ y(t, x1, u(t)), where y(t, xi, u(t)) denotes the output function of the system for the input u(t) and initial state xi(i = 0, 1).
The concept of observability was initially formulated by Kalman for linear systems [34], and then extended to the nonlinear case by Hermann and Krener [23]. For a nonlinear system given by Eq (1) it is possible to obtain information about the states x from its outputs y by calculating the derivatives y˙,y¨,…. These differentiations are performed by taking Lie derivatives of the output function g. The Lie derivative of g with respect to f is:
L f g ( x ) = ∂ g ( x ) ∂ x f ( x , u ) (3)
For a system with n states and m outputs, ∂ ∂ x g ( x ) is an m × n matrix, and L f g ( x ) = ∂ g ( x ) ∂ x f ( x , u ) is an m × 1 column vector. The ith order Lie derivatives are recursively defined as follows:
L f 2 g ( x ) = ∂ L f g ( x ) ∂ x f ( x , u ) ⋯ L f i g ( x ) = ∂ L f i - 1 g ( x ) ∂ x f ( x , u ) (4)
Stacking n sub-matrices, we obtain the nonlinear observability matrix:
O ( x ) = ∂ ∂ x g ( x ) ∂ ∂ x ( L f g ( x ) ) ∂ ∂ x ( L f 2 g ( x ) ) ⋮ ∂ ∂ x ( L f n - 1 g ( x ) ) (5)
We can now formulate the Observability Rank Condition (ORC) as follows: if the system given by Eq (1) satisfies rank ( O ( x 0 ) ) = n, where O is defined by Eq (5), then it is (locally) observable around x0 [35].
The rank condition provides a result about local observability of any possible state x0. That is, if the matrix is full rank then for every state x0 there exists a neighbourhood N(x0) in which x0 can be distinguished from any other state x*. In other words, every state can be distinguished from its neighbours, but not necessarily from other distant states. In contrast, global observability is a property that must hold for every possible N(x0). The difference is clearly shown with the following example [23]:
x ˙ = u , y 1 = cos ( x ) , y 2 = sin ( x ) (6)
While this system satisfies the observability rank condition and is therefore locally observable, it is not globally observable because it is impossible to distinguish between x0 and xk = x0 + 2kπ, for any integer k.
We remark that the observability rank condition does not require the assumption of constant inputs u; analytic differentiable input functions can be used [36, 37]. As noted in [38], this entails that u can be treated symbolically in rank calculations.
While identifiability problems can be addressed by a number of techniques not explicitly related to nonlinear observability, it is possible to consider the parameters p as additional states with trivial dynamics p ˙ = 0 and, in this way, the identifiability problem can be recast in the framework of observability [17, 21, 24]. Thus, by augmenting the state variable vector so as to include model parameters, x ˜ = [ x , p ], we obtain a generalized observability-identifiability matrix, O I ( x ˜ ):
O I ( x ˜ ) = ∂ ∂ x ˜ g ( x ˜ ) ∂ ∂ x ˜ ( L f g ( x ˜ ) ) ∂ ∂ x ˜ ( L f 2 g ( x ˜ ) ) ⋮ ∂ ∂ x ˜ ( L f n + q - 1 g ( x ˜ ) ) (7)
With this formulation we can define a generalized Observability-Identifiability Condition (OIC) as follows: if the system given by Eq (1) satisfies rank ( O I ( x ˜ 0 ) ) = n + q, it is (locally) observable and identifiable in a neighbourhood N ( x ˜ 0 ) of x ˜ 0.
Since we have recast the analysis of structural identifiability as a particular case of observability, the same remark that was made in the preceding subsection about the difference between local and global properties applies here.
It has been noted [39] that in certain cases a system may become unreachable for specific values of the initial conditions, leading to the impossibility of determining the values of parameters classified as identifiable by structural identifiability methods. This situation can be detected if rank ( O I ( x ˜ 0 ) ) is calculated using a vector of specific initial conditions instead of a generic symbolic vector.
Finally, we note that the idea of treating parameters and state variables similarly is also adopted by estimation methods such as extended Kalman filtering [40]. However, the context is different, since the goal of such techniques is to determine the value of states and parameters from data, while structural identifiability analysis aims at establishing whether such estimation is theoretically possible.
In practice, checking the aforementioned Observability-Identifiability Condition (OIC) is often computationally inefficient (or even infeasible) because building O I and calculating its rank is a highly demanding, memory-consuming task. Fortunately, sometimes this cost can be decreased by building a smaller matrix. Let us first note that each of the n + q sub-matrices vertically stacked in the generalized observability-identifiability matrix of Eq (7) has dimension m × (n + q), and the full matrix O I has dimensions (m ⋅ (n + q)) × (n + q). Therefore it may not be necessary to calculate the n + q − 1 Lie derivatives in order to test whether O I is full rank, since full rank may be achieved with a lower number of derivatives. The minimum number of Lie derivatives for which the matrix may be full rank is
n d = n + q m - 1 (8)
that is, the smallest integer not less than (n + q)/m − 1, where n, q, and m are the numbers of states, parameters, and outputs, respectively. The maximum number of Lie derivatives is also known a priori: derivatives of order higher than n + q − 1 cannot increase the matrix rank [38]. Having lower and upper bounds for the necessary Lie derivatives is an advantage of this methodology compared to, e.g., power series approaches, for which the maximum number of derivatives is in principle infinite [10].
Our method builds O I recursively. Once nd is reached, addition of a new Lie derivative is followed by calculation of the rank. This process is repeated until the maximum number n + q − 1 is reached, or until adding a new Lie derivative does not increase the matrix rank; in both cases no further derivatives are necessary [38]. At that point, if O I is full rank the corresponding model is observable and identifiable, as seen in the previous subsection. If O I is not full rank, the algorithm proceeds to find identifiable parameters, as explained in the following subsection.
Further improvements in the computational burden can be obtained by calculating the rank numerically instead of symbolically. A way in which this can be performed is by replacing the symbolic variables in the O I with prime numbers to minimize the risk of accidental cancellations, which would reduce the rank.
If O I is not full rank, the Observability-Identifiability Condition (OIC) does not inform us about which parameters are identifiable and which are not. This can be achieved by realizing that each column of O I corresponds to a parameter-to-output relation (or state-to-output):
∂ ∂ x 1 g ( x ˜ ) ∂ ∂ x 2 g ( x ˜ ) ⋯ ∂ ∂ p q g ( x ˜ ) ∂ ∂ x 1 ( L f g ( x ˜ ) ) ∂ ∂ x 2 ( L f g ( x ˜ ) ) ⋯ ∂ ∂ p q ( L f g ( x ˜ ) ) ⋮ ⋮ ⋮ ⋮ ∂ ∂ x 1 ( L f n + q - 1 g ( x ˜ ) ) ∂ ∂ x 2 ( L f n + q - 1 g ( x ˜ ) ) ⋯ ∂ ∂ p q ( L f n + q - 1 g ( x ˜ ) )
Therefore, if deleting the ith column of the generalized observability-identifiability matrix does not change its rank, then the corresponding ith state (parameter) is non-observable (unidentifiable). This fact can be exploited to determine which of the parameters in an unidentifiable model are identifiable and which are not, using a sequential procedure: after the matrix rank has been calculated and the model has been found to be unidentifiable, each of the columns in O I corresponding to a particular parameter is removed one by one and the rank is recalculated. In this way the identifiability of each of the parameters is evaluated.
The procedure outlined in the preceding subsections classifies the model parameters as either identifiable of unidentifiable. A question that naturally follows is: are there combinations of the unidentifiable parameters which are themselves identifiable? If the answer is affirmative, the model can be reparameterized and converted to a structurally identifiable model. However, this is a difficult problem, which few methods can address, and only for models of moderate size. An example is COMBOS [41, 42], which is based on differential algebra. Here we suggest an approach based on ideas presented in [43, 44] and on the method for finding symmetries proposed by [38]; related work has been recently presented in [45].
The procedure is as follows: if O I is rank-deficient, remove the columns corresponding to identifiable parameters and obtain a reduced submatrix, O U [38]. Then, obtain a basis for the kernel (null space) of this matrix, N ( O U ) (step 2 in [44]). Its coefficients define one or several partial differential equations whose solution(s) are the identifiable combinations (step 3 in [44]). This procedure is illustrated in the Methods section with the JAK/STAT signalling pathway, for which an identifiable combination of two parameters is found. While this example shows the potential of this procedure, it must be acknowledged that the computational complexity of calculating the kernel of O U limits its applicability to models with a moderate number of unidentifiable parameters.
The methodology described in the previous subsections can be used to analyse the identifiability of whole models and, if the model is unidentifiable, of its parameters individually. However, since it relies heavily on symbolic operations, it may be computationally infeasible for large or complex models. It should be noted that the main limiting operations are:
The minimum number of derivatives necessary for building O I ( x ˜ ) is given by nd as defined in Eq (8). The limit of what is computationally possible is difficult to quantify a priori, since it depends on the model equations and the machine used in the calculations. As a rule of thumb, analyses involving nd ≥ 10 are infeasible except for very small models. As model size or complexity increases, this upper bound decreases; some examples will be shown in the Results section.
One solution is to decompose those models into smaller submodels whose analysis is possible computationally. Thus, we seek to decompose a model M into submodels {M1, M2, …} which require few Lie derivatives for their analysis, that is, they have a small nd. Each submodel Msub includes a subset of the model states, xsub. Its outputs, ysub, are the outputs of M which are functions of at least one state included in xsub. The submodel parameters and inputs are those appearing in the equations of xsub and ysub. There may be states that appear in the equations of xsub or ysub but are not part of xsub; they are considered as additional unknown parameters of Msub.
The submodels can be found by optimization as follows. For each submodel Mi we select a subset of the states in M by performing a combinatorial optimization where we minimize nd:
min s n d ( s ) (9)
where s = {s1, s2, …, sn} is a binary vector of size n, whose entries sj ∈ {0, 1} denote inclusion (sj = 1) or exclusion (sj = 0) of the corresponding state. The combinatorial optimization is performed with the Variable Neighbourhood Search metaheuristic [46]. We carry out n optimizations (one per state); in the jth optimization we force sj = 1, so that each state appears in at least one solution. This, in turn, guarantees that all the parameters will eventually be evaluated. A penalty term is included in the objective function to penalize solutions that have more states than a chosen maximum.
Apart from this optimization-based decomposition, it may sometimes be useful to specify a particular submodel in order to explore the identifiability of a specific part of the model.
Let us clarify how we can conclude identifiability of a parameter from analysis of a submodel. As an example, consider M to be the model of Arabidopsis thaliana described in the Results section; its equations are given in the Supplementary Information (S1 Text). Let us consider a submodel Msub consisting of two states, xsub = {x1, x7}. The equations of Msub are those that correspond to the states {x1, x7}, that is:
{ x˙1=n1x6ag1a+x6a−m1x1k1+x1+q1x7u(t),x˙7=p3−m7x7k7+x7−(p3+q2x7)u(t),x1(0)=0,x7(0)=0 (10)
The outputs of Msub are those outputs of M which are functions of at least one of the states in Msub (in this example, y1 = x1). The parameters and inputs of Msub are those present in Eq (10): respectively, {n1, g1, a, m1, k1, q1, p3, m7, k7, q2} and u. Additionally, we must also include as parameters the states that do not belong to xsub but appear in Eq (10) or in ysub (in this case, x6). Thus in this example the submodel parameters would be {n1, g1, a, m1, k1, q1, p3, m7, k7, q2, x6}. By including states such as x6 as parameters we are considering them as unknown and constant. In contrast, if they were included as inputs to the submodel, we would be implicitly assuming that they provide sufficient excitation for identification purposes. Thus, including them as parameters is a conservative assumption in terms of identifiability. Therefore, if a parameter is classified as identifiable in a submodel under these conditions, it will also be identifiable when considering the whole model.
When the nd of the full model is so high that it is not feasible to build O I, one solution is to decompose the model into smaller submodels as described in the previous subsections. Another possibility is to build O I with i < nd derivatives. In this case we know that full rank cannot be achieved, so even if the model is identifiable we will not be able to determine it in this way. However, it may be possible to determine identifiability of at least some of the parameters. Note that this procedure can be helpful exactly in the same circumstances as decomposition. In some cases one approach will be more appropriate than the other, but both can be used to determine the identifiability of different parameters, and may therefore be complementary.
Fig 1 shows a diagram of the methodology presented so far. The next sections discuss the types of analyses that can be performed with this methodology and show how to refine the solutions iteratively in order to obtain more complete diagnoses.
By assessing identifiability as explained in sections “Assessing the OIC efficiently” and “Determining identifiability of individual parameters” we are performing a “Complete Case Analysis” (CCA): every parameter in the model is either classified as identifiable or as unidentifiable.
However, it may not always possible to carry out the aforementioned procedure due to computational limitations, as explained in sections “Decomposing large models into submodels to facilitate their analysis” and “Building O I with less than nd Lie derivatives”, which presented two different alternatives. In certain cases these alternatives can yield incomplete results, that is, they may fail to determine the (un)identifiability of some parameters. For example, this may happen in the following scenarios:
The two cases mentioned above will be called “Partial Analyses for Identifiability” (PAI): some parameters are conclusively classified as identifiable, but nothing can be said about the rest. It is also possible to perform similar analyses to guarantee unidentifiability of some parameters, leading to what we will call “Partial Analyses for Unidentifiability” (PAU). In such tests, some parameters are classified as unidentifiable while the analysis of the rest is not conclusive. This can happen in at least two situations:
The different types of analyses that can be performed are summarized in Table 1.
As shown in the preceding subsection, for some complex problems a complete analysis—that is, classifying all the parameters as identifiable or unidentifiable—may not be possible, at least in a first approach, due to computational limitations. In such cases, one can try to obtain more complete diagnoses by running the procedure iteratively. At each time, the computational cost can be reduced by removing from the augmented state vector, x ˜ = [ x , p ], those parameters that were already found to be identifiable in previous steps. This operation, which leads to a smaller O I matrix, does not alter the result of the identifiability test, because the resulting O I is identical to the one obtained with the original vector x ˜ = [ x , p ] after removing the columns corresponding to identifiable parameters—which results in a decreased rank. Note that this equivalence is made possible by the fact that p ˙ = 0, so the procedure cannot be applied to the model states, since x ˙ ≠ 0.
In summary, if a model M is too large to be analysed as a whole—i.e. to directly calculate the rank of its identifiability matrix and perform a complete case analysis (CCA)—identifiability analysis can be approached as follows:
The present method has been implemented as a MATLAB toolbox named STRIKE-GOLDD (STRuctural Identifiability taKen as Extended-Generalized Observability using Lie Derivatives and Decomposition). It is an open source tool licensed under the GNU General Public License version 3 (GPLv3). It is freely available from https://sites.google.com/site/strikegolddtoolbox/ and as supplementary information accompanying this article (S1 File). It requires a MATLAB installation with the Symbolic toolbox. Additionally, to use optimization-based decomposition it is necessary to install the MEIGO toolbox [48]. The usage of the STRIKE-GOLDD software is discussed in detail in the manual (S2 Text); in the following lines we provide a brief description of the key options.
The toolbox allows limiting the number of Lie derivatives that are calculated when building O I ( x ˜ ). This is useful to prevent the algorithm from getting stuck in excessively lengthy calculations. To adapt this limit to the computer where the algorithm is running, it is specified as a machine-dependent criterion: the user can set a limit on the time invested in calculating these derivatives by entering it in opts.maxLietime (that is, the algorithm will not calculate the ith + 1 derivative if the time spent in obtaining the ith one was ti > opts.maxLietime).
Furthermore, the user can choose what to do if this time limit is reached without O I ( x ˜ ) being full rank: by setting opts.decomp = 0, STRIKE-GOLDD will perform a partial analysis of the whole model with the current O I ( x ˜ ); if opts.decomp = 1, it will decompose the model. It is also possible to enforce the use of decomposition from the start, i.e. without checking whether the time limit is reached, with the option opts.forcedecomp. The submodels can be found by optimization or specified by the user; this choice is made by opts.decomp_user.
Another option, opts.numeric, allows deciding whether to calculate rank(O I ( x ˜ )) numerically or symbolically. The symbolic calculation is always exact. It is possible to perform a numerical calculation by replacing the symbolic variables with prime numbers. This usually reduces the computational cost, although in some cases it might lead to accidental cancellations that decrease the rank artificially. However, the risk of obtaining a spurious result is low, and it can be minimized by running the procedure several times, since the substitutions are random. In all of the case studies analyzed in the Results section we found agreement between numeric and symbolic rank calculations.
Finally, it is possible to assess identifiability for specific values of the system’s initial conditions. As mentioned in subsection “Structural identifiability as augmented observability: the OIC”, this can be useful in order to detect situations in which loss of reachability from particular initial conditions leads to loss of identifiability. Such pathological cases are not detected if rank ( O I ( x ˜ 0 ) ) is calculated using a generic symbolic vector of initial conditions. However, they can be tested by setting the option opts.knowninitc = 1 and entering the corresponding vector of initial conditions in the script that creates the model.
We applied the proposed methodology to a number of published models of varying size and complexity [19, 27–32], some of which have been recently used to benchmark identifiability analysis methods [10, 19, 20]. The main features of the models are summarized in Table 2, and schematic diagrams are shown in Figs 2–4. Their equations are given in the supplementary information. Calculations were carried out on a computer running Windows7 SP1 64bit, with an Intel processor at 3.40 GHz and 16 GB of RAM, using MATLAB R2015b.
In [19], Grandjean and coworkers proposed 18 alternative pharmacokinetic nonlinear compartmental models of the uptake process of Pitavastatin (a drug used to treat hypercholesterolaemia) into hepatocytes. They applied five different methods to analyse their structural identifiability: similarity transformation, differential algebra, Taylor series, and two approaches based on a non-differential input/output observable normal form and an algebraic input/output relationship approach. With these techniques they established the identifiability of most of the models. However, for several model formulations none of the methods was able to produce results. This was the case for two candidate models (with or without pseudo-state assumption) that accounted for drug metabolising within the cell.
A diagram of these Pitavastatin uptake models is shown in panel A of Fig 2. The upper part of the panel shows the system’s functional diagram. The lower part shows a graph drawn following the same convention as in [47], in which a directed arrow from A to B indicates that B appears in the dynamic equation of A. This graphical approach was originally proposed to study observability, and hence in [47] only the states were shown in the graphs. Since here we use it for identifiability purposes, we extend it to include both states and parameters (see figure caption for more details).
The method presented here determines that both Pitavastatin uptake models (with and without pseudo steady state assumption) are structurally identifiable.
The classical model of oscillations in enzyme kinetics proposed by Goodwin in 1965 [27] and shown in panel B of Fig 2 is still the subject of analyses [28]. It was selected by [10] to compare the performance of several structural identifiability methods, considering two different scenarios or variations of the model: when the three states are measured, or when only one of them—the enzyme concentration, x1—can be measured. The latter situation is more realistic, but its analysis is particularly challenging, and none of the methods tested by [10] managed to reach a conclusion due to computational complexity.
According to Eq (8), the minimum number of Lie derivatives for which the identifiability matrix may be full rank is nd = 10 for this model. While the subsequent rank calculation is very demanding, the computational cost is substantially reduced by building O I ( x ˜ ) with only 9 Lie derivatives. In this way the method classifies four parameters as identifiable: b, σ, β, δ. Then, removing these parameters from the model as explained in “Iterative refinement of the diagnosis” enables the analysis of the remaining parameters (a, A, α, γ), which are found to be unidentifiable.
Thus this model is unidentifiable. It can be made identifiable by considering two parameters as known, one from each of the pairs {A, a} and {α, γ}. For example, if we fix the values of {A, α}, the remaining six unknown parameters in the model are identifiable. An alternative solution is to measure more states, if it is experimentally possible. In this case, if all three states are outputs, the model is structurally identifiable. Measuring only two of the three states, however, increases the number of identifiable parameters but does not render the model fully identifiable. The subsets of unidentifiable parameters for y = {x1, x2}, y = {x1, x3}, and y = {x2, x3} are, respectively, {a, A, γ}, {α, γ} and {a, α}.
This model was presented in [28] as an example of a system exhibiting both oscillation and bistability. It is a three-layer signalling cascade with positive and negative feedback loops and Hill nonlinearities, shown in panel C of Fig 2. It has three states, which are the phosphorylated forms (x1, x2, x3), and 14 parameters (k1, k2, k3, k4, k5, k6, s1t, s2t, s3t, K1, K2, n1, n2, α).
This system requires that all its three states are measured in order to be identifiable. However, if just one of the states is left unmeasured, some parameters become unidentifiable: if x1 is not measured, k3 and s1t are unidentifiable; if x2 is not measured, k5 and s2t are unidentifiable; and if x3 is not measured, K1, K2, and s3t are unidentifiable.
This model was presented by [30] and was used by both [10] and [16] as a benchmark for structural identifiability methods. In the formulation of [10], only 13 parameters are considered unknown. In that case, all of them are identifiable. The general case, in which all 29 parameters are in principle unknown, is more challenging. For this case STRIKE-GOLDD classifies 5 parameters as unidentifiable: c1c, c2c, c3c, c4, and k2, and the remaining as identifiable. Part of this diagnosis can be confirmed by inspection of the connection diagram in the right side of Fig 3, which shows that c1c, c2c, and c3c only appear in the equation of state x15. Since x15 is in turn “disconnected” from the rest of the model (i.e. it does not appear in the equation of any other state), and it is not measured, there is clearly no way of determining its value. Hence x15 is unobservable, and the three parameters associated with it are unidentifiable. In contrast, the unidentifiability of c4 and k2 is by no means apparent from the figure. However, it can be determined with the methodology that they are not only unidentifiable, but related: fixing any of the two renders the other one identifiable. In summary, this 29-parameter model can be converted into a structurally identifiable 25-parameter model by fixing the values of four parameters: c1c, c2c, c3c, and either c4 or k2.
This model of the IL13-Induced JAK/STAT signalling pathway was presented in [31] and later used in [20] to benchmark three identifiability analysis methods. The network interaction diagrams are shown in panel A of Fig 4. The results of our method coincide with those reported in [20], that is, five of the 23 parameters are unidentifiable, pu = [θ11, θ15, θ17, θ21, θ22]. Following the procedure outlined in the Methods section, it is possible to find an identifiable combination of unidentifiable parameters. To do this we remove the columns corresponding to identifiable parameters and obtain a reduced submatrix, O U. Calculation of a basis of the kernel of O U yields the following vector:
v = [ 0 , 0 , - θ 17 θ 22 , 0 , 1 ] (11)
which in turn leads to the following PDE:
- θ 17 θ 22 · ∂ Φ ∂ θ 17 + ∂ Φ ∂ θ 22 = 0 ⇒ Φ = θ 17 · θ 22 (12)
Thus, Φ = θ17 ⋅ θ22 is an identifiable parameter combination. The methodology does not report any combination involving θ11, θ15, θ21. If, additionally, we fix the value of θ11 a priori, we obtain a structurally identifiable model with 21 unknown parameters.
The genetic subnetwork that controls the circadian clock in the plant A. thaliana was modelled in [29]; its diagram is shown in Fig 4. This model uses both Michaelis-Menten and Hill kinetics. Two Hill coefficients of transcription (a, b) were considered as known parameters in the original publication [29]. Although it was argued in [29] that there is evidence that b = 2, coefficient a was fixed to a = 1 without experimental evidence. In [10] it was reported that (for the case of a = 1) no global structural identifiability method was capable of successfully analysing the model; at most, identifiability of five parameters was established.
While the choice of a = 1 makes the system rational and reduces the problem dimension, here we consider the more general case in which a is an unknown parameter. According to Eq (8), the minimum number of Lie derivatives for which O I ( x ˜ ) may be full rank is very high for this model (nd = 16). This is the same situation as with the previously analysed Goodwin model, that is, the computational cost of the construction and subsequent rank calculation of O I ( x ˜ ) with nd derivatives is too high. Furthermore, we found that the approach adopted for the Goodwin model—building the matrix with less than nd derivatives—was not successful in the case of this example, at least when performed with few derivatives. Hence we turned to the alternative procedure, i.e. decomposing the model using optimization. In this way, identifiability of ten parameters was established: a, k1, k4, m1, m4, n1, n2, q2, r2, and r4. Removing these parameters from the model decreases the number of required derivatives nd to 12, which is still very high; however, building O I ( x ˜ ) with 9 derivatives reports identifiability of an additional parameter, r1.
By performing partial analyses for unidentifiability (PAUs) we confirmed that the model is indeed unidentifiable. This can be remedied in several ways. A possible solution is to measure more states, if it is experimentally feasible. In the model it has been assumed that only mRNA concentrations are measured (i.e. states x1 and x4); however, if protein concentrations (i.e. the remaining states) are also measured, then all the parameters become structurally identifiable. Alternatively, if we assume that only the original outputs can be measured, it is possible to obtain an identifiable reformulation of the model by fixing some parameters. For example, choosing fixed values for the five degradation constants that were not found to be identifiable (k2, k3, k5, k6, k7) yields a structurally identifiable model with 23 parameters.
This large-scale model was taken from the BioPreDyn-bench collection [32], where it was included as benchmark B4. It models a batch fermentation process for protein production using Chinese Hamster Ovary cells. Its diagrams are shown in Fig 5. It contains 34 states (which are metabolites present in three compartments: fermenter, cytosol, and mitochondria), of which 13 are measured outputs. Its 32 reactions include protein product formation, the Embden-Meyerhof-Parnas pathway (EMP), the TCA cycle, a reduced amino acid metabolism, lactate production, and the electron transport chain. The reactions are modelled using lin-log kinetics [49], resulting in non-rational equations with 117 unknown parameters. While it was noted in [32] that the parameter estimation results suggested practical identifiability issues, possible deficiencies in structural identifiability were not mentioned. Given the size of this model, its analysis is very challenging.
Using decomposition it is possible to classify most of the parameters in the model as identifiable. However, we also found that at least four parameters are structurally unidentifiable: they are the parameters numbered 47, 48, 55, and 57, which correspond to the following kinetic constants (elasticities): {e54, − e55, − e62, e64}. After inspecting the model, we realised that it is possible to rewrite its dynamic equations in such a way that these parameters appear as (e54 + e55) and (e62 + e64); clearly, the individual parameters appearing in these sums are not identifiable. Thus, we replaced these four parameters with two new ones, en1 = e54 + e55 and en2 = e62 + e64. In this way we obtained a new model with 115 parameters, and confirmed that the newly introduced ones are structurally identifiable. Overall, we determined the identifiability of 97 parameters. While we did not manage to assess the identifiability of the remaining 18, we did find that fixing six of them (e.g. {p28, p72, p77, p101, p105, p115}) results in a structurally identifiable model. This action is slightly conservative, since those parameters can in principle be s.l.i. However, since the model has practical identifiability deficiencies [32] (as is typical of models of this type and size [49]), and given that it would be necessary to perform many Lie derivatives to relate these parameters to the model outputs, it is likely that in practice their values will be difficult to estimate. Therefore, fixing a subset of them appears as a reasonable solution.
In summary, we found that: (i) this model is structurally unidentifiable, (ii) there exist two identifiable combinations of parameters, which convert 4 unidentifiable parameters into 2 identifiable ones, (iii) of the remaining 113 parameters, at least 95 are identifiable, and (iv) fixing the values of 6 parameters ensures that the remaining 12 (and the model as a whole) are identifiable.
We have presented a methodology for analysing the structural identifiability of dynamic models described by a system of ordinary differential equations. It builds on concepts and techniques originally presented in the context of nonlinear observability analysis. More specifically, it adopts a differential geometry approach, which is based on building an augmented observability matrix—with the parameters considered as additional state variables—and calculating its rank. This formulation, as opposed to other approaches based on differential algebra, allows handling any analytic models, without requiring them to be in rational or polynomial form. If a model is structurally unidentifiable the method determines the identifiability of each parameter individually, by recalculating the matrix rank after removing the corresponding column.
Realising that the structural identifiability analysis of nonlinear dynamic models is a challenging task, and that this difficulty increases rapidly with the problem size, our method is geared towards computational efficiency. To this end it includes several algorithmic developments to facilitate the analysis of models of larger size. One is the possibility of decomposing the model into smaller submodels, which can be found by optimization or specified by the user. Another is the calculation of the matrix rank with a reduced number of Lie derivatives. These alternatives lead in some cases to partial analyses, whose result is only conclusive if a parameter is classified as identifiable, but not as unidentifiable (or vice versa, depending on the type of analysis). In these situations the method also allows for an iterative refinement of the diagnosis: by removing parameters already classified as identifiable, the problem size is reduced and more complete analyses are made possible.
To facilitate the application of this methodology, we have provided it as a free MATLAB (The MathWorks, Natick, MA) toolbox called STRIKE-GOLDD (STRuctural Identifiability taKen as Extended-Generalized Observability with Lie Derivatives and Decomposition), available under the GNU General Public License from https://sites.google.com/site/strikegolddtoolbox/. We expect that STRIKE-GOLDD will contribute to fill the gap between the complexity of current systems biology models and their usability, which can be compromised unless structural identifiability is assessed.
We have validated the methodology using a set of nonlinear systems biology models whose size and/or complexity make them challenging case studies. They range from a classic model of enzymatic oscillations with 8 parameters proposed by Goodwin in 1965 [27] to a metabolic model of more than 100 parameters published in 2015 [32]. Interestingly, we found structural identifiability issues even in models of relatively small size, such as the aforementioned Goodwin model. Indeed, the results show that identifiability issues are likely to appear in over-parameterized models (with many parameters per state), specially if only few of their states are available for measurement (in order words, if there are few outputs).
A large parameter-to-output ratio also implies that the structural identifiability of the model will be difficult to analyse, because it will be necessary to perform many Lie derivative calculations in order to build the augmented observability matrix, thus incurring a high computational cost. Could this common cause mean that the difficulty in analysing a model is a hint of lack of identifiability? We ask this question because we know that, on the other hand, it is possible to analyse models with many parameters as long as sufficient measurements are available.
Among the models analysed here, the JAK/STAT pathway had already been studied [20], and for that case our method confirmed previously reported results. In other cases we established the identifiability of systems that had not been analysed before, such as the mixed feedback MAPK pathway [28] or the model of Pitavastatin hepatic uptake (which had been reported to resist analysis when attempted with other methods, although it was suspected that it was identifiable [19]). Perhaps more interestingly, we also found some unidentifiabilities that had not been previously reported. An example is the Goodwin oscillator [27], for which it was established that half of its parameters are structurally unidentifiable. Despite the relatively small size of this model (3 states and 8 parameters), the fact that it is not a rational system, combined with the high parameter-to-output ratio (given that only one of its states is measured) make it a very challenging problem. Similar issues were found in the NF-κB signalling pathway [30] and in the genetic subnetwork of the circadian clock in Arabidopsis thaliana [29]. In these cases it can be noted that the ratio of unidentifiable parameters is larger in models with a lower ratio of measured outputs. Finally, we also detected unidentifiabilities in a recently presented large-scale dynamic model of metabolism of Chinese Hamster Ovary cells [32] with 117 parameters.
We have also shown how to turn unidentifiable models into identifiable ones. With the procedure described in this paper it is sometimes possible to combine several unidentifiable parameters into a single identifiable combination. More often the solution is to reparameterize the model by considering some of the unidentifiable parameters as known constants, fixing them to values that appear reasonable according to available knowledge. In this way the remaining unknown parameters are rendered identifiable. Finally, a model can also be made identifiable by increasing the number of its outputs, if it is experimentally possible to measure more of its states.
|
10.1371/journal.pgen.1007532 | A genome-wide association study identifies a susceptibility locus for biliary atresia on 2p16.1 within the gene EFEMP1 | Biliary atresia (BA) is a rare pediatric cholangiopathy characterized by fibrosclerosing obliteration of the extrahepatic bile ducts, leading to cholestasis, fibrosis, cirrhosis, and eventual liver failure. The etiology of BA remains unknown, although environmental, inflammatory, infectious, and genetic risk factors have been proposed. We performed a genome-wide association study (GWAS) in a European-American cohort of 343 isolated BA patients and 1716 controls to identify genetic loci associated with BA. A second GWAS was performed in an independent European-American cohort of 156 patients with BA and other extrahepatic anomalies and 212 controls to confirm the identified candidate BA-associated SNPs. Meta-analysis revealed three genome-wide significant BA-associated SNPs on 2p16.1 (rs10865291, rs6761893, and rs727878; P < 5 ×10−8), located within the fifth intron of the EFEMP1 gene, which encodes a secreted extracellular protein implicated in extracellular matrix remodeling, cell proliferation, and organogenesis. RNA expression analysis showed an increase in EFEMP1 transcripts from human liver specimens isolated from patients with either BA or other cholestatic diseases when compared to normal control liver samples. Immunohistochemistry demonstrated that EFEMP1 is expressed in cholangiocytes and vascular smooth muscle cells in liver specimens from patients with BA and other cholestatic diseases, but it is absent from cholangiocytes in normal control liver samples. Efemp1 transcripts had higher expression in cholangiocytes and portal fibroblasts as compared with other cell types in normal rat liver. The identification of a novel BA-associated locus, and implication of EFEMP1 as a new BA candidate susceptibility gene, could provide new insights to understanding the mechanisms underlying this severe pediatric disorder.
| The etiology of biliary atresia (BA) is unknown and likely complex. Environmental, infectious, and genetic risk factors have all been proposed, and the leading hypothesis in the field is that a combination of these factors is responsible for disease manifestation. To identify susceptibility loci for BA, we performed a genome wide association study on two groups of BA patients (one composed of patients with isolated BA (n = 343) and one of patients with BA and other extrahepatic anomalies (n = 156)) and genetically matched controls. We detected a set of SNPs within the EFEMP1 gene associated with BA in both cohorts. We further showed by immunohistochemistry that EFEMP1 protein was expressed in cholangiocytes and vascular smooth muscle cells in BA livers, and that EFEMP1 RNA expression levels were elevated in both BA and other cholestatic disease livers. These findings suggest that EFEMP1 should be considered as a new candidate susceptibility gene for BA.
| Biliary atresia (BA; OMIM 210500) is a progressive, necro-inflammatory disease, affecting the extra- and intrahepatic biliary system, leading to bile flow obstruction, cholestasis, and jaundice in infants [1–3]. If left untreated, the condition progresses to hepatic fibrosis, cirrhosis, liver failure, and eventual death within the first two years of life [4]. BA is a rare disease with varying incidence from approximately 1/18,000 live births in Western Europe to 1/8,000 in Asia [5–8]. About 85% of BA patients have a non-syndromic form, in which BA is an isolated finding, while roughly 15% present with other congenital anomalies, including laterality defects in some patients [2, 9]. The primary treatment for BA includes surgical restoration of bile flow (the Kasai hepatoportoenterostomy), which is only successful in about 50% of patients [10]. With or without successful bile drainage, most patients have progressive liver disease. Fifty percent of patients require a liver transplant by age 2 and most of the remaining before reaching adulthood [4, 11]. BA is the most frequent indication for liver transplantation in children worldwide [12].
Although the etiology of BA is not well understood, it is proposed to be multi-factorial and heterogeneous. The current theory posits that BA arises from a combination of genetic predisposition, cholangiocyte damage from environmental factors such as a toxin exposure [13] or viral infection [14], and an inflammatory response to damage [15, 16]. A genetic susceptibility for BA is supported by reports of familial cases, including both parent to child transmission and affected siblings [17–20], and disparate incidences among populations, even after controlling for environmental influences in regional epidemiologic studies [6, 21].
In order to identify genes implicated in susceptibility to BA, candidate gene association studies have been performed, mainly on genes involved in immune or inflammatory responses, and suggestive associations have been found with the genes ITGB2, ADIPOQ, IFNG, VEGFA, MIF [22–27]. Additional studies, including genome-wide copy number variant (CNV) and single nucleotide polymorphism (SNP) association studies, have identified a few other candidate susceptibility genes. The GPC1 gene was implicated in BA following a genome-wide CNV association study and its role was further supported by functional studies showing that gpc1 knockdown in zebrafish led to impaired development of the biliary network [28]. A susceptibility locus on 10q25 near the XPNPEP1 and ADD3 genes was identified by a genome-wide SNP association study conducted in Han Chinese BA patients and controls [29], and was replicated in an independent Chinese cohort [30] and in a Thai cohort [31]. Fine-mapping of this signal in North-American patients of European descent [32] and functional studies in an animal model [33] suggested that ADD3 was the most likely BA susceptibility gene at this locus. A second genome-wide association study (GWAS) performed in a Caucasian cohort identified a signal in the 3' enhancer of ARF6, and subsequent studies in zebrafish suggested that knockdown of this gene resulted in biliary defects [34]. Together, these studies have exposed a spectrum of genetic associations in BA and highlight a complex, heterogeneous genetic susceptibility as an emerging feature of the disease etiology.
In an effort to add to our understanding of the genetic susceptibility to BA, we performed a GWAS on 343 patients with isolated BA and 1716 controls of European descent. This analysis identified a novel signal on 2p16.1, in the intronic region of the gene EFEMP1, that showed the highest association with BA and reached genome-wide significance after meta-analysis with data from a second cohort of 156 patients with BA and other extrahepatic anomalies and 212 controls, also of European descent. Downstream expression analysis of EFEMP1 RNA and protein localization in liver specimens suggest that EFEMP1 may have functional relevance not only in BA, but also in other cholestatic liver diseases.
We carried out a GWAS in a cohort of 343 isolated BA patients ascertained through the NIDDK-funded Childhood Liver Disease Research Network (ChiLDReN) and 1716 genetically-matched controls of European descent on 1,171,073 common markers (minor allele frequency (MAF) ≥ 5%). Although none of the tested markers reached standard genome-wide significance (P < 5 × 10−8), 19 markers reached suggestive significance (P < 1 × 10−5). These 19 markers were located in genomic regions 1p31.1, 1q32.2, 2p23.2, 2p16.1, 2q37.3, 3p24.2, 6p22.3, 6q24.3, 16p13.1, and 20q13.2 (Fig 1A and S1 Table).
The most significant BA-associated SNP was rs10865291, located in the fifth intron of the EFEMP1 gene on 2p16.1 (P = 5.85 ×10−7; OR = 1.56) (S1 Table and S1A Fig). The minor allele A was overrepresented in cases (43%) compared to controls (33%). Five neighboring markers in linkage disequilibrium (LD) with rs10865291 (r2 > 0.5) also reached suggestive significance (S1 and S2 Tables), indicating that this association signal was unlikely to be due to genotyping errors. After imputation, seven additional SNPs in the same region showed suggestive significance, although the originally identified SNP, rs10865291, remained the most significant (S2A Fig). Association tests conditional on rs10865291 decreased significance for all genotyped and imputed markers within and downstream of the EFEMP1 gene (S2 Table, S1B and S2B Figs), suggesting that these SNPs contribute to the same association signal.
As a complementary approach to the single SNP association test, we performed a genome-wide gene-based association test using VEGAS2 [35]. The results showed that EFEMP1 was the top BA-associated gene with empirical P-value estimated as 3.9 × 10−5 (S3 Table).
We previously replicated, in a smaller subset of the same cases and a different set of controls, the association of BA with SNPs from the 10q25 region identified in a Chinese cohort [29, 32]. To confirm this finding in the current, larger cohort, we imputed the genotypes at rs17095355, the most significant BA-associated SNP identified in the Chinese GWAS, and neighboring markers, since rs17095355 was not present on the Illumina genotyping array used in this study. A few markers, including rs17095355, reached nominal significance (P < 0.05). The most significant SNP at the 10q25 locus was rs59804002 (P = 0.003), located in the gene ADD3-AS1, which encodes a long non-coding RNA (ncRNA) (S3 Fig). In contrast, we were not able to replicate the association of ARF6 SNPs with BA in our dataset (rs3126184 P = 0.85, and rs10140366 P = 0.82 after imputation) [34].
To confirm the association with BA of the 2p16.1 SNPs, we performed a second GWAS in an independent group of 156 patients with BA and other extrahepatic anomalies and 212 genetically-matched controls of European descent. All subjects in this cohort were genotyped with the Illumina OmniExpress array, which has a lower SNP density than the Illumina Omni2.5 array used in the isolated BA cohort. Therefore, in order to combine results from the two datasets, we performed genome-wide imputation using 1000 Genomes Project Phase 3 data as the reference panel followed by meta-analysis. The results showed that all SNPs in the 2p16.1 region reaching the threshold for suggestive significance in the isolated BA cohort were more significant following meta-analysis. Importantly, three highly correlated markers (rs10865291, rs6761893, and rs727878; r2>0.8) reached genome-wide significance, with imputed SNP rs6761893 showing the highest significance (P = 3.39 ×10−8) (Table 1 and Fig 1B). No other genomic region, including the other nine regions that had reached suggestive significance in the isolated BA cohort GWAS, contained SNPs that reached genome-wide significance (S4 Fig and S4 Table).
To investigate whether the associated SNPs on 2p16.1 were expression quantitative trait loci (eQTLs) with a potential functional role in regulating EFEMP1 expression in liver tissues, we tested the correlation between the six genotyped SNPs with P < 1x10-5 (Table 1) and EFEMP1 expression, using data from a published liver transcriptome study of BA [36]. Genotyping and expression data from 20 BA patients included in both the liver transcriptome study and our GWAS were combined to examine whether the BA-associated SNP genotypes were correlated with EFEMP1 expression. No correlation was found between the genotypes of any of the six SNPs and EFEMP1 expression. The results for the two top BA-associated genotyped SNPs (rs10865291 and rs727878) are shown in S5 Fig. Inability to detect correlation may be due to the small sample size (n = 20) or the advanced liver disease status of the BA patients, which may have masked BA-specific effects.
We therefore sought to analyze data from the Genotype-Tissue Expression (GTEx) database V7 (gtexportal.org/), which contains eQTL data from a variety of tissue types from 620 healthy human donors. The level of EFEMP1 expression was very low in GTEx human liver tissue and we did not detect any liver eQTLs for EFEMP1 among the BA-associated SNPs within 2p16.1. However, we found that EFEMP1 expression in GTEx human thyroid tissue was positively correlated with the risk alleles of one of the 2p16.1 BA-associated SNPs, and that this correlation was highly significant (rs1346786, GTEx eQTL P = 3.6 × 10−8).
We next examined the gene expression level of EFEMP1 in snap-frozen liver samples collected at the time of liver transplant or surgery from patients with BA (n = 5), other cholestatic diseases (n = 7), and non-cholestatic controls (n = 5) using droplet digital PCR (ddPCR) (S5 Table). We found that EFEMP1 expression was significantly different among the three groups (one-way ANOVA P = 0.007), being increased in both BA livers and other cholestatic disease livers compared to non-cholestatic control livers (Fig 2A). This finding is consistent with a previous observation that EFEMP1 expression is upregulated in BA patients by 2.85 fold when compared to controls without cholestatic liver diseases [36]. We observed variable gene expression levels among BA patient samples, and there seemed to be a correlation between age at transplant and transcript expression levels (S6 Fig). Patients who received a transplant at an earlier age showed the highest levels of EFEMP1 RNA expression, while patients who received a transplant at an older age showed the lowest levels of RNA expression.
To examine the cell type-specific expression pattern of EFEMP1, we quantified the orthologous Efemp1 transcripts in six rat liver cell populations using ddPCR. In normal rats, we found that Efemp1 is significantly enriched in cholangiocytes and portal fibroblasts compared to hepatocytes, Kupffer cells, sinusoidal endothelial cells, and hepatic stellate cells (one-way ANOVA P<0.0001) (Fig 2B).
To investigate the expression and localization of EFEMP1 protein in human liver, we performed immunohistochemistry (IHC) on normal control and BA livers, and we also included disease control livers from patients with total parenteral nutrition-associated (TPN) cholestasis and autosomal recessive polycystic kidney disease (ARPKD). We found that EFEMP1 is specifically expressed in α-smooth muscle actin (α-SMA)-positive vascular smooth muscle cells (Fig 3A and S7A and S7B Fig), but not in cytokeratin 19 (CK19)-positive intrahepatic cholangiocytes (Fig 3B and S7C and S7D Fig) in normal control livers. In contrast, it is expressed in both vascular smooth muscle cells and intrahepatic cholangiocytes in BA livers and in livers of patients with TPN cholestasis and ARPKD (Fig 3C–3H and S7E–S7P Fig). Immunohistochemical staining for EFEMP1 in the extrahepatic bile duct remnant removed from a BA patient during Kasai hepatoportoenterostomy showed that it is also expressed in extrahepatic cholangiocytes, as well as in vascular smooth muscle cells (Fig 4 and S8 Fig).
Using a GWAS approach, we identified a novel candidate susceptibility locus for BA that maps within the EFEMP1 gene on 2p16.1. EFEMP1 encodes the EGF-containing fibulin-like extracellular matrix protein, Fibulin-3 (EFEMP1), which is one of seven members of the fibulin family and has roles in extracellular matrix remodeling, tissue regeneration, and organogenesis [37, 38]. EFEMP1 has also been reported to activate Notch signaling in vitro, although with less efficiency than JAG1 [39], and to exert both tumor suppressive and oncogenic effects in various cancers by participating in multiple signaling pathways [40–43]. A missense mutation in EFEMP1 (R345W) has been found in patients with Doyne honeycomb retinal dystrophy (DHRD) [44]. Consistent with this finding, Efemp1-R345W knock-in mice develop a similar ocular phenotype, supporting the causal role of the EFEMP1 R345W mutation in DHRD [45]. In contrast, targeted inactivation of Efemp1 in mice resulted in premature aging, shortened lifespan and reduced reproductive capacity compared to wild-type mice, without signs of macular degeneration [46]. In aged Efemp1-/- mice (18 to 24 months of age), the authors observed overall decreased body mass, muscle and fat mass compared to littermate controls. In addition, the aged mutant mice had significantly decreased mass of internal organs including liver, spleen and kidney. A biliary phenotype was not reported. Of note, adult Efemp1-/- mice on the C57BL/6 background developed herniation of abdominal and pelvic organs. On further examination, this phenotype was found to be due to reduction of elastic fibers in the abdominal fascia [46]. Variants in the EFEMP1 locus have been associated with a variety of human traits and diseases, such as human height [47], forced vital capacity [48], inguinal hernia [49], and chronic venous disease (CVD) [50]. We note that it is not unusual for genes to be associated with both complex and Mendelian disorders affecting different organ systems [51].
Using two different cohorts in two separate GWAS, we were able to identify a region of interest within 2p16.1 in a European-American population of isolated BA patients, and we subsequently observed additional support for this association in an independent European-American cohort of patients with BA and other extrahepatic anomalies. When results from these two cohorts were combined using meta-analysis, three markers (rs10865291, rs6761893, and rs727878) reached genome-wide significance, providing evidence towards disease relevance. The associated alleles at these three SNPs are very common in the East Asian population (rs10865291:A = 0.77; rs6761893:T = 0.88; and rs727878:T = 0.77; 1000 Genomes Project Phase 3 data), which is consistent with a higher documented incidence of BA in Asia [52]. We also confirmed replication in our European American cohort of an association signal within 10q25, originally identified in a GWAS performed in a Chinese cohort [29].
All of the BA-associated SNPs within 2p16.1 identified in our GWAS, including the top three markers (rs10865291, rs6761893, and rs727878), are located in the 3' UTR or within intronic regions of the EFEMP1 gene. Although we did not detect any correlation between the genotypes of BA-associated SNPs and EFEMP1 gene expression in livers from BA patients, analysis of data from the GTEx project showed that one of the BA-associated SNP (rs1346786) is an eQTLs for EFEMP1 in healthy human thyroid tissue. This result suggests that the BA-associated SNPs (or other SNPs in linkage disequilibrium with them) may play a regulatory role in EFEMP1 expression in specific tissues or at specific time-points. RNA expression data showed that EFEMP1 was strongly upregulated in BA liver when compared to healthy liver. This result is consistent with a published comparative liver transcriptome analysis, which showed that EFEMP1 expression was upregulated in BA patients by 2.85 fold in comparison with controls without cholestatic liver diseases [36]. We also found EFEMP1 protein expression in vascular smooth muscle cells in both control and disease liver specimens, an expression profile that is consistent with a published report of EFEMP1 as a candidate susceptibility gene for CVD [50].
The identification of EFEMP1 as a susceptibility locus in both isolated BA and BA with other anomalies, which are clinically-distinct classifications of BA, suggests that it may be important in liver and bile duct development. Recent evidence has shown that direct bilirubin levels are elevated in the first days of life even in infants with isolated BA, supporting a model of prenatal onset of disease pathogenesis in the absence of other congenital anomalies [53]. Fibulins are known to have extensive interactions with laminins and other extracellular matrix (ECM) -related proteins [38]. Given the close proximity and cellular communication of the developing bile ducts with the ECM of the portal mesenchyme, it is possible that EFEMP1 may have a previously unrecognized function in bile duct development.
We found Efemp1 transcripts to have higher expression in cholangiocytes and portal fibroblasts as compared with other cell types in normal rat liver. In human whole liver tissue, EFEMP1 RNA expression was high not only in BA, but also in non-BA cholestatic disease as compared with control liver tissue. It is interesting that expression of Efemp1 transcripts was high in rat cholangiocytes and low in human control whole liver tissue, although we note that the major cell type in the liver is hepatocytes, which express very low levels of Efemp1 RNA in the rat (Fig 2). We further investigated cell type-specific expression of EFEMP1 protein in human liver tissue by IHC, which showed EFEMP1 expression in CK19 positive cholangiocytes in both BA and other cholestatic disease patient liver, in contrast to healthy control liver where EFEMP1 protein was not expressed in bile ducts. This distinct expression pattern suggests that EFEMP1 may have a bile duct-specific role in cholestatic liver disease. More specifically, the similar expression pattern of EFEMP1 in both BA and other cholestatic diseases suggests a common function for EFEMP1 in cholestatic liver disease. Given that progressive fibrosis is a common feature of disease pathogenesis for these various conditions, EFEMP1 may have a role in fibrosis. Indeed, there was advanced fibrosis in all the BA and disease control samples used in this study. This is further supported by known roles for EFEMP1 as an ECM protein, and the importance of the ECM in a variety of biological processes, including fibrosis [37]. Moreover, it has been reported that portal fibroblasts are important sources of matrix proteins in hepatic fibrosis [54]. Our liver cell type-specific expression analysis of Efemp1 in normal rats showed it was highly expressed in portal fibroblasts compared with other liver cell populations, which also supports a possible role in hepatic fibrosis.
Identification of the causal gene and the underlying causal mechanism following a significant GWAS signal is a challenging task [55]. Given the intronic localization of our genome-wide significant associated SNPs, we focused initially on evaluating a potential role for EFEMP1 in BA by gene expression and protein localization studies. It is still possible that other genes may be involved, in addition to or in place of EFEMP1, through complex long-range interactions or regulatory mechanisms. Other studies have shown that the genes targeted by GWAS-identified SNPs are not always or only the nearest ones [56]. Our data support a role for EFEMP1 in the etiology of BA, and its possible involvement in other cholestatic liver diseases. However, with the available data it is difficult to speculate on what the underlying mechanism linking the GWAS identified EFEMP1 common variants to susceptibility to BA might be, and whether their effect could be mediated through a gain or loss of function of the EFEMP1 protein. Significant evidence supports a model of BA resulting from a combination of an environmental insult (such as exposure to toxins or viruses) and host susceptibility to the insult and/or to a damaging and progressive response to it [57]. In this context, it is perhaps not surprising that, in the absence of an environmental insult, the available Efemp1 mouse models do not show a biliary phenotype [45, 46]. Further studies will be necessary to define the possible role of EFEMP1 in the development of BA, and in progression of fibrosis in BA or other cholestatic liver diseases. Ultimately, its identification as a candidate BA susceptibility gene could provide new insights to understanding the mechanisms underlying BA pathogenesis.
BA patient samples were obtained from the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK)-funded Childhood Liver Disease Research Network (ChiLDReN). Patients were enrolled into the ChiLDReN study at each participating site under IRB-approved protocols, and patients’ data and blood samples had been obtained with informed consent from patients’ parents or legal guardians. DNA samples from participants were isolated and banked at the Rutgers University NIDDK biorepository. De-identified DNA samples were sent to our laboratory for analysis. All patients were diagnosed with BA based on clinical presentation, liver histology, and intraoperative cholangiogram. The majority had their diagnosis confirmed by examination of the biliary remnant from portoenterostomy. A total of 450 self-reported white patients with isolated BA (260 females and 190 males), including 74 Hispanics, and 170 self-reported white, non-Hispanic, patients with BA and other extrahepatic anomalies (104 females and 66 males) were genotyped.
Control genotypes for the isolated BA samples were obtained from de-identified samples of 1981 self-reported white, healthy individuals (1167 females and 814 males) from the NIH-funded Age-Related Eye Disease Study (AREDS) [58]. Controls for the patients with BA and other anomalies were de-identified samples of 246 self-reported white children (187 females and 59 males) genotyped at the Center for Applied Genomics (CAG) at the Children’s Hospital of Philadelphia (CHOP). These children had no diagnosis of congenital anomalies or any diseases in the digestive system.
De-identified human liver specimens used for the gene expression assay were obtained from a CHOP liver repository. All samples were collected at time of liver transplant or surgery, and included non-cholestatic livers (n = 5), cholestatic disease livers (n = 7) and BA livers (n = 5). Non-cholestatic controls included patients with citrullinemia (n = 2) and propionic acidemia (n = 1), and normal liver tissue adjacent to tumor (n = 2). The cholestatic disease control group consisted of children with Alagille syndrome (n = 3), cystic fibrosis (n = 1), primary sclerosing cholangitis (n = 2), and autoimmune hepatitis (n = 1). The characteristics of these 17 human liver specimens are summarized in S5 Table.
DNA samples from isolated BA patients and AREDS control subjects were genotyped with the Illumina Omni2.5 BeadChip (Illumina, San Diego, CA). Per-sample and per-marker quality control was performed using PLINK [59]. SNPs with genotype missing rate >5%, minor allele frequency (MAF) <5%, missing rate significantly different between cases and controls (P <0.00001), or significant deviation from Hardy-Weinberg equilibrium in controls (P <0.00001) were excluded. A total of 1,171,073 markers were included in the association analysis. Samples with a genotype call rate < 97%, discordance between reported and genotyped sex, and an outlying heterozygosity rate (defined as outside of 3 standard deviation from the mean) were excluded. Duplicates and samples with hidden relatedness were identified by estimating identity-by-descent (IBD) on the basis of the genome-wide identity-by-state (IBS) information, and one from each pair of samples with proportion of IBD larger than 0.105 was excluded.
To control for population stratification, principal component analysis (PCA) was carried out in the remaining 432 BA cases and 1876 controls by incorporating the genotypes of HapMap3 individuals from 11 populations (S9 Fig). Genetically-matched cases and controls were selected based on the first two principal components (PCs) using an algorithm called Ordering Points to Identify the Clustering Structure (OPTICS) [60] with the maximum distance in a cluster set as 0.003 (S10 Fig). A total of 343 cases and 1716 controls that clustered with the HapMap3 samples of European ancestry were selected for the association test. A Tracy-Widom test was performed to estimate the number of statistically significant PCs to be included in the association test to correct for any potential residual population stratification.
DNA from 170 patients with BA and other anomalies and 246 CAG controls were genotyped with the Illumina OmniExpress BeadChip. Quality control was applied exactly as described for the isolated BA cohort. A total of 156 BA and 212 genetically-matched controls of European ancestry were selected for association test on 579,213 markers.
Adjusted logistic regression under the additive model, with the statistically significant PCs as covariates, was carried out for case-control association analysis in the isolated BA cohort for 1,171,073 genotyped common SNPs (MAF ≥ 5%). Genome-wide significance was set at P < 5x10-8, and suggestive association was defined as an average of 1 false-positive association per GWAS in European populations, or P < 1x10-5 [61]. No deviation from the expected P-values was observed in the Q-Q plot and the genomic inflation factor λ was estimated as 1 (S11 Fig). The P-values of each marker within 20kb upstream and downstream of a gene were used for gene-based association analysis in the isolated BA cohort using VEGAS2 [35].
IMPUTE2 [62] v2.3.2 was used for imputation of all SNPs and indel variants annotated in the 1000 Genomes Project Phase 3 in both BA cohorts and their respective controls. Variants that were imputed with low confidence (info < 0.8), or with MAF < 0.05 were removed from subsequent association test. Testing for association with BA under an additive genetic effect model, adjusting for the statistically significant PCs, was performed using the frequentist likelihood score method implemented in SNPTEST [63] v2.5 in the two datasets separately. Meta-analysis across the two cohorts was performed with METAL [64] using the inverse variance method, which weights the effect size estimates (β-coefficients) by their estimated standard errors.
Regional association plots were obtained with LocusZoom [65].
Total RNA was extracted from 17 human liver specimens using TRIzol reagent (Invitrogen, Carlsbad, CA) and the RNeasy RNA extraction kit (Qiagen, Venlo, The Netherlands) according to the manufacturer’s instructions. Subsequent cDNA was synthesized using the TaqMan Reverse Transcription Kit under normal conditions (Applied Biosystems, Foster City, CA). ddPCR was performed on a Bio-Rad QX100 ddPCR system (Bio-Rad, Hercules, CA) using standard methods. Droplets were generated from reactions containing 60 ng of cDNA from each of the 17 liver biopsy specimens using the TaqMan fluorescent reporter assay (Thermo Fisher Scientific, Waltham, MA). Human EFEMP1 primer and probe set labeled with FAM (Hs00244575_m1) and human TBP primer and probe set labeled with VIC (Hs00427620_m1) were used. Samples were multiplexed and run under standard ddPCR conditions [66]. One-way analysis of variance was performed to test for a difference among the three groups (non-cholestatic control, cholestatic disease control and BA) and P < 0.05 was considered as statistically significant.
Hepatocytes, Kupffer cells, portal fibroblasts, sinusoidal endothelial cells, and hepatic stellate cells were isolated from normal rats [67], and cholangiocytes were isolated as the non-adhering cells after 1 hour of plating, using the same protocol as portal fibroblasts. RNA was extracted from single specimens for each cell type as described [67]. cDNA was generated using the TaqMan Reverse Transcription Kit (Applied Biosystems, Foster City, CA). Droplets for ddPCR were generated from reactions containing 30 ng of cDNA from each of the cell populations using TaqMan (Thermo Fisher Scientific, Waltham, MA) rat Efemp1 primer and probe set labeled with FAM (Rn01434325_m1) and TaqMan rat Rps12 primer and probe set labeled with VIC (Rn00821373_g1) as the control. One-way analysis of variance was performed to test for a difference among three cell type groups (cholangiocytes, portal fibroblasts and all others) and P < 0.05 was considered as statistically significant.
Formalin-fixed, paraffin-embedded liver slides were obtained from the Department of Pathology at CHOP. These slides were taken from de-identified liver samples not required for clinical use. Samples included BA tissue from liver explants, BA tissue from extrahepatic bile ducts (from Kasai hepatoportoenterostomy), control liver tissue from an autopsied infant with no known liver disease, tissue from an infant with TPN, and tissue from an infant with clinically-diagnosed ARPKD.
Tissue was de-paraffinized using standard procedures. All tissue was pre-treated by exposing to 3% hydrogen peroxide in methanol for 5 minutes followed by heat-mediated antigen retrieval in citrate buffer (Sigma-Aldrich, St. Louis, MO) for 10 minutes. Slides were blocked for one hour at room temperature (10% donkey serum, 5% milk, 4% BSA, 0.1% Triton X-100 in PBS) and incubated in primary antibody overnight at 4°C. All primary antibodies were purchased from Abcam (Cambridge, MA), and included CK19 (1:500; ab7754), α-smooth muscle actin (1:50; ab7817), and EFEMP1 (1:100; ab151976). Slides were incubated in secondary antibodies at a dilution of 1:500 for 1 hour at room temperature (Alexa Fluor 488 α-mouse and Alexa Fluor 555 α-rabbit, Life Technologies, Carlsbad, CA). All slides were mounted with Vectashield (Vector Laboratories, Inc., Burlingame, CA) and photographs were taken on a Leica DMi8 inverted microscope using Leica Application Suite (LAS X) software (Wetzlar, Germany).
|
10.1371/journal.pntd.0000297 | The Potential Impact of Density Dependent Fecundity on the Use of the Faecal Egg Count Reduction Test for Detecting Drug Resistance in Human Hookworms | Current efforts to control human soil-transmitted helminth (STH) infections involve the periodic mass treatment of people, particularly children, in all endemic areas, using benzimidazole and imidothiazole drugs. Given the fact that high levels of resistance have developed to these same drugs in roundworms of livestock, there is a need to monitor drug efficacy in human STHs. The faecal egg count reduction test (FECRT), in which faecal egg output is measured pre- and post-drug treatment, is presently under examination by WHO as a means of detecting the emergence of resistance. We have examined the potential impact of density dependent fecundity on FECRT data. Recent evidence with the canine hookworm indicates that the density dependent egg production phenomenon shows dynamic properties in response to drug treatment. This will impact on measurements of drug efficacy, and hence drug resistance. It is likely that the female worms that survive a FECRT drug treatment in some human cases will respond to the relaxation of density dependent constraints on egg production by increasing their egg output significantly compared to their pre-treatment levels. These cases will therefore underestimate drug efficacy in the FECRT. The degree of underestimation will depend on the ability of the worms within particular hosts to increase their egg output, which will in turn depend on the extent to which their egg output is constrained prior to the drug treatment. As worms within different human cases will likely be present at quite different densities prior to a proposed FECRT, there is potential for the effects of this phenomenon on drug efficacy measurements to vary considerably within any group of potential FECRT candidates. Measurement of relative drug efficacy may be improved by attempting to ensure a consistent degree of underestimation in groups of people involved in separate FECRTs. This may be partly achieved by omission of cases with the heaviest infections from a FECRT, as these cases may have the greatest potential to increase their egg output upon removal of density dependent constraints. The potential impact of worm reproductive biology on the utility of the FECRT as a resistance detection tool highlights the need to develop new drug resistance monitoring methods which examine either direct drug effects on isolated worms with in vitro phenotypic assays, or changes in worm genotypes.
| Current efforts to control soil-transmitted helminth (STH) infections in humans in endemic countries involve the mass administration of drugs. The use of these same drugs for many years to control livestock nematodes has resulted in the emergence of significant levels of resistance. Hence, there is a need to manage the use of drugs against human STHs in order to reduce the likelihood of resistance developing. An important component of managing drug use will be an ability to detect drug resistance should it emerge. WHO and the World Bank are presently supporting initiatives to develop tools for detecting drug resistance in human STHs. The tool to be assessed in the short term is the faecal egg count reduction test (FECRT). We have examined literature on an aspect of worm reproductive biology with potential to impact significantly on the FECRT. We describe the potential effects that density dependent egg production by female hookworms may have on interpretation of FECRT data. This study highlights a potential weakness in reliance on the FECRT for assessment of drug resistance in human hookworms, hence emphasising the need to develop more advanced worm bioassay and molecular methods.
| Soil transmitted helminth (STH) infections (Ascaris lumbricoides, Trichuris trichiura, and the hookworms Necator americanus and Ancylostoma duodenale) contribute significantly to morbidity in humans in endemic countries. The major means of controlling these infections is by the periodic administration of anthelmintic drugs [1]. Given the lessons from livestock, where resistance to anthelmintics is widespread [2], the potential for anthelmintic resistance in human parasites has been recognized [3]. There are a number of differences in drug use patterns between the human and livestock fields which may aid in reducing the rate of resistance development in the former [4],[5], however, the potential for resistance in human parasites is a significant issue [6].
It is clear that monitoring of drug resistance will be an important part of ensuring the success of current efforts to control STH infections. Although sensitive molecular tests are able to detect SNPs associated with benzimidazole resistance in livestock nematodes, they are still not available for human STHs. Phenotypic assays of worm eggs or larvae in vitro are useful for detection of resistance to some drug groups in livestock nematodes, however, their use with human STHs requires substantial validation and field testing. Hence, the faecal egg count reduction test (FECRT), in which egg output per gram of faeces (epg) is compared in individuals before and after drug treatment to estimate drug efficacy, is considered to be the most useful technique currently available for detecting the emergence of resistance in STH infections in humans.
A recent study by Kopp et al. [7], in which dogs infected with the hookworm Ancylostoma caninum were treated with pyrantel, showed a poor relationship between drug efficacy and changes to epg. A mean drug efficacy of 71% in two dogs infected with an isolate showing low level resistance to the drug was associated with a 41% increase in epg. That is, egg output increased markedly despite the removal of a significant proportion of the adult worm burden by the drug. They suggested that this was due to an increase in egg production in surviving adult female worms after drug treatment due to the relaxation of density-dependent constraints on egg production.
Density dependent fecundity has been demonstrated many times with Helminth parasites [reviewed by 8],[9]. The effect is seen as a decrease in egg output as the parasite burden increases. An additional feature typical of reports on density dependent fecundity is the presence of a greater degree of variation in fecundity at low worm burdens. However, the phenomenon is not associated with all Helminth infections, for example, several reports have indicated an absence of density dependent effects on the fecundity of Haemonchus contortus in sheep [10],[11], while the evidence for Ostertagia circumcincta is equivocal [12], and Shaw and Moss [13] found no evidence of a density dependent effect on fecundity in Trichostrongylus tenuis in red grouse.
In terms of human STHs, the density dependent effects have been reported for Ascaris lumbricoides [14]–[17], Necator americanus [18],[19], and mixed N. americanus and Ancylostoma duodenale infections [20]. Anderson and Schad [20] noted that while the intrinsic fecundities of the two human hookworm species were similar, the density dependent constraints on egg production were more severe for N. americanus. For Trichuria trichiura the evidence is less clear. Bundy et al. [21] reported a strong density dependent effect on fecundity, while a later study [22] found that no significant association between fecundity and worm burden. The later report did however note that very high per capita fecundities were only observed at low worm burdens. Michael and Bundy [23] reported density dependent effects on fecundity in Trichuris muris. Density dependent effects on egg output have also been reported for the canine hookworm A. caninum [24],[25].
It has been suggested that density dependent fecundity results from either competition for resources between parasites, or from immunological responses by the host [8],[9], or from direct parasite-parasite interactions [26]. In the case of Strongyloides ratti infections in rats, the density dependent effects on the parasite's establishment, survivorship, as well as fecundity, are mediated by host immune responses [27],[28].
A number of reports have described the impact of density dependent fecundity on transmission dynamics [20],[29], the rate of parasite reinfection following chemotherapy [30], and the spread of drug resistance [31]. We were interested in examining the effect that the phenomenon may have on the use of the FECRT to detect anthelmintic resistance in human STHs. If egg-laying by female hookworms is constrained by density dependent effects, the question arises as to whether this phenomenon is dynamic in nature if the worm density within a host is changed by, for example, a drug treatment. Survivors of a drug treatment will be present at a lower density after the treatment compared to before. Will the constraints on egg production be removed as the worm density decreases? The present study aimed to address this issue in terms of its potential to distort a FECRT by examining recent evidence on the response of the canine hookworm to drug treatment [7], and by examining an early N. americanus dataset describing the relationship between egg output and worm infection levels [18].
Kopp et al. [7] examined the response of canine hookworms to pyrantel drug treatment. The hookworms were from a highly resistant isolate, and an isolate showing a low level of resistance. Infections were established in four dogs (two per worm isolate), and faecal egg counts were measured daily. Egg counts stabilised by day 24, and all dogs were treated with pyrantel. Egg counts were monitored for a further 6 days, and faeces was collected daily. The dogs were then euthanased and all worms present in the small intestine were counted. Total pretreatment worm counts were calculated as the sum of those present in the animal after drug treatment and those recovered from the daily feacal collections. Egg output (epg) data over the time course of the experiment was reported. For the present study we reanalysed the data from Kopp et al. [7] in terms of egg output per day (epday) per worm, rather than epg. Daily faecal weights had not been recorded so in order to do estimate daily egg output per dog we used a mean value for faeces output of dogs of similar size to those used in the Kopp et al. [7] study that were housed under identical conditions for an earlier experiment. This represents an approximation of the faeces output for each of the 4 study dogs; however the absolute amount of output is not crucial to our analysis, as long as it remained approximately constant before and after drug treatment. Kopp (pers. comm.) noted that there was no observable change in faecal output in each dog post treatment.
We examined the data of Hill [18] who reported egg output (epg, and epday per female) and worm burden (counting worms in faeces after treatment with carbon tetrachloride and oil of chenpodium) in human cases infected with N. americanus in Porto Rico in 1926. Complete data was obtained for 93 cases. Hill [18] noted a threshold of 500 females above which egg production per female was always low. We grouped the human cases separately: those above the 500 threshold which we term as ‘constrained’ in their egg output by density effects (n = 11), and the remainder of the cases with less than 500 worms (n = 82). We examined the effects on a FECRT if these ‘constrained’ worms are able to increase their egg output upon relaxation of the density dependent constraints by a drug treatment. We assumed that as a result of the removal of significant proportions of the adult worm burden (60–90%), all female worms in the ‘constrained’ group would be able to increase their egg output to the median value of those worms lying below the 500 female threshold. This represents an estimation of the post drug egg output as the worms in the 11 cases will most likely adopt a range of egg outputs after drug treatment as reflected by the great degree of variability shown by worms at low densities [8],[9]. We assumed that all worms below the 500 threshold (n = 82) would not change their egg output per female after drug treatment, as the drug treatment did not represent the removal of fecundity constraints. We calculated the numbers of female worms remaining in all 93 human cases at each drug efficacy value, and then expressed their egg output on the basis of the faecal output of each human case prior to the drug treatment (mean of 2 or 3 faecal output values given for each human by Hill [18]). The % decrease in faecal egg count (% FECR) for each person was then calculated using the formula: [(T1-T2)/T1]×100, where: T1 = epg before treatment,T2 = epg after treatment. Mean % FECR was calculated for the constrained (n = 11) and unconstrained (n = 82) groups, as well as for the whole population (n = 93).
We also examined the relationship between epg and infection level for the Hill [18] data set. The human cases were subdivided into those showing epg values above or below 10,000, 5,000 and 2,500. The relationship between these groupings and actual infection levels (number of female worms per host) was then examined graphically. This allowed assessment of the impact of removing certain epg groups from the full data set on the ability to preferentially omit high infection level cases from a FECRT analysis.
Figure 1 shows an analysis of the egg output and adult worm numbers for the 4 dogs before and after treatment with pyrantel using the data of Kopp et al. [7]. Dogs 1 and 2 were infected with worms from an isolate with a high level of resistance to the drug, while dogs 3 and 4 were infected with an isolate which showed only a low level of resistance. All four dogs had similar worm burdens before the drug treatment, and all showed very similar rates of egg production per adult worm. Following drug treatment, all dogs showed an increase in their egg output per worm alongside reductions in worm numbers. The change in worm reproductive output was most dramatic for dogs 3 and 4 which increased egg output per worm by 4.2- and 3.6-fold, respectively.
The production of eggs per day per female worm from the study of Hill [18] is shown in Figure 2A. The two trends noted earlier as common to such egg output data are apparent, namely, 1) as the number of worms harboured increases there is a decrease in the number of eggs produced per female, with no cases of high egg production per female occurring at high worm densities (>500 females per person), and 2) at lower densities (<500 females per person) there is a large degree of variation in egg production between females. Egg production per day per female at densities less than 500 females varied from 351 to 11,904.
The result of our simulated FECRT on the human cases described in the Hill [18] data set is shown in Figure 2B. The actual drug efficacy was varied from 60–90%, while observed efficacy was calculated by assuming that the ‘constrained’ cases (those shown as squares in Figure 2A) could increase their egg output to the median of the unconstrained cases (indicated by crosses) due to the relaxation of density dependent constraints. It is apparent that a grouping of the 11 highest infection level human cases shows a % FECR significantly less than the actual removal of worms by the drug, and that this component of the population has affected the ability of the whole data set (93 cases) to accurately reflect drug efficacy.
Figure 3 highlights the cases from Hill [18] which fall into various epg groupings in order to demonstrate the effect of using such pre-treatment epg values to select appropriate candidates for a FECRT. It is apparent that the highest epg cases (epg >10,000, n = 10) do not coincide with the heaviest infection levels (Figure 3A). As the epg cut-off decreases, the highest infection level cases are mostly highlighted, alongside many cases with much lower infection levels (Figures 3B, 3C). For this data set, removal of most of the heaviest infections (ie. >500 females) from a FECRT would require the omission of all cases with epg >2,500, representing the additional omission of many low infection (<500 female) cases. It is clear that the relationship between egg output for each human case (ie, epg) and infection level (number of females) is poor due to the extent to which egg output per female can increase at low worm densities.
The study by Kopp et al. [7] suggests that the reproductive behaviour of adult hookworms before and after a drug treatment will likely be far from static. The significant increase in egg production by the worms following drug treatment is most likely a consequence of the relaxation of density dependent constraints on their egg output. The degree to which egg output increased coincided with the extent to which the worm populations were reduced by the drug treatment in the dogs harbouring the two worm isolates showing different levels of resistance to the drug. Given the widespread occurrence of density dependent worm fecundity constraints in human STHs [14]–[22] it is most likely that the phenomenon noted here with the canine hookworm will apply more widely to the human STHs.
The effect of the host immune response on fecundity in S. ratti has been demonstrated to be reversible [32],[33], with the immune mediated constraints on egg production being quickly reversed upon removal or suppression of the host immune effects. The time scale of the reversibility in these studies is similar to that in the report of Kopp et al. [7] (6 days), and as would occur in a FECRT (approximately 10 days between drug treatment and faecal egg count). However, the role of host immune mechanisms in the increased fecundity observed by Kopp et al. [7] is unclear as no specific suppression or removal of host immunity was imposed as in the experiments described above for S. ratti.
It is likely that among a group of people involved in a FECRT will be some cases in which female worms are producing more eggs post-treatment than they did prior to the drug treatment due to the relaxation of density constraints following the drug treatment. Therefore epg in these cases would be expected to decrease by a lower percentage than worm burden in response to the drug. Hence, a comparison of epg before and after drug treatment will underestimate efficacy. For dogs 3 and 4 from Kopp et al. [7] removal of 71% of the worm burden by the drug was associated with an increase in epg of 41%. The degree to which an increase in fecundity can distort FECRT data will depend on the extent to which the egg production of female worms is constrained prior to the FECRT. It is clear that while worms which are constrained in their egg output by density-dependent effects prior to the treatment will have the potential for increasing their egg output per worm as worm numbers decrease, cases whose worms are at low densities and hence may be producing eggs at nearer to their reproductive potential within that host prior to the drug, will have a significantly lower scope to further increase egg output. It can be seen from Figure 2B that the overall FECR data from a pool of cases will depend on the relative proportions of worms that are able to increase egg output significantly in response to drug treatment, compared to those which do show little or no change, and the extent of any output increases in individual cases. For FECRT data to be a useful tool for detecting changes in drug efficacy (that is, drug resistance), it is important for such population based variables to be minimised. If FECRTs are conducted on populations with similar balances of cases which show a presence or absence of relaxation of egg output constraints, then they will be comparable despite the % FECR values underestimating actual drug efficacy. However, if the degree of pre-treatment constraint is different between two populations, then the potential exists for % FECR values to falsely indicate different drug efficacy between the two populations.
Removal of the most heavily infected human cases whose worm populations are significantly constrained by density dependent effects from a FECRT study would seem desirable in order to reduce the potential distorting effects that relaxation of the constraints in these cases may have on the FECRT data. Anderson and Schad [20], working with mixed infections of N. americanus and A. duodenale, showed that, although density dependent fecundity resulted in a poor relationship between epg and worm burden, it was possible to discriminate between people with very low and very high infections on the basis of epg. This indicates that omission of high burden cases with potential to distort FECRTs would be possible. However, our analysis of the Hill [18] N. americanus data set in this regard shows that such selection of suitable candidates for a FECRT may not always be so clear. For the Hill [18] dataset, omission of the highest epg cases would not result in preferential removal of the highest worm burden cases (from Figure 3A). It is apparent that the increase in fecundity seen at lower worm burdens is of such a magnitude that the total egg output by these less dense populations (as measured by epg) can greatly exceed that of populations containing significantly higher numbers of worms. However, for the Hill [18] data set, as the epg cut-off for omission from a FECRT was further reduced, the removal of most heavy infections was evident (Figures 3B, 3C). A negative outcome is that many lighter infection cases were also removed. The cost, in terms of necessitating the omission of many apparently ‘unconstrained’ populations, may be outweighed by the benefit of removing the potential distorting effects of the heavy infection cases. Decisions on omission of cases for FECRTs may depend on local infection levels, and availability of suitable numbers of people to satisfy the statistical requirements of the test. Further study may identify epg cut-off values that could be applied widely to FECRTs.
The FECRT is used in livestock industries as a measure of drug resistance. The question arises as to whether the density dependent effects we have described will have a significant potential to distort FECRTs with the human parasites given the acceptance of the utility of the test for livestock. While reports on some livestock species have shown an absence of density dependent effects [10],[11], other livestock species show strong density dependence (for example, Teladorsagia circumcincta [34]). A difference in the application of FECRTs to livestock and humans, which may influence the impact of density dependency, is the efficacy levels of anthelmintics in the two systems. In livestock, the expected drug efficacy for susceptible parasites is greater than 99%. Resistance is suspected if this value drops below 95% (and if the lower 95% confidence interval is below 90%) [35]. Hence, very little of a susceptible worm population within each host is expected to survive the drug treatment, leaving little scope for egg output in the remaining worms to be significantly amplified by relaxation of density dependent constraints. Hence, a susceptible worm population is clearly indicated by the test. The effect of a relaxation of density dependent fecundity constraints in amplifying egg counts from worms remaining after drug treatment may only be expected if a significant proportion of the population is drug resistant. In humans, however, the situation is quite different. Efficacies for some human anthelmintics would never have originally met the criteria required by the livestock FECRT as indicating susceptibility, suggesting the use of suboptimal dosing regimes in some cases. While egg reduction rates for A. lumbricoides after treatment with albendazole or mebendazole are generally close to 100%, the values for hookworms and Trichuris are much lower [36],[37]. For example, in the analysis of a number of studies by Keiser and Utzinger [37], hookworm egg reduction rates in response to albendazole varied form 64.2% to 100%, with an overall cure rate of 78.4%. Hence, with significant proportions of susceptible worms expected to survive current treatment regimes, the capacity for the relaxation of fecundity constraints in drug susceptible worms to distort efficacy measurements may be significantly greater than would be expected for the livestock parasite species.
Given the potential for distortion of FECRT data by density dependent fecundity, it may be difficult to identify apparent drug efficacy changes (that is, the development of drug resistance) amongst a background of dynamic female worm reproductive biology. The impact of worm biology on the utility of the FECRT as a resistance detection tool highlights the need to remove this influence by developing methods which examine either direct drug effects on isolated worms with in vitro phenotypic assays, or changes in worm genotypes, as drug resistance monitoring tools.
|
10.1371/journal.pgen.1004503 | Recombination in the Human Pseudoautosomal Region PAR1 | The pseudoautosomal region (PAR) is a short region of homology between the mammalian X and Y chromosomes, which has undergone rapid evolution. A crossover in the PAR is essential for the proper disjunction of X and Y chromosomes in male meiosis, and PAR deletion results in male sterility. This leads the human PAR with the obligatory crossover, PAR1, to having an exceptionally high male crossover rate, which is 17-fold higher than the genome-wide average. However, the mechanism by which this obligatory crossover occurs remains unknown, as does the fine-scale positioning of crossovers across this region. Recent research in mice has suggested that crossovers in PAR may be mediated independently of the protein PRDM9, which localises virtually all crossovers in the autosomes. To investigate recombination in this region, we construct the most fine-scale genetic map containing directly observed crossovers to date using African-American pedigrees. We leverage recombination rates inferred from the breakdown of linkage disequilibrium in human populations and investigate the signatures of DNA evolution due to recombination. Further, we identify direct PRDM9 binding sites using ChIP-seq in human cells. Using these independent lines of evidence, we show that, in contrast with mouse, PRDM9 does localise peaks of recombination in the human PAR1. We find that recombination is a far more rapid and intense driver of sequence evolution in PAR1 than it is on the autosomes. We also show that PAR1 hotspot activities differ significantly among human populations. Finally, we find evidence that PAR1 hotspot positions have changed between human and chimpanzee, with no evidence of sharing among the hottest hotspots. We anticipate that the genetic maps built and validated in this work will aid research on this vital and fascinating region of the genome.
| Recombination is a fundamental biological process, which shuffles genes between pairs of chromosomes during the production of eggs and sperm. After shuffling, the chromosomes consist of alternating sequences of genes from each parent, where the switches are the result of ‘crossovers’. Recombination is essential for eggs and sperm to receive the correct number of chromosomes, failure in which is an important cause of miscarriage, birth defects and mental retardation. Males have the particular challenge of recombining between the X and Y chromosomes. Unlike the other 22 chromosome pairs, the X and Y chromosomes do not match up, except for a small special region called PAR1, which must host a crossover. We investigate recombination in PAR1 by building a ‘map’ of where it occurs in African-American families. We use a variety of approaches, both analytical and experimental, to demonstrate the role of a protein called PRDM9 in marking crossovers in this region. PRDM9 has previously been shown to position crossovers on the other chromosomes, but a role in PAR1 was unexpected based on research in mice. We also show that the recombination map has changed in the evolutionary history of PAR1, both among human populations, and between human and chimpanzee.
| Pseudoautosomal regions (PARs) are segments of sequence homology between the X and Y (or Z and W) chromosomes, which are otherwise non-homologous. Uniquely, PARs are inherited in the same manner as autosomes, while also being partially linked with X-specific and Y-specific loci. They have a critical role in the successful progression of meiosis in mammalian males and in the heterogametic sex in many other plant and animal species [1]–[10]. Correct segregation of chromosomes into gametes during meiosis requires that homologous chromosomes pair up and undergo exchange of chromosomal material known as recombination or ‘crossing over’. In females, the two homologous X chromosomes pair up and can recombine along their entire length [3]. In males, however, pairing and recombination are restricted to the homologous PAR regions. PARs in most mammals are typically a few hundred kilobases to several megabases in length [11]–[14] and make up only a small fraction of the Y chromosome, imposing an extraordinary pressure to achieve recombination in a short genomic segment. Humans have two PARs – PAR1, which is at the tip of the short arm (Xp/Yp) of the sex chromosomes, and PAR2, which is at the tip of the long arm (Xq/Yq). Deletion of PAR1 is associated with total male sterility in humans [5], [15]. Reduced recombination in PAR1 can lead to aneuploid sperm, which can cause X-chromosome monosomy (Turner syndrome) or XXY (Kleinfelter syndrome) in the offspring [7], [16].
In addition to their vital role in fertility, PARs contain genes in all mammals whose sequence has become available so far. The human PARs together contain at least 29 genes, with diverse roles in cell signalling, transcriptional regulation and mitochondrial function [17]. Thus far, SHOX is the only PAR gene which has been definitively associated with a role in normal development [18]. More recently, associations have also been reported with PAR1 loci for schizophrenia and bipolar affective disorder [19], [20].
Studies in viable human sperm and pedigrees have shown that the recombination rate in PAR1 is consistent with one obligatory crossover per male meiosis, accompanied very rarely by a second crossover [2], [21]. PAR1 is approximately 2.7 Mb long, and this leads to PAR1 having a crossover rate 17-fold greater than the genome-wide average, over four times greater than the next most recombinogenic region of comparable size in the genome. In contrast, the female recombination rate in PAR1 is comparable to the genome-wide average [22]–[24]. Human PAR1 shares homology with other mammalian PARs [14], [25]. While PARs in several mammals, including human, horse, cattle, dog and sheep, appear to descend from the same ancestral region [25], the boundary between the PAR and X-specific and Y-specific regions has shifted dramatically, leading to highly variable gene content. The mouse PAR does not share homology with human or any other known mammalian PAR (the ancestral PAR appears to have been lost from the mouse X chromosome). Instead, mice have a different, considerably shorter PAR on the q-arm of the X chromosome, which spans only 700 kb [26], [27].
The second human pseudoautosomal region, PAR2, is much smaller at approximately 330 kb and specific to the human lineage, having likely arisen due to a translocation between the X and Y chromosomes [28]. Crossovers in PAR2 occur rarely, at a rate similar to the genome average, in both sexes [24], suggesting behaviour similar to many autosomal regions. For the rest of this work, we focus our attention on PAR1, the evolutionarily and biologically more significant region.
Despite the critical role of PAR1 in fertility and disease, an understanding of its biology remains highly incomplete. In the reference human genome, the PAR1 sequence is not yet fully assembled, likely because of the exceptionally high GC-content and density of repetitive regions it contains. Since the publication of the X chromosome sequence [17], updates in the human genome release GRCh37 by the Genome Reference Consortium have closed some of the gaps, resulting in a sequence that is complete. Nevertheless, PAR1 has a far lower density of single nucleotide polymorphisms that are included on genotyping arrays relative to other parts of the genome [29], despite the much shorter extent of linkage disequilibrium (LD) in this region. PAR1 has also largely been neglected in linkage studies and genome-wide association scans, possibly due to the lack of both polymorphism and linkage information. For other mammalian species with otherwise high-quality reference genomes, the PAR sequence is similarly either absent entirely or only partially represented [30].
Even less is known about recombination, which lies at the heart of PAR1 biology. For instance, it is not known how the extraordinarily high rate of recombination in this region is achieved biologically. In the autosomes, recombination clusters into short 1–2 kb segments known as ‘recombination hotspots’, which are flanked by regions with very low recombination rate [31]–[35]. That hotspots are also a feature of PAR1 recombination is implied by the characterisation of a single recombination hotspot within the SHOX gene, which is one of the hottest hotspots measured thus far using high resolution sperm-typing in the genome [36]. However, no further hotspots in PAR1 have yet been characterized. The utility of the fine-scale genetic map based on LD [29] in this region is unclear [37], due to the very rapid breakdown of LD in this region [36]. Other currently available genetic maps for PAR1 that have been built using low resolution sperm-typing and genotyped pedigrees are based on a small number of markers, typically in small sample sizes [2], [21]–[24], [38], [39]. This, along with technical difficulties linked to the relatively small size of PAR1, leads to imprecise estimates, and insufficient resolution to understand the drivers of recombination. The most fine-scale map available to date from directly observed crossovers was built in 28 European ancestry pedigrees genotyped at 22 polymorphic markers in PAR1, corresponding to roughly one marker per 100 kb [24]. The most detailed human pedigree-based map built to date [40], with 15,000 meioses in the Icelandic population, did not include any markers in PAR1. The PAR was also not included in the recent work that built LD-based maps in the chimpanzee [41].
An intriguing study [42] found that pairing of homologous chromosomes occurs significantly later in the PAR than in the autosomes in male mice. They also found that chromosomal axes were significantly longer in the PAR relative to the autosomes during meiosis, and that a different isoform of a key recombination protein (Spo11) is active in this region, implying that distinct recombination machinery may operate here.
The role of another key recombination protein, PRDM9, is also unclear in the PAR. Several lines of evidence have shown recently that PRDM9 positions sites of recombination in human and mice autosomes [43]–[45] by direct binding to recombination hotspots. However, whether PRDM9 plays any role in the male PAR1 is controversial. Recent work in mice [46] has shown that male mice with different Prdm9 variants have completely different autosomal recombination patterns, yet show similar recombination landscapes in and adjacent to the PAR region. Brick et al. [46] have therefore suggested that a mechanism independent of Prdm9 may be positioning crossovers in the mouse PAR.
In this work, we aim to characterise the patterns of recombination in PAR1 to learn more about the biology of this region, and provide a resource for medical genetics research. We have built the most fine-scale genetic map containing directly identified crossovers to date in this region. This map contains more meioses, and an order of magnitude greater markers than the densest PAR1 map so far [24]. This allows us to analyse recombination in this region at a finer scale than has been possible in the past. It also enables us to assess the accuracy of the LD-based map built using HapMap2 variation data in this region [29]. We use evidence of direct PRDM9 binding in human cells to examine the role of this protein in specifying recombination in PAR1. Finally, we measure the impact and evolution of recombination using observed biases in the allele frequency spectra for different types of mutations due to recombination. We leverage these resources to explore the role of PRDM9, and to infer evolution of recombination in PAR1 within human populations and between human and chimpanzee.
We have leveraged the genotype data of 220 markers from 135 African-American families with two or more children to build a new pedigree-based genetic map (Materials and Methods, Text S1, Dataset S1). These data comprise a total of 672 meioses (336 paternal and 336 maternal), in which we could directly detect crossovers between parent and child. Amongst these families, 19 families included genotype data for both parents, and the rest for only one parent. We used methods that we have previously published [47] to detect crossovers in such incomplete pedigrees (Materials and Methods).
Figure 1 shows the recombination rates estimated in both males and females (Dataset S2). We inferred a total genetic distance of 136 paternal and 18 maternal crossovers in PAR1. The average number of detected events in males (0.4 events per meiosis) is less than the expected number of events (0.5 events per meiosis). This may be due to the paucity of markers in the sub-telomeric 250 kb region of PAR1, which reduces our power to detect crossovers in this region. The number of female events (0.05 events per meiosis) is consistent with previous studies, which have detected between 0.03 to 0.06 events per female meiosis [2],[23],[24],[39]. 126 paternal and 17 maternal crossovers have both endpoints mapping within our region of marker coverage (Datasets S3 and S4). No double crossovers were identified in either sex. Table S1 summarizes the resolution of paternal and maternal events.
We found intense crossover activity throughout PAR1 in males. Only a few loci have an estimated recombination rate that is lower than the genome-wide average rate of approximately 1.2 cM/Mb [47], with little evidence for truly cold regions anywhere in the male PAR1. The previously identified SHOX hotspot [36] is at a peak of male recombination rate (Figure 1). Consistent with the pattern in other chromosomes in males [48], [49], we observed a significant trend of reduction in rate away from the telomere (Tables S2 and S3). In contrast, in females, we observed the lowest rate near the telomeres and the highest rate near the pseudoautosomal boundary, and the differences are significant (Tables S2 and S4). The male rate increases somewhat in the vicinity of the pseudoautosomal boundary (Figure S1).
In the rest of this work, we use these maps to validate the sex-averaged HapMap2 LD-based map, and to learn about the biological drivers of recombination in this region.
The HapMap2 LD-based map is the most fine-scaled map currently available for PAR1 with rates inferred between nearly 1,400 markers [29]. This map was built using genotypes from unrelated individuals from three HapMap Phase II populations – European ancestry individuals from Utah (CEU), Yoruba individuals from West Africa (YRI) and Asian individuals from China and Japan (JPT+CHB). Maps specific to each of these populations have also been built, and are referred to as the CEU, YRI and JPT+CHB maps respectively. LD-based maps are built by inferring recombination from the observed breakdown of linkage disequilibrium between markers, and capture information from tens of thousands of meioses over thousands of generations of human history. They have been found to be reliable estimates of historical recombination rates in the autosomes, in comparisons with numerous pedigree-based maps and high-resolution sperm-typing experiments [40], [50].
In PAR1, however, the use of LD-based maps raises special concerns specific to this region. The first concern is that rate estimates in the map may be biased downwards, which we call ‘saturation’ of rates. This is because recombination is inferred from the breakdown of LD between markers. If the recombination rate is very high, nearby markers may segregate practically independently. Since further recombination cannot meaningfully reduce the LD in this situation, it may not be possible to infer any difference between very high rates, in practice. The second concern is that the role of selection in PAR1, to ensure male fertility, is unknown, and strong selection might bias the estimation of rates. Therefore, it is vital to empirically confirm the map using a resource which is not influenced by these factors. Finally, LD-based maps are sex-averaged. Since male recombination in PAR1 is of particular interest, we also assess how informative this map is for male recombination.
To check the accuracy of the HapMap2 population-averaged LD-based map, we compared it with the sex-averaged rates from our pedigree map, and found good agreement between the two maps (Figure 2a). The correlation between the maps is high despite considerable statistical uncertainty in the estimation of the pedigree-based map (Spearman's at 50 kb scale, ). Further, there is no evidence of downward bias among high rate regions in the LD-based map (Figure 2a). This suggests that saturation of rates is not a significant concern.
Approximately 90% of the historical crossover events in PAR1, which influence LD patterns in the region, are expected to have occurred in males. Therefore, we anticipate that the LD-based maps are dominated by male recombination. This is confirmed by the correlation of the male-specific pedigree-based map with the population-averaged LD-based map (Spearman's at 50 kb, ), which is approximately the same as that of the sex-averaged map.
Next, we assessed how accurately hotspots in the HapMap2 population-averaged LD-based map are localised by comparing them with the location of crossovers in the pedigrees. Specifically, we calculated the average rate around the centres of the best-resolved 10% of crossovers in pedigree fathers, whose resolution ranged from 13 kb to 45 kb. We found that the LD-based map has a clear peak precisely centred at the sites of crossovers (Figure 2b). This rate elevation to 14.6 cM/Mb above the average rate of 9.1 cM/Mb is significant (, 5000 bootstrap iterations over the crossovers). We conclude that the LD-based map predicts rate peaks at crossover sites in African-American fathers.
Recombination in African Americans has previously been modelled using a linear combination of the CEU and YRI maps in the autosomes [47], [51]. The ratio of the two maps (79%:21%) for the best linear combination of the two maps was similar to the average underlying ancestry proportions (80%:20%) in the admixed individuals [51]. We applied the same approach to the PAR1 map of our African-African fathers. If the CEU, YRI and the pedigree-based maps in PAR1 are the same, we would expect the best linear combination to be an equal 0.5∶0.5 weighting of the CEU and YRI maps, while differences between the maps should result in a higher YRI contribution. We found that, at the 10 kb scale, the best map is a weighted average of 70% (s.e. = 8%) YRI map and 30% (s.e. = 8%) CEU map. It is significantly different from an equal weighting of the two maps (). We also performed a model-free analysis by bootstrapping over the pedigree fathers, and calculating the mean squared difference of each bootstrap map with the CEU and YRI maps. We found that the YRI map is significantly more similar to the pedigree map than the CEU map (). This indicates that the LD-based approach has power to detect differences in the populations, and also suggests that the two populations have systematic differences in the first place. Although this analysis is suggestive, departure from the assumption of equal error in the CEU and YRI maps may also explain the results, in particular if the CEU map is less informative than the YRI map. However, other forms of evidence also support a population difference, but do not support lower error in the YRI map, as shown below.
These analyses show that the LD-based approach is reliable, accurate, and informative specifically about male recombination. This allows us to use both the pedigree-based and the LD-based maps in the rest of this work.
Recent work has shown that the chromatin-modifying protein PRDM9 positions the sites of practically all recombination hotspots in human and mouse autosomes [43]–[45]. PRDM9 contains a domain of C2H2 zinc fingers, which is remarkable for being the fastest evolving zinc finger domain in the genome [52]. There are, for example, no PRDM9 zinc fingers known to be present in more than one of the great ape species [44], and dozens of different zinc finger arrays have been characterized in humans [53]. Changes in the PRDM9 zinc-finger array are accompanied by shifts in the recombination landscape: multiple groups have shown that nearly all autosomal recombination is controlled by PRDM9 [46], [47].
A previous study [54] analysed over 30,000 LD-based hotspots and identified a 13-bp motif CCnCCnTnnCCnC (where ‘n’ may be any of the four bases) that marks approximately 40% of human hotspots. In the autosomes, only a fraction of the instances of this motif become hotspots [54]. More recently, the role of this motif has been understood through the realization that certain alleles of PRDM9, including the most common human allele, called allele A, bind this motif via the PRDM9 zinc finger array [43]. It has been shown that individuals with PRDM9 alleles binding to significantly different motifs have no shared autosomal hotspots [46], [47]. However, as discussed above, recent research suggests that Prdm9 may not have a role in specifying recombination in the PAR in mice [46].
To investigate whether PRDM9 is activating recombination in the human PAR1, we examined the recombination rate near exact matches to the motif CCnCCnTnnCCnC. We observed a sharp increase in the rate in the HapMap2 population-averaged LD-based map in the immediate vicinity of the motif (Figure 3a), comparable in magnitude to the increase observed previously in the autosomes [44]. In the autosomes, the likelihood of the motif resulting in a hotspot is several times greater in THE1A/B and L2 repeat elements, relative to other occurrences of the motif. While there are no copies of the motif within THE1A/B elements currently assembled in PAR1, there are 4 copies of L2 elements that contain the motif and around which rates could be measured. The recombination rate around these elements is nearly twice the regional rate (Figure S2), and the rate elevation is over 5 times greater as compared with other occurrences of the motif in PAR1. This weakly supports a greater increase in rate in such elements, consistent with the autosomes. Moreover, because PRDM9 binds the motif, the observation of a highly localized crossover rate increase around the motif conclusively demonstrates a role for this protein in PAR1.
While the bioinformatically predicted and inferred motif CCnCCnTnnCCnC narrows down the scope of PRDM9 binding sites in the genome, the relationship between motifs, binding sites and recombination hotspots is not perfect [53]–[55]. For example, zinc-finger proteins can bind DNA in a large variety of possible configurations, which are not fully understood [53], [56]. As a result, DNA sequences that appear unlikely to be bound in silico have been shown to bind in vitro [57].
To address this for PAR1, we measured PRDM9 binding experimentally via chromatin immunoprecipitation followed by high-throughput sequencing (ChIP-seq) in human cells (Materials and Methods). Specifically, we measured the binding of PRDM9 allele B, which is the human reference allele, and is predicted to have binding properties similar to PRDM9 allele A [43]. We identified 185 PRDM9 binding peaks in PAR1 (Materials and Methods). The LD-based map shows a sharp increase in rates at these peaks (Figure 3b), directly connecting PRDM9 binding with local recombination rate increases in this region. Notably, the rate elevation is more than two-fold the increase observed for the 13-bp motif alone (Figure 3a). This is consistent with the fact that the PRDM9 binding peaks constitute direct evidence of binding. Further, PAR1 peaks containing close matches to the motif are more strongly signalled and show a stronger increase in the LD-based rate than peaks without the motif (Figure S3), suggesting that strength of PRDM9 binding is correlated with recombination rate.
Finally, we report an intriguing characteristic of the binding peaks in PAR1. Approximately 42% of PAR1 peaks contain close matches to the motif, which is consistent with the expected number of hotspots containing the motif in the autosomes [54]. Nearly a fifth of the peaks contain 5 or more and 5% of the peaks contain 12 or more copies of the motif. Many of these peaks are composed of low complexity minisatellite-like tandem repeat structures of periodicity varying from 4 bases to 101 bases. Other tandem repeats containing matches to the PRDM9 binding motif have been observed to be unstable and biased towards gain of repeat units in the human male germline [58]–[60], and this might present an interesting counterbalancing mechanism to the loss of motifs due to preferred transmission of recombination-suppression alleles.
The PRDM9 zinc finger array is highly variable in humans, with around 40 different alleles that have been identified so far [53], [61]. Alleles can be grouped into 5 categories, depending on the number of bases at which their bioinformatically predicted binding sequence matches the 13-bp motif CCnCCnTnnCCnC (known alleles match between 4 and 8 out of the 8 non-degenerate bases in the motif). These categories have differing allele frequencies across different human populations [53]. Variants predicted to match the 13-bp motif exactly (8/8 match) are predominant in Europeans (91%) and Asians (also approximately 91%), but occurred at only about 58% frequency in an African sample [53]. In Africans, approximately 35% of PRDM9 alleles (5/8 match) are strongly predicted not to bind the 13-bp motif [47], [53]. This leads to Africans having reduced activity, on average, in the hotspots activated by alleles most common in Europeans. Instead, they are recombinationally active at novel hotspots not active in most Europeans [47], [53].
As shown in a previous section, African-American pedigree fathers have a significantly greater usage of the African (YRI) map than the European (CEU) map (P = 0.009). This suggests that recombination has evolved within the human lineage in PAR1, in a manner similar to the evolution observed in the autosomes.
To test this further, we examined rates across PAR1 in three population-specific maps, the European (CEU), African (YRI), East Asian (JPT+CHB) LD-based maps at the ChIP-seq binding sites of allele B, which is predicted to bind the 13-bp motif. As expected, the increase in rate in both the Asian and European maps near the binding sites is greater than that in the African map (P = 0.002 and 0.02 respectively) (Figure 4). This suggests that the CEU map is unlikely to be systematically less informative than the YRI map. As expected from the similar allele frequencies of the variants matching the 13-bp motif in Europe and Asia, there is no significant difference between the increase in rate in the European and Asian maps near B-allele binding sites.
Programmed double-strand breaks leading to recombination may be resolved in one of two ways, as crossovers, which involve reciprocal exchange of chromosomal material, or as non-crossovers, which do not [62], [63]. Both of these outcomes are accompanied by the non-reciprocal copying of a tract of DNA from one participating chromosome to another, known as gene conversion [63]. This process is said to be biased if one of the two chromosomes is systematically more likely to be used as the template for copying than the other chromosome, and this phenomenon is referred to as biased gene conversion (BGC). Several types of bias have been observed in different eukaryotes [64]–[68], among which is a bias favouring GC over AT alleles, referred to as GC-biased gene conversion (gcBGC) [66], [67], [69], [70]. gcBGC tends to increase the frequency of GC bases in the pool of gametes relative to 50%:50% Mendelian segregation.
A broad range of evidence, across several eukaryotic taxa, indicates that bias towards GC bases is associated with recombination[41], [66], [67], [69]–[74]. A detailed study of gene conversion tracts in yeast directly demonstrated the over-transmission of GC alleles [66], and a recent re-analysis of the data indicates that the bias may be specific to recombination events that are resolved as crossovers [70]. Patterns of variation both within and between species have shown a skew towards GC alleles that correlates strongly with recombination rates in primates, and particularly with recombination hotspots [41], [73]–[76]. The mouse gene Fxy presents a particularly striking case study, indicating that GC-bias may operate in the mouse PAR as well. This gene has translocated from the non-recombining part of the mouse Y-chromosome to its PAR within the last 3 million years [77]. This translocation has been followed by an extremely rapid increase in GC content at both coding and non-coding sites [69], [77]. While the molecular mechanisms causing gcBGC are not well understood, recombination is the only known force producing this bias [67], [70].
We investigated whether such a bias is observed in the human PAR1, both in the frequency of segregating sites and for the fixation of alleles leading to substitutions between human and chimpanzee. We reasoned that such a bias, if present, should act as an indirect marker of sites undergoing recombination in the two species, even in the absence of direct evidence on recombination sites in PAR1 in the chimpanzee. We investigated these patterns in (relatively) hot and cold regions of PAR1, and around copies of the 13-bp motif CCnCCnTnnCCnC, which marks peaks of recombination in PAR1 as shown above. Finally, we compared the distribution of GC-altering substitutions between human and chimpanzee to understand the evolution of recombination hotspots between the two species.
PAR1 in humans has a far higher GC content than the rest of the X chromosome (48% vs 39%) [18]. This is also true in chimpanzee (Pan troglodytes), which again has 48% GC content in the PAR. We used 1000 Genomes data [78] in PAR1 to obtain a set of sites segregating in human populations at a minor allele frequency of at least . We restricted the set to those sites where the chimpanzee allele is known, and assigned the chimpanzee allele to be the ancestral allelic state. Further, we filtered out all sites where either the ancestral or derived allele is part of a CpG dinucleotide to reduce noise due to repeat mutations resulting from the deamination of methylated CpGs.
Figure 5a shows the allele frequency distribution of all six classes of segregating sites in PAR1: GCAT transitions and transversions (which reduce GC content), ATGC transitions and transversions (which increase GC content), and AT and CG transversions (which leave GC content unchanged). We observed that mutations that increase GC content are enriched at the top-end of the frequency spectrum, while mutations that decrease GC content are more concentrated at the bottom end of the frequency spectrum. Specifically, we noted that a significantly greater proportion of mutations that increase GC content segregate with allele frequency than GC-reducing and GC-neutral mutations (). Correspondingly, GC-increasing mutations are less likely to segregate with allele frequency than GC-neutral mutations (), while the opposite is true of GC-decreasing mutations (). Among GC-increasing (or GC-decreasing) mutations, no significant difference was observed between transitions and transversions at any allele frequency. This is consistent with the expectation of gcBGC in the autosomes, however the ‘U-shape’ of the distribution is much more pronounced in PAR1 than in Chr 20, which is the autosome with the highest chromosome-wide recombination rate in the human genome [47] (Figure S4).
Figure 5b shows a comparison of the full allele frequency spectra of GCAT and ATGC mutations in the form of a quantile-quantile plot (details in figure legend). ATGC mutations in PAR1 segregate at significantly higher allele frequencies, on average, than GCAT mutations (). We compared this with the pattern in Chr 20. The hottest 15% of loci of size 1 kb in Chr 20 have an average rate of 8.2 cM/Mb, which is comparable to the sex-averaged rate in PAR1. ATGC mutations segregate at higher frequencies than GCAT mutations at these loci, to an extent similar to PAR1 (Figure 5b). This suggests that the mechanism causing the bias towards GC alleles operates similarly in PAR1 as it does in the autosomes, and that the strength of gcBGC may be similar in males and females. The coldest 15% of Chr 20, with an average rate of 0.02 cM/Mb, does not show a significant excess of GC-mutations, confirming that recombination is causing the bias towards GC-mutations. We note that a quantitative relationship between recombination rate and gcBGC is also confirmed in PAR1, where we observe that the more telomeric 200–700 kb of the PAR has a significantly stronger gcBGC effect than the 500 kb nearest the pseudoautosomal boundary (Figure S5), consistent with its higher average recombination rate.
We examined the role of PRDM9 by examining the allele frequency distributions of GCAT and ATGC mutations near the motif CCnCCnTnnCCnC. A prediction of the recombination-driven gcBGC hypothesis is that the effect should be strongest near recombination hotspots. As shown in Figure 6a, we compared the allele frequency spectrum of ATGC mutations near the motif relative to that class of mutations in PAR1 as a whole. We observed that the elevation of the allele frequencies of GC mutations near the motif is extreme, and far stronger, over and above the rest of PAR1 (which already shows a strong GC bias). The signal is local to the motif, and weakens rapidly with distance away from it (It is significantly stronger within 25 bases of copies of the motif relative to PAR as a whole (P = 0.008), and also relative to within 500 bases of copies of the motif (P = 0.01)). The lowering of allele frequencies of AT mutations is also extremely strong near the motif relative to the rest of PAR1. The effect is strongest within 25 bp of the motif, and weakens with distance from it (P = 0.02 relative to PAR as a whole).
We expect that, due to the much higher male recombination rate in PAR1, the GC-bias in PAR1 is driven mainly by male recombination. We confirmed this by comparing two regions with opposite trends in male and female recombination rates (Figure S5). Therefore, the patterns of GC-bias near the motif and throughout PAR1 cannot be explained by female recombination alone.
Brick et al. [46] have proposed that, in the mouse PAR, there is a cline of PRDM9 activity – with no activity in the most telomeric region and increasing activity with distance away from the telomere. We found no evidence for such a trend in humans. In the human PAR1, the elevation of GC allele frequencies and suppression of AT allele frequencies near the PRDM9 motif are at least as strong in the most telomeric region of PAR1 where rates could be estimated (200 kb–700 kb), as it is near the pseudoautosomal boundary (Figure S6). This region excludes the most telomeric 200 kb, where rates could not be reliably estimated due to lack of markers.
We examined whether gcBGC has an effect on substitution rates in PAR1. Figure 5a suggests that a segregating GC variant in PAR1 is about 1.9 times more likely to be near fixation as a segregating AT variant. To estimate bias in the overall rate of fixation of ATGC and GCAT variants while accounting for differences in mutation rates [79], we count segregating sites of each type using only derived alleles with allele frequencies between 95% and 100%. We found that, for Chr 20, the higher rate of being near fixation of individual GC alleles is offset by the greater number of GCAT segregating sites (bias estimate = 0.97). However, in PAR1, the number of GC bases near fixation exceeds that of AT bases by almost 20% (bias estimate = 1.19, P = 0.05). We note that this estimate is conservative since a subset of variants will have the wrong ancestral allele assigned due to polymorphism or errors in the chimpanzee (assuming that ATGC and GCAT mutations are equally likely to have the wrong ancestral allele).
Within 25 bases of the 13-bp motif CCnCCnTnnCCnC (Figure 6b), the fixation bias towards GC is extremely high – 8 times as many GC bases are near fixation as AT bases (bias estimate = 8.0 and P = 0.003, and compared with Chr 20 bias estimate = 1.15). Another way to estimate the fixation bias close to the motif, in a conservative way, is to model the allele frequency distribution of derived GC alleles as a mixture of the PAR-wide allele frequency distribution of GC alleles, and a perfectly symmetric U-shaped distribution representing a situation where derived alleles are either newly arisen or completely fixed. Such an analysis indicates that 28.4% of motifs in the PAR are extremely active. This contrasts with an estimated 3% of motifs in Chr 20, which is consistent with previous autosomal estimates [54]. This suggests that the higher recombination rate in PAR1 may be supported by nearly an order of magnitude greater availability of motifs for binding via PRDM9.
In the section above, we showed that recombination in PAR1 strongly accelerates the fixation of ATGC mutations relative to GCAT mutations. While the overall GC content is similar in the PAR in human and chimpanzee, we ask if the location of substitutions differs in the two species. A region that is a hotspot in one species but not in another is likely to accumulate more GC-substitutions in the first species. In other words, if two species are significantly different in their hotspot landscape, we would expect to see a corresponding signature in the location of their respective GC substitutions.
We test this hypothesis by comparing human and chimpanzee PAR sequence. While no fine-scale genetic map is available for the chimpanzee PAR, we compare substitutions in the two species in regions which are hotspots in humans. Specifically, we consider substitutions in syntenic regions using a human-chimpanzee sequence alignment (Materials and Methods). If hotspots are the same in both species, we expect to see comparable numbers of and substitutions in regions identified as human hotspots. If the hotspots are completely different, we expect to see an excess of over substitutions in human hotspots. Determining which species experienced the mutation, however, requires the DNA sequence of a related species as outgroup. For PAR1, however, the sequence assembly is less than 4% complete for any primate other than human and chimpanzee. Therefore, while the inability to determine the direction of the mutation reduces power to detect differences, we would still expect to observe an excess of over substitutions in human hotspots (if they are different from those in chimpanzee).
To quantify the relationship between substitution and recombination rate, we modelled substitution rates using a linear model with recombination rate, GC content and CpG content as explanatory variables. We performed this analysis in 2 kb intervals, the approximate size of a hotspot [80], using the HapMap2 LD-based map [29]. We considered all six mutational possibilities separately: the two types of transition ( and ) and four types of transversion (, , , and ). Substitution rates between the different mutational types are highly correlated with each other, and may reflect systematic differences between loci, such as variable mutation rate and chromatin context, some of which may also influence recombination rate [35], [81], [82]. To control for such systematic differences in mutation rates between loci, we modelled the substitution rate in each mutational class as the dependent variable, and included the substitution rate in all other mutational classes as explanatory variables (together with human recombination rate, GC-content and CpG content). This approach is likely to be conservative, if recombination influences both transitions and transversions towards GC bases.
Table 1 summarizes the effect size and p-value of the human recombination rate explanatory variable for each mutational class in unique DNA. Human recombination rate correlates with the rate of transitions, independently of the other factors we considered. This is consistent with previous studies [72], [73], and is expected based on our results above for sites segregating in human populations. Specifically, these results suggest that recombination is a driver of fixed substitutions towards GC in the PAR, even measured over millions of years, a result observed previously for the autosomes [41], [74]. A significant effect of transversions was not observed. This may be because there are 2.6 fold fewer transversions, leading to lower power to detect true associations. It may also be because allowing transitions as an explanatory variation in the regression reduces our power further.
However, while human recombination rate is strongly correlated with GC-biased transitions in humans, there is no evidence that it is correlated with GC-biased transitions in chimpanzee (Table 1) in the same way, because recombination does not show a symmetric association with transitions. Since our results above establish that human recombination hotspots in the PAR are associated with elevation of GC substitution rates, if these sites were also hotspots in chimpanzee, we would expect to see a similar signal in that species also. Because we do not, we deduce that recombination patterns have changed strongly in the PAR between humans and chimpanzee. To investigate this further, we estimated the increase in the rate of GC-biased transitions in each species in the hottest and coldest 15% of human loci in the PAR, relative to regions with intermediate rates.
Figure 7 shows that the hottest human regions have significantly greater accumulation of GC-biased transitions than the coldest regions (), which is not the case for the chimpanzee (). The coldest human regions have a comparably reduced rate of GC-biased transitions in both humans (−0.05% per base) and chimpanzees (−0.09% per base), suggesting that the coldest regions may be shared between the two species. This is consistent with previous work in the autosomes [41], [83], showing that certain regions (e.g. genic regions) show reduced recombination rate in both human and chimpanzee but that no shared hotspots exist. Finally, human hotspots show significantly greater rate of GC-biased transitions in human than in chimpanzee (Figure 7, ). In fact, in agreement with the idea of no chimpanzee hotspot activity at human hotspots, the hottest human regions have no increase in GC-biased transitions in the chimpanzee (estimated excess in chimpanzee is −0.01% per base, relative to +0.21% per base in human). This observation that hotspots are almost certainly different in PAR1 between humans and chimpanzees is consistent with our finding that PRDM9 positions hotspots in this region.
Finally, we investigated whether hotspot heat can be predicted using the observed substitution patterns. Current approaches, such as the building of LD-based maps, require multiple individuals from a species to be genotyped or sequenced. Since such data are currently not available for the PAR in most organisms, an ability to build recombination maps using only the reference sequence of closely related species could provide a preliminary method to analyse recombination. We found that the ‘optimal’ linear model using the human-chimpanzee divergence patterns (Materials and Methods) explains 23% of the variance in the LD-based map (Table S5). While the variance explained may seem low at first, it is, in fact, in line with expectations. This is because LD-based maps capture recombination in the last thousands of generations [84] while the rate predicted from substitution patterns averages recombination since the human-chimpanzee split. If hotspots are turning over at the same rate in the PAR as in the autosomes, they are being replaced every 1 to 2 million years [54]. Given a human-chimpanzee speciation time between 5.5 and 7 million years ago [85], the LD-based maps are expected to comprise only about a third to a seventh of the recombination reflected in the substitution-based approach. We found that diversity data can also be used to estimate a genetic map, albeit at a broader scale (Figure S7).
In this work, we have built the most fine-scale genetic map to date from directly inferred crossovers for the human PAR1. We used this map to validate, for the first time, the previously built LD-based genetic map in this region, which localises recombination to a resolution close to the size of a hotspot. We also show the existence of biological differences between LD-based maps in different populations. We hope that these resources will promote research in this gene-rich and fast-evolving region, which currently remains under-represented in both linkage studies and on genotyping chips used in large-scale disease association scans.
Our analysis indicates that, in contrast with evidence currently available for the mouse [46], PRDM9 indeed plays a powerful role in positioning recombination events in the human PAR1. PRDM9 binding sites, and target motifs, mark crossover hotspots. In turn, these hotspots are sites of very rapid – much more rapid than on the autosomes – evolution of base content towards becoming more GC rich. Thus, as has been seen in other species [77], recombination is a rapid and powerful driver of sequence evolution in the PAR. Moreover, by using GC change as a marker of recombination sites, we observe indirectly that chimpanzee hotspots and human hotspots must show little or no overlap in PAR1, without being able to directly identify such hotspot positions in chimpanzee. This signal cannot be due to recombination only in female meiosis, because our PAR1 maps are dominated (90%) by male recombination. Moreover, the exceptionally rapid sequence evolution we see in PAR1 implies evolution driven by male meiosis, because recombination in female meioses does not occur at an unusually high rate in this region.
In many ways, PAR1 has a recombination profile in male meiosis resembling a miniature autosome, with an elevated crossover rate near the telomere. However, we observe a key difference in that a relatively high rate appears maintained throughout most of the region, without recombination coldspots as seen in the autosomes. A clue to what might be going on is perhaps given by the examination of mutations near the positions of the 13-bp motif CCnCCnTnnCCnC in PAR1, which revealed extreme skews in frequency spectra with almost no high frequency mutations toward AT bases and a U-shaped distribution of mutations towards GC bases, particularly for mutations within 25 bp of the motif (Figure 6). Recombination is the only known force able to produce such a strong skew, and our analysis shows that an order of magnitude higher fraction of these motifs form hotspots in PAR1, relative to the autosomes. This hypothesis has implications for how PAR1 manages to maintain such a uniquely high crossover rate. Firstly, it may imply a chromatin configuration in meiosis that facilitates access by PRDM9 to a high fraction of its binding sites. For instance, mouse chromatin axes are physically longer in PAR1 than the autosomes, also by an order of magnitude, potentially enabling greater access to recombination-initiating proteins [42]. Secondly, it would imply that a high fraction of bound sites go on to become recombination-promoting loci. Thus, we suggest that in humans, PRDM9 remains responsible for positioning recombination events, but that other factors may aid this protein in producing a high overall crossover rate.
We note that it is not clear our results are in contradiction with the finding of Prdm9-independent hotspots in the mouse PAR. For example, it may be that a back-up mechanism, independent of PRDM9, exists to ensure crossover occurs in the PAR. This back-up mechanism might, speculatively, be identical in the two mammals, but play a much larger role in mouse meiosis than in humans. This seems plausible to us based on PRDM9 binding target characteristics in the two species – the human PRDM9 target is GC-rich [54], like the PAR, and accordingly the PAR has many PRDM9 binding motifs. In contrast, studied mouse Prdm9 alleles recognize much more AT-rich motifs [46]. There were no matches, for instance, to the mouse motif TCnTGnTnCTT [86] in the section of mouse PAR assembled so far ( kb), whereas there were 9 matches to the human motif CCnCCnTnnCCnC, which has the same number of specified bases. The mouse motif is thus potentially rare or absent in its PAR, and likely to become rapidly eroded due to the phenomenon of gcBGC we have discussed here. Recombination in humans has been shown to lead to loss of PRDM9-binding motifs that become hotspots, via biased gene conversion (with a mechanism distinct from that of gcBGC). This phenomenon has been proposed to place evolutionary pressure on PRDM9 to evolve rapidly, as it is observed to do [52], to avoid eventual depletion of crossover locations essential for meiosis. The PAR represents an obvious genomic location where this problem might be especially acute, due to its small size and high recombination rate, perhaps even contributing to the rapid evolution of PRDM9. However, whether such rapid loss is occurring in the PAR in humans has not been possible for us to test, due to lack of statistical power. Interestingly, the force of gcBGC could even oppose the loss of PRDM9 target motifs, by creating other motifs, because human PRDM9 binding target motifs are GC-rich. Similarly, minisatellite mutation mechanisms may expand the number of PRDM9 binding sites in PAR1, by duplicating motif copies [58]–[60]. It is not clear, however, if these mechanisms can dominate over motif loss, and more study is required to better understand the evolutionary properties of PRDM9 binding sites, and more generally the DNA sequence, through time, in this intriguing region.
We have used genotype data from 135 previously published African-American pedigrees [47]. The pedigrees were drawn from cohorts in the CARe consortium: 70 families from the Jackson Heart Study (JHS) and 65 families from the Cleveland Family Study (CFS). After quality control filtering, 209 markers were available for CFS samples and either 215 or 180 or 192 markers for different subsets of JHS samples (more details are provided in Text S1). A union of these SNPs was performed, resulting in 220 SNPs, which were used to build the map in PAR1. A listing of these SNPs is provided in Dataset S1.
Each family had at least two children, and at least one parent genotyped. Crossovers were identified using an adaptation of the Lander-Green algorithm [87] that accommodates genotyping error and significant degrees of missing data, and has been published previously [47]. The algorithm has been summarized in Text S1 for completeness.
To increase power to detect crossovers near the pseudoautosomal boundary, we have included 100 SNPs from the X chromosome (Text S1). Fathers and sons were modelled to have one X-specific chromosome proximal to the pseudoautosomal boundary, and one ‘dummy’ chromosome with a fixed genotype sequence and no recombination. This improves the detection of both paternal and maternal crossovers near the pseudoautosomal boundary.
The algorithm estimates the posterior probability of crossover in each SNP interval across all parents. To build a male map, we add the probability of crossover for each SNP interval for all fathers, and divide by the total number of male meioses. We repeat this process for mothers to produce a female map. We post-process the cumulative posterior probability distribution of crossover over all SNP intervals for each parent to identify individual crossovers (Text S1).
The male and female genetic maps are provided in Dataset S2. The crossovers where both endpoints mapped into our regions of marker coverage are provided in Dataset S3 (male) and Dataset S4 (female).
The HapMap2 population-averaged LD-based map for PAR1 was downloaded from:
https://mathgen.stats.ox.ac.uk/impute/impute_v1.html#Download
Population-specific recombination maps were kindly provided by Colin Freeman from the Wellcome Trust Centre for Human Genetics, Oxford University.
LiftOver tool [88] was used to convert maps in builds 35/36 to builds 36/37.
A cDNA for the human PRDM9 B-allele was synthesised and cloned into a transient expression vector (pLEXm [89]) with an N-terminal Venus YFP tag. Large-scale transfections were performed in HEK293T cells as described [89]. Cells were harvested 72 hours after transfection and processed for ChIP-seq according to an online protocol used for the ENCODE project by the laboratory of Rick Myers [90]. Immunoprecipitation was performed using an Abcam rabbit polyclonal ChIP-grade anti-GFP antibody (ab290), and two technical replicates were performed. Uncrosslinked total chromatin DNA (without immunoprecipitation) was sequenced as a control sample. ChIP-DNA and control DNA were sequenced using 180 million paired 51 bp Illumina reads per replicate. Reads were aligned to hg19 and PCR duplicates were removed. Peak calling was performed using an in-house, maximum-likelihood-based peak calling algorithm that uses fragment coverage information from both sequencing replicates and the total chromatin control. Peaks were called at a p-value cutoff of . Further details of the protocol are provided in Text S1. The peaks are listed in Dataset S5. A separate manuscript describing the ChIP-seq results for the rest of the genome is in preparation.
To detect substitutions on the human and chimpanzee lineages, we downloaded the GRCh37-CHIMP2.1.4 (release 70) alignment available from Ensembl. The alignment was restricted to regions with accurate expected LD-based map rates (we removed the first and last 50 markers in the HapMap2 LD-based map, out of a total of 1385 markers, since power is reduced to detect the breakdown of LD there.). After this, the alignment contains approximately 1.2 Mb of sequence. For this analysis, we divided PAR1 into 2 kb regions, and included only those regions for analysis where at least 1 kb of the sequence was not repeat-masked and aligned without deletions or missing data on either lineage. A small number of regions were observed with total human/chimpanzee divergence greater than 5% and up to 11%. They were strongly clustered and represented clear outliers in the divergence distribution. These were filtered out from the analysis as they are not representative of PAR1 in general, and because we suspect that they represent mismapped or misaligned regions.
A stepwise search was performed to predict recombination rate using a linear model. The Aikake Information Criterion (AIC) was used to perform model selection and minimize overfitting. The full set of explanatory variables considered were the GC-content fraction, CpG content fraction and divergence rates for each of , , , , , and substitutions. Models were fit for substitutions in non-repeat DNA only.
Informed consent was provided by all the individuals participating in the study, and was approved by all of the institutions responsible for sample collection.
|
10.1371/journal.pbio.0050234 | Deletion of Ultraconserved Elements Yields Viable Mice | Ultraconserved elements have been suggested to retain extended perfect sequence identity between the human, mouse, and rat genomes due to essential functional properties. To investigate the necessities of these elements in vivo, we removed four noncoding ultraconserved elements (ranging in length from 222 to 731 base pairs) from the mouse genome. To maximize the likelihood of observing a phenotype, we chose to delete elements that function as enhancers in a mouse transgenic assay and that are near genes that exhibit marked phenotypes both when completely inactivated in the mouse and when their expression is altered due to other genomic modifications. Remarkably, all four resulting lines of mice lacking these ultraconserved elements were viable and fertile, and failed to reveal any critical abnormalities when assayed for a variety of phenotypes including growth, longevity, pathology, and metabolism. In addition, more targeted screens, informed by the abnormalities observed in mice in which genes in proximity to the investigated elements had been altered, also failed to reveal notable abnormalities. These results, while not inclusive of all the possible phenotypic impact of the deleted sequences, indicate that extreme sequence constraint does not necessarily reflect crucial functions required for viability.
| It is widely believed that the most evolutionarily conserved DNA sequences in the human genome have been preserved because of their functional importance and that their removal would thus have a devastating effect on the organism. To ascertain this we removed from the mouse genome four ultraconserved elements—sequences of 200 base pairs or longer that are 100% identical among human, mouse, and rat. To our surprise, we found that the mice lacking these elements are viable, fertile, and show no apparent abnormalities. This completely unexpected finding indicates that extreme levels of DNA sequence conservation are not necessarily indicative of an indispensable functional nature.
| Evolutionary conservation has become a powerful means for identifying functionally important genomic sequences [1,2]. Ultraconserved elements have been defined as a group of extremely conserved sequences that show 100% identity over 200 bp or greater between the human, mouse, and rat genomes [3]. This category of extreme evolutionary sequence conservation is represented by 481 sequences in the human genome, of which over half show no evidence of transcription. Further analysis of the distribution of these noncoding ultraconserved elements demonstrates that they tend to cluster in regions that are enriched for transcription factors and developmental genes [3], and a limited number of functional studies suggest a role for some of these noncoding elements in gene regulation [4–6].
Several hypotheses have been proposed to explain the extreme sequence constraint of ultraconserved elements, including strong negative selective pressure and/or reduced mutation rates [3]. The negative selection hypothesis postulates that crucial functions such as vital gene regulatory information is embedded within these sequences, while the reduced mutation rate hypothesis suggests that these sequences exist in a hyperrepaired or hypomutable state [3]. Recent analysis of human variation in these noncoding ultraconserved elements provides compelling evidence supporting negative selection as contributing to their extreme evolutionary conservation [7]. Furthermore, noncoding ultraconserved elements have also been shown to be significantly depleted in human segmental duplications and copy number variants, suggesting that disruption of their normal copy number may lead to reduced fitness [8]. In this study, we removed four carefully chosen noncoding ultraconserved elements in the mouse genome to directly explore a functional role for these elements in vivo.
To increase the probability of observing an associated phenotype in the ultraconserved null mice, we employed a variety of criteria in selecting the noncoding ultraconserved elements for deletion. We chose elements that showed tissue-specific in vivo enhancer activity in a mouse transgenic reporter assay that tended to recapitulate aspects of the expression pattern found in genes that were in their proximity (Figure 1) [6]. Other factors that were taken into account in prioritizing elements for deletion included their proximity to genes whose inactivation or alteration in expression result in specific phenotypes that we could screen for in the ultraconserved element deletion mice (Table 1). Elements meeting most of these criteria were chosen for removal and included: uc248, uc329, uc467, and uc482 (Figure 1) [3], representing 222, 307, 731, and 295 bp, respectively, of 100% identity between human, mouse, and rat.
All four noncoding ultraconserved elements were deleted from the mouse genome using standard mouse genetic engineering techniques, and removal was confirmed by PCR and Southern blot hybridization (Protocol S1). We first examined each line for the viability of homozygous/hemizygous knockout mice in mixed crosses, and observed that all four lines showed no reduction in the expected number of homozygous/hemizygous mice that were generated (Tables 2 and 3). Homozygous matings within the four lines revealed no significant differences in viability and litter size compared to the wild-type littermates (Table 4). We next examined body weight (up to 10 wk of age; Figure 2) and survival (up to 25 wk; see Materials and Methods), and found no significant differences compared to the wild-type littermates. Further analysis of a standard panel of 16 different clinical chemistry assays in each of the mouse lines detected only a few small differences compared to the wild-type littermates (Figure S1). Expression analysis of genes adjacent to each element by whole-mount in situ hybridization at embryonic day 11.5 (e11.5) revealed no apparent differences between null embryos and their wild-type littermates, except for a moderate reduction in forebrain expression of SRY-box containing gene 3 (Sox3) in uc482 null embryos (Figure S2). Quantification by real-time PCR suggested a slight reduction in Sox3 e11.5 head expression that was, however, insignificant (29.63 in wild-types compared to 23.66 in nulls, corresponding to 18S RNA expression; p = 0.64, unpaired t-test). General pathological analysis of 6-wk-old mice revealed no distinct differences compared to the wild-type littermates (Table S1), with one exception. The exception was one uc329 homozygous male having unilateral renal agenesis. Additional analysis of 102 uc329 homozygous null mice revealed a total of two mice (including the initial propositus) with one kidney, compared to none within the 30 uc329 wild-type littermates that were screened. Unilateral renal agenesis is estimated to occur in 1 to 1,000 live births in humans [9] and is asymptomatic and unassociated with a reduction in survival rate [10]. Possible explanations for unilateral renal agenesis in ∼2% of uc329 homozygous null mice in this study include a spontaneous event unassociated with the deleted element or a low penetrance phenotype caused by the absence of this element.
In addition to the above general screens, we screened each of these mouse lines for phenotypes specifically associated with the inactivation or dosage abnormality of the genes in proximity to the deleted ultraconserved elements. uc248 is bracketed by the genes doublesex and mab-3 related transcription factor 1 (DMRT1) and doublesex and mab-3 related transcription factor 3 (DMRT3) (Figure 1A). In humans, haploinsufficiency due to chromosomal aberrations within this region leads to XY sex reversal [11]. In mice, Dmrt1 homozygous knockouts exhibit defects in testicular development [12], while Dmrt3 function is unknown. In order to identify the phenotype associated with Dmrt3 deficiency for these studies, we deleted Dmrt3 from the mouse genome. All Dmrt3 null homozygous mice died from starvation at 2 mo of age due to dental malocclusions, and in addition some of the males exhibited male sexual development abnormalities (N. Ahituv, unpublished data). Based on these results, we extensively phenotyped uc248 homozygous null mice for sexual and dental abnormalities. Pathological analysis of both male and female sexual organs and teeth in 6-wk-old uc248 null mice revealed no obvious defects (Table S1). In addition, heterozygous and homozygous crosses exhibited no reduction in expected homozygous offspring (Tables 2 and 4).
uc467, the longest solitary noncoding ultraconserved element in the human genome (731 bp), lies inside the last intron of polymerase (DNA directed), alpha 1 (POLA1) adjacent to the aristaless related homeobox (ARX) gene (Figure 1C). Mutations in ARX in humans lead to a wide range of neurological and sexual development disorders [13,14], while hemizygous Arx null male mice die shortly after birth and have small brains and male sexual development abnormalities [15]. In addition, a duplication of this region in mice, caused by insertional mutagenesis, leads to embryonic lethality due to exencephaly accompanied by anophthalmia [16]. Detailed pathological examination of the reproductive organs and neuroanatomical examination of the brains of uc467 null mice revealed no apparent abnormalities (Table S1). In addition, the mice showed no obvious differences in the offspring expected from the hemizygous × heterozygous and hemizygous × homozygous crosses (Tables 3 and 4).
uc329 lies in the middle of an 80-kb intronic region of the hypothetical gene 0610012H03Rik in a region adjacent to the reticulocalbin 1, EF-hand calcium binding domain (RCN1) gene and two developmental transcription factors, Wilms tumor 1 (WT1) and paired box 6 (PAX6) (Figure 1B). Mutations in humans in WT1 and PAX6 respectively cause Wilms tumor and type 2 aniridia, while chromosomal deletions encompassing all four genes lead to WAGR syndrome. Mouse knockouts generated for Wt1 and Pax6 have a variety of phenotypes, the most notable being kidney and eye abnormalities, respectively. Detailed pathological analysis of the kidneys and eyes of the uc329 null mice revealed no significant differences compared to the wild-type littermates (Table S1), other than the ∼2% unilateral renal agenesis discussed above. Clinical chemistry tests revealed slightly higher urea nitrogen levels compared to the wild-type littermates (33.16 versus 26.16 mg/dl; p = 0.032, unpaired t-test; Figure S1), while creatinine levels, which are a more specific measure for kidney function, were similar to those in the wild-type littermates (0.25 versus 0.22 mg/dl; p = 0.165, unpaired t-test; Figure S1).
uc482 resides in a gene desert between the ATPase, Class VI type 11C (ATP11C) and SRY (sex determining region Y)-box 3 (SOX3) genes (Figure 1D). Human SOX3 mutations lead to X-linked mental retardation with isolated growth hormone deficiency [17] and hypopituitarism [18], while SOX3 dosage defects are suggested to cause hypopituitarism [18] and hypoparathyroidism [19]. In mice, deletion of Sox3 results in sexual development and pituitary abnormalities [20,21]. Pathological analysis of the reproductive organs of uc482 null mice revealed no significant abnormalities (Table S1), and hemizygous × heterozygous and hemizygous × homozygous crosses exhibited no reduction in expected homozygous/hemizygous offspring (Tables 3 and 4). Growth hormone abnormalities would be expected to lead to body weight irregularities, none of which were detected (Figure 2). Calcium levels were also normal (Figure S1), supporting a lack of marked abnormalities in parathyroid gland function.
Based on the compelling evidence that ultraconserved elements are conserved due to functional constraint, it has been proposed that their removal in vivo would lead to a significant phenotypic impact [7,8]. Accordingly, our results were unexpected. It is possible that our assays were not able to detect dramatic phenotypes that under a different setting, for instance, outside the controlled laboratory setting, would become evident. Moreover, possible phenotypes might become evident only on a longer timescale, such as longer generation time. It is also possible that subtler genetic manipulations of the ultraconserved elements might lead to an evident phenotype due to a gain-of-function-type mechanism. All four elements examined in this study demonstrated in vivo enhancer activity when tested in a transgenic mouse assay (Figure 1) [6], which would suggest regulatory element redundancy as another possible explanation for the lack of a significant impact following the removal of these specific elements. Just as gene redundancy has been shown to be responsible for the lack of phenotypes associated with many seemingly vital gene knockouts, regulatory sequence redundancy [22] can similarly provide a possible explanation for the lack of a marked phenotype in this study. While our studies have not defined a specific need for the extreme sequence constraints of noncoding ultraconserved elements, they have ruled out the hypothesis that these constraints reflect crucial functions required for viability.
The basic technology used for gene targeting and screening has been described previously [23]. Briefly, the four selected ultraconserved elements were removed in W4/129S6 mouse embryonic stem cells (Taconic, http://www.taconic.com/) by standard replacement of a LoxP-flanked neomycin cassette. To avoid potential regulatory effects due to the neomycin gene cassette, we subsequently removed it by Cre-mediated recombination of LoxP sites in the embryonic stem cells. All positive colonies in each stage were confirmed by PCR and Southern analysis (Protocol S1) and then injected into C57BL/6J blastocyst stage embryos. Chimeric mice were subsequently crossed to C57BL/6J mice, generating agouti offspring that were heterozygous/hemizygous for the ultraconserved element deletion and were intercrossed to generate homozygous ultraconserved null mice. Genotyping was carried out using standard PCR techniques (Protocol S1).
Eight males and eight females from each line and wild-type littermates were analyzed for survival up to 25 wk. Mice were housed in a temperature-controlled room under a 12-h light/dark cycle, given free access to water, and fed ad libitum on a standard chow. No lethality was observed for any of the strains during the period of study.
Serum samples from at least six males and six females at 10–14 wk of age from each line were analyzed using the automated spectrophotometric chemistry analyzer Hitachi 917 at Marshfield Laboratories (http://www.marshfieldlaboratories.org/) following standard protocols.
Four e11.5 wild-type embryos were analyzed for each gene. Genes that were positive for expression at this time point were further analyzed for expression differences using four homozygous null and four wild-type littermates at e11.5. Briefly, embryos were fixed overnight in 4% paraformaldehyde followed by methanol washes. Whole-mount RNA in situ hybridization was carried out using standard protocols [24] with antisense digoxigenin-labeled riboprobes. The following vectors were used as templates for probes: Dmrt1 (kind gift from D. Zarkower, University of Minnesota), Dmrt2 (IMAGE 1248080, http://image.llnl.gov/), Dmrt3 (IMAGE 6404988), Pax6 (IMAGE 4504106), Rcn1 (IMAGE 6414128), 0610012H03Rik (IMAGE 5042053), Wt1 (RNA probe 777, GenePaint.org, http://genepaint.org/), Pola1 (IMAGE 894396, 30063811, and 30103897), Arx (IMAGE 5707995), Atp11c (IMAGE 30843359), and Sox3 (IMAGE 5717161). Stained embryos were analyzed using a Leica (http://www.leica.com/) MZ16 microscope and photographed with a Leica DC480 camera.
Total RNA was extracted using TRIzol reagent (Invitrogen, http://www.invitrogen.com/) from the heads of four uc482 homozygous/hemizygous null and four wild-type littermates at e11.5, and pooled separately. Following reverse transcription with SuperScript First-Strand Synthesis System (Invitrogen), real-time PCR was performed using Sox3-specific primers (forward: agcgcctggacacgtacac; reverse: atgtcgtagcggtgcatct), QuantumRNA Universal 18S (Ambion, http://www.ambion.com/), and the SYBR Green PCR Master Mix (Applied Biosystems, http://www.appliedbiosystems.com/) on a 7500 Fast Real-Time PCR System (Applied Biosystems). All procedures and calculations were carried out according to manufacturers' recommendations.
Two male and two female 6-wk-old mice from each knockout line and wild-type littermates were submitted to the Comparative Pathology Laboratory at the University of California Davis. Tissues were fixed in 10% phosphate buffered formalin for at least 24 h and processed using routine methods to Hematoxylin-and-Eosin-stained sections and subsequently analyzed for any abnormalities.
The Entrez Gene (http://www.ncbi.nlm.nih.gov/sites/entrez?db=gene) accession numbers for the mouse genes discussed in this paper are Arx (11878), Atp11c (54668), Dmrt1 (50796), Dmrt2 (226049), Dmrt3 (240590), Pax6 (18508), Pola1 (18968), Rcn1 (19672), Sox3 (20675), and Wt1 (22431). The Entrez Gene accession numbers for the human genes discussed in this paper are ARX (170302), ATP11C (286410), DMRT1 (1761), DMRT2 (10655), DMRT3 (58524), PAX6 (5080), POLA1 (5422), RCN1 (5954), SOX3 (6658), and WT1 (7490). The GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) accession number for hypothetical gene 0610012H03Rik is NM_028747. The OMIM (http://www.ncbi.nlm.nih.gov/sites/entrez?db=OMIM) accession numbers for the genetic disorders discussed in this paper are type 2 aniridia (106210), WAGR syndrome (194072), and Wilms Tumor (194070).
|
10.1371/journal.ppat.1001110 | Cellular Entry of Ebola Virus Involves Uptake by a Macropinocytosis-Like Mechanism and Subsequent Trafficking through Early and Late Endosomes | Zaire ebolavirus (ZEBOV), a highly pathogenic zoonotic virus, poses serious public health, ecological and potential bioterrorism threats. Currently no specific therapy or vaccine is available. Virus entry is an attractive target for therapeutic intervention. However, current knowledge of the ZEBOV entry mechanism is limited. While it is known that ZEBOV enters cells through endocytosis, which of the cellular endocytic mechanisms used remains unclear. Previous studies have produced differing outcomes, indicating potential involvement of multiple routes but many of these studies were performed using noninfectious surrogate systems such as pseudotyped retroviral particles, which may not accurately recapitulate the entry characteristics of the morphologically distinct wild type virus. Here we used replication-competent infectious ZEBOV as well as morphologically similar virus-like particles in specific infection and entry assays to demonstrate that in HEK293T and Vero cells internalization of ZEBOV is independent of clathrin, caveolae, and dynamin. Instead the uptake mechanism has features of macropinocytosis. The binding of virus to cells appears to directly stimulate fluid phase uptake as well as localized actin polymerization. Inhibition of key regulators of macropinocytosis including Pak1 and CtBP/BARS as well as treatment with the drug EIPA, which affects macropinosome formation, resulted in significant reduction in ZEBOV entry and infection. It is also shown that following internalization, the virus enters the endolysosomal pathway and is trafficked through early and late endosomes, but the exact site of membrane fusion and nucleocapsid penetration in the cytoplasm remains unclear. This study identifies the route for ZEBOV entry and identifies the key cellular factors required for the uptake of this filamentous virus. The findings greatly expand our understanding of the ZEBOV entry mechanism that can be applied to development of new therapeutics as well as provide potential insight into the trafficking and entry mechanism of other filoviruses.
| Filoviruses, including Zaire ebolavirus (ZEBOV), are among the most pathogenic viruses known. Our understanding of how these viruses enter into host cells is very limited. A deeper understanding of this process would enable the design of better targeted antiviral therapies. This study defines in detail, key steps of ZEBOV cellular uptake and trafficking into cells using wild type virus as well as the host factors that are responsible for permitting virus entry into cells. Our data indicated that the primary mechanism of ZEBOV uptake is a macropinocytosis-like process that delivers the virus to early endosomes and subsequently to late endosomes. These findings aid in our understanding of how filoviruses infect cells and suggest that disruption of macropinocytosis may be useful in treatment of infection.
| Zaire ebolavirus (ZEBOV, Genbank:AF086833), a member of the family Filoviridae, genus Filovirus, causes a highly fatal hemorrhagic fever in humans and non-human primates. Over the past three decades numerous human outbreaks have occurred in Central Africa involving hundreds of cases with fatality rates ranging from 50–89% [1]. In addition, outbreaks of ZEBOV infection have been implicated in deaths of tens of thousands of gorillas, chimpanzees and duikers in Central and Western Africa posing a considerable threat to the wildlife and ecology in those areas [2]. Due to a very high case fatality rate in humans, significant transmissibility of the virus, lack of effective preventive or therapeutic measures against the disease, ZEBOV is considered a serious emerging viral pathogen. Currently no specific therapy or vaccine is approved for human or animal use against this pathogen.
As for other members of Filoviridae, ZEBOV is morphologically distinct from other animal viruses. The virions are long and filamentous with an average length of 800–1000 nm and a diameter of about 80 nm but can form a variety of shapes ranging from straight rods to closed circles [3]. Virions are surrounded by a host cell-derived lipid envelope. The envelope contains virally-encoded glycoprotein (GP) spikes composed of homotrimers of two virally-encoded glycoproteins, GP1 and GP2. The approximately 19 kb single-stranded, negative-sense genomic RNA complexed with nucleocapsid, VP35, VP30 and L proteins form the nucleocapsid, while VP40 forms the matrix that underlies the viral membrane [4].
Like all viruses, ZEBOV largely relies on host cell factors and physiological processes for key steps of its replication cycle. Identification of these processes and factors will not only allow a better insight into pathogenic mechanism, but may identify novel targets for future therapeutic development. As the first step of replication, entry into the host cell is an attractive target for therapeutic intervention as infection can be stopped before virus replication disrupts cellular functions. However, the entry mechanism of ZEBOV, and that of other large enveloped viruses is very limited.
Many enveloped viruses, including ZEBOV, require endocytosis to infect cells. The internalized virus is transported through successive endocytic vesicles to reach a vesicle/compartment where conditions are conducive (low pH and/or presence of proteolytic enzymes) for the GP to attain a suitable conformation needed for membrane fusion [5], [6], [7]. Upon fusion of viral and endocytic membranes, the capsid moves into the cell cytoplasm to begin genome replication. Several distinct endocytic mechanisms exist in mammalian cells. They are distinguished from each other on a number of criteria including the size and morphology of endocytic vesicles, the type of cargo they carry, the cellular factors involved in their control and their origins and destinations [8]. Different viruses employ different routes of endocytosis, and the route taken by a given virus largely depends on the receptor it interacts with.
Clathrin-mediated endocytosis (CME) is the best understood endocytic pathway. A number of viruses including Influenza A (Genbank:M73524), Semliki forest (Genbank:X04129) and vesicular stomatitis viruses (VSV; Genbank:J02428) employ this pathway for entry [9], [10]. The internalization of the virus-receptor cargo occurs in specialized areas of cell membrane called clathrin-coated pits (CCPs). CCPs are formed on the cytoplasmic face of the plasma membrane through sequential assembly of proteins including clathrin that form a cage-like structure lining the cytoplasmic side of the pit. The pit then invaginates and buds from the plasma membrane forming a clathrin-coated vesicle approximately 120 nm in diameter containing the internalized cargo. Subsequently, the vesicle sheds its clathrin coat, a prerequisite for further trafficking and merging with other compartments. A number of accessory, adaptor and signaling molecules participate in this process, and provide a tight regulation of the pathway. Some, such as accessory protein-2 (AP2) and Eps15, are specifically associated with CCPs, while others such as dynamin, which is responsible for vesicle budding from the plasma membrane, are shared with other endocytic pathways [8].
Caveolin-mediated endocytosis (CavME), first observed for the cellular uptake of simian virus 40 (SV40) [11], differs from CME in terms of internalization mechanism and vesicular transport route. Caveolae are flask-shaped invaginations in the plasma membrane that are rich in caveolin protein, and are predominantly associated with cholesterol-rich plasma membrane microdomains termed lipid rafts. Therefore, extraction or perturbation of membrane cholesterol severely impedes entry of viruses that use CavME. Vesicles derived from CavME are indicated by the presence of caveolin and are termed caveosomes. Other cellular factors such as Eps15-related (Eps15R) protein are thought to be specific for CavME, but as with CME, dynamin is still required for severing of caveolae from the plasma membrane. Another distinguishing feature is that caveloae are smaller than CCPs and have an average diameter of approximately 60–80 nm [8].
Recently, the importance of macropinocytosis, as a distinct endocytic uptake mechanism for virus infection, has started to be realized for some viruses [9]. Macropinocytosis is associated with membrane ruffles such as those formed by filopodia and lamellipodia, which are outward extensions of the plasma membrane driven by actin polymerization underneath the membrane surface [12]. When a ruffle folds back upon itself a cavity can be formed. Subsequent fusion of the distal end of the loop with the plasma membrane results in formation of a large vesicle called a macropinosome. These can range in size from 200 to 10,000 nm across and take up cargoes of similar dimensions [9]. Morphological and regulatory characteristics that distinguish macropinocytosis from other endocytic processes have also begun to emerge [8], [9], [13], [14]. Macropinosomes are best characterized for uptake of fluid phase markers such as high molecular weight dextran and horse-radish peroxidase and is sensitive to inhibitors of Na+/H+ exchangers, such as amilorides [14]. As for CavME, they are dependent on cholesterol-rich lipid rafts, but dynamin is not required. Instead, scission of macropinosomes appears to require CtBP/BARS [9], [15], [16]. Other work indicates the involvement of cell signaling factors PI3K, Akt, PKC, PLCγ and PLC-A2 that act to promote membrane ruffling by stimulating actin remodeling through Rac and cdc42 [9].
All endocytic pathways used by viruses serve to deliver virus to vesicles and compartments conducive to virus membrane fusion and release of the core into the cell cytoplasm at a site where replication proceeds optimally. Many endocytic pathways share common features, such as acidification, yet each virus type appears to prefer one trafficking pathway over others and misdirection into alternative pathways can result in inhibition of infection. For most enveloped viruses, the point at which membrane fusion occurs appears to be at the early or late endosome stage. This evidence has been gathered by comparing pH-sensitivity of the GP to known pH of the endosome at different stages of maturation. More recently, the use of dominant negative GTPases, that are involved in endosomal maturation, have been used in determining virus exit points from the endosome [17], [18], [19]. In general, the two methods agree but provide little detail as to whether viruses have additional requirements in terms of site of release other than the early or late endosome.
Currently, a detailed understanding of ZEBOV endocytosis and trafficking is lacking. Each of the previous studies on understanding ZEBOV entry pathway have indicated involvement of different pathways, including CME [20], [21], CavME [20], [22] and a Rho GTPase-dependent pathway that may suggest involvement of macropinocytosis [23]. These conflicting findings may be due to the use of surrogate models of ZEBOV such as pseudotyped retroviruses, which are morphologically and biochemically distinct from wild type filamentous ZEBOV and/or reliance on one analytical approach, such as use of pharmacological agents, which are likely to act on more than one cellular target [24].
Here we have used multiple independent approaches employing replication-competent, infectious ZEBOV and/or morphologically comparable virus-like particles (VLPs). We have examined the contribution of each endocytic pathway to ZEBOV entry and infection by quantitative analysis. The work involves measuring the impact of drugs, siRNA and/or expression of well characterized dominant negative (DN) mutants of cell trafficking proteins on virus entry and infection. We also use fluorescently-labeled virus-like particles (VLPs) to follow virus internalization and trafficking through different endocytic compartments. The product of combining all these approaches provides, for the first time, an accurate and detailed description of ZEBOV uptake mechanism. Our data clearly indicate that wild type ZEBOV enters and infects Vero and HEK293T cells independently of clathrin, caveolae and dynamin. Instead, virus entry required the presence of cholesterol in the cell membrane and was inhibited by the amiloride derivative, EIPA. A marked induction in fluid-phase uptake was also observed shortly after virus binding to cells and internalized virus particles showed significant colocalization with high molecular weight dextran. In addition, inhibition of p53-activated kinase (Pak1) or CtBP/BARS resulted in significant reduction in virus entry and infection. Importantly, the virus particles appear to stimulate uptake through this pathway directly by promoting localized actin polymerization and this is consistent with our previous work where the GP triggered the PI3 kinase signaling cascade and Rac1 activity [25]. No evidence for involvement of clathrin- or caveolae-dependent endocytosis was seen. Instead, the primary mechanism of virus uptake appears closely related to macropinocytosis. Subsequent to internalization, the virus utilizes the conventional endolysosomal pathway and is trafficked through early and late endosomes before membrane fusion takes place. This study provides novel information regarding ZEBOV entry, and is likely to be useful in understanding the entry mechanism of other filoviruses.
A recent study suggested that Ebola virus uses CME for cellular entry [21], while an earlier study had implicated both CME and CavME [20], [21]. However, these studies utilized either pseudotyped virus, which due to morphological and/or biochemical differences may not accurately depict the ZEBOV entry pathway, and/or relied solely on the use of pharmacological agents, which may alter multiple processes important for membrane trafficking. To more closely examine the role of the each endocytic pathway we used replication-competent infectious virus and morphologically comparable ZEBOV virus-like particles (VLPs) to determine the impact of specific dominant-negative (DN) forms of Eps15 (OMIM:600051) and caveolin-1 (cav-1, OMIM:601047) on infection and VLP uptake into cells. Eps15 and caveolin-1 (cav-1) are required for the formation and trafficking of CME and CavME vesicles respectively, and their DN forms inhibit the respective endocytosis with high specificity [26]. HEK293T cells were transfected with plasmid encoding GFP alone or GFP-tagged forms of DN-Eps15 or DN-Cav1, and subsequently infected with ZEBOV. Infected cells were detected by immunofluorescence staining for ZEBOV matrix protein (VP40) protein and the proportion of cells that co-expressed the transfected protein and ZEBOV VP40 (infection marker) was calculated. It was found that the proportion of ZEBOV-infected cells in cultures expressing DN-Eps15-GFP or DN-Cav1-GFP was not significantly different (P>0.05) to that in cultures expressing GFP alone, indicating that neither DN protein had a significant impact on ZEBOV infection (Fig. 1A). This lack of effect of either DN protein on ZEBOV entry was confirmed using a sensitive contents-mixing entry assay that measures virus-endosomal membrane fusion by monitoring luciferase release from VLPs into the cell cytoplasm [25]. Expression of either DN protein failed to have any significant effect on entry of ZEBO-VLP (P>0.05; Fig. 1B). As control, VLPs bearing VSV envelope glycoprotein (VSV-VLP) were used. VSV is known to use CME for cellular entry [27]. Entry of VSV-VLP was significantly inhibited only in cells expressing DN-Eps15 (Fig.1B, P<0.001). To confirm that DN-Cav1 expression impacted caveolar endocytosis, murine leukemia virus 10A1 (MLV-10A1) infection and cholera toxin B subunit (CTxB, Pubchem:53787834) uptake were measured as both processes are known to require CavME [28], [29]. 10A1 infection of cells expressing DN-Cav-1 was reduced by half (Fig. 1C) and is consistent with previously reported observations [28]. Uptake of CTxB (Fig. 1D) was more strongly inhibited, as cells expressing DN-Cav1 had little CTxB inside the cytoplasm, indicating that DN-Cav1 was functional, blocking caveolar endocytosis.
As a further test, colocalization of internalized GFP-labeled ZEBO-VLPs (gfpZEBO-VLP) with established markers of CME (clathrin light chain A; OMIM:118960 and transferrin; OMIM:190000) or CavME (caveolin-1) pathways was examined. Confocal microscopy revealed no significant colocalization of gfpZEBO-VLP with any of the markers used (Fig. 1E). Similar results were obtained when Vero cells were used (not shown). Taken together, the above findings indicated that neither CME nor CavME plays a major role in entry and infection of ZEBOV into HEK293T or Vero cells.
Dynamin (OMIM:602377) is a large GTPase and plays a critical role in numerous endocytic pathways including CME and CavME as well as some of the non-clathrin/non-caveolin-dependent (NC) pathways [8]. Dynamin acts by mediating the release of newly-formed endocytic vesicles from the plasma membrane. To determine ZEBOV dependence on dynamin, the effect of dynasore (Pubchem:56437635), a potent and specific dynamin inhibitor [30] was tested. A recombinant infectious ZEBOV that encodes GFP (gfpZEBOV) was used. This virus is comparable to wild-type ZEBOV in terms of replication and cytopathic effects (CPE) in cultured cells but has been engineered to express GFP as an infection marker [31]. As control, a recombinant infectious VSV that encoded red fluorescent protein (rfpVSV) was used. Dynasore treatment of Vero cells greatly reduced rfpVSV infection but failed to have any significant effect on infection by gfpZEBOV even at the highest concentration tested (Fig. 2A, B). Similar results were obtained in HEK293T cells (not shown). This result was confirmed using the VLP-based entry assay. Just like the gfpZEBOV, entry of ZEBO-VLP was unaffected at any of the doses used while VSV-VLPs were strongly inhibited by dynasore in a dose-dependent manner (Fig. 2C). To ensure that dynasore inhibited dynamin-mediated endocytosis, its effect on internalization of transferrin (CME marker) or CTxB (CavME marker) was determined. Confocal microscopy revealed that treatment reduced internalization of both markers by >80% and 96% respectively in Vero cells (Fig. 2D).
As a further test of the dynamin independence of ZEBOV infection cells were made to express a DN form of dynamin-2 (Dyn2-K44A) and VLP entry assays were performed. As with dynasore, there was a significant drop in the entry of VSV-VLP in cells transfected with Dyn2-K44A (P<0.05). In contrast, the entry of ZEBO-VLPs actually increased significantly (P<0.05), suggesting that the suppression of dynamin function may enhance entry of ZEBOV (Fig. 2E). Furthermore, the majority of GFP labeled ZEBO-VLPs (gfpZEBO-VLPs) did not colocalize with endogenous dynamin at any point up to 60 min after cell contact (Fig. 2F). These findings indicated that cell entry of ZEBOV is independent of dynamin and that dynamin activity may actually redirect virus to a non-productive pathway.
Many of the cellular endocytic pathways including CavME, macropinocytosis and certain NC pathways occur in cholesterol-rich membrane microdomains such as lipid rafts. An earlier study suggested that lipid rafts may play a role in ZEBOV infection [32]. Consistent with this, we found significant co-localization of gfpZEBO-VLPs with lipid rafts during entry (Fig. 3A). Furthermore, methyl-β cyclodextrin (Pubchem:3889506) or nystatin (Pubchem:6433272), which disrupt lipid rafts by extracting or sequestering cholesterol out of the plasma membrane, respectively were able to block gfpZEBOV infection in a dose-dependent manner (Fig. 3B). Similar effects of both drugs were observed when tested using the ZEBO-VLP-based entry assay (not shown). These data indicated that cholesterol-rich lipid raft domains are the likely site of ZEBOV entry.
The above findings clearly indicated that ZEBOV uptake occurs through a dynamin-independent, lipid raft-dependent, non-clathrin/non-caveolar endocytic mechanism but remains cholesterol dependent. Macropinocytosis is one pathway that is known to be cholesterol-dependent, but independent of clathrin, caveolin and dynamin and has been shown important for uptake of vaccinia virus into cells as well as bacteria [9]. Also, our previous work indicated involvement of PI3K and Rac1 (a rho family GTPase) in ZEBOV entry and infection [25]. Work by others had also indicated involvement of Rho GTPase in Ebola virus entry [23]. Each of these signaling proteins is also thought to be important for macropinocytosis [13], [33].
To assess the involvement of macropinocytosis, the effect of EIPA (5-(N-ethyl-N-isopropyl amiloride; Pubchem:1795) on ZEBOV infection was determined. EIPA, an amiloride, is a potent and specific inhibitor of Na+/H+ exchanger activity important for macropinosome formation [34], [35], [36], [37]. Consistent with this activity, EIPA caused a significant reduction (>80%) in the uptake of high molecular weight dextran, a marker of macropinosomes (Fig. 4A and B). When tested in Vero cells, a dose-dependent inhibition of gfpZEBOV infection was observed in the presence of EIPA (Fig. 4C, top panels; Fig. 4D), while infection by VSV was not significantly affected (Fig. 4C, middle panels; Fig. 4D). As used, EIPA had no significant cytotoxic effect as assessed by cell monolayer integrity (Fig. 4C, bottom panels). Counting of infected cells revealed that VSV infection was inhibited by 30% but increasing the dosage of the drug did not further reduce infection, indicating a small portion of VSV infection may occur through an EIPA-sensitive pathway. In contrast, the majority of ZEBOV infection inhibition was dosage dependent, potent and indicative of inhibition of a single uptake pathway (Fig. 4D). Similar results were observed in HEK293T cells (data not shown).
To rule out the possibility that EIPA blocked ZEBOV infection at a post-entry step, the VLP entry assay was used. Here, EIPA treatment had no significant effect on entry of VSV-VLP (P<0.05) while the level of ZEBO-VLP entry was inhibited similarly to that seen for infectious virus (Fig. 4E). The impact of EIPA on virus binding to cells was also tested. Cells were pretreated with EIPA and then incubated with luciferase-containing ZEBO-VLPs for 10 min on ice. Unbound particles were washed away and then the amount of VLP associated with cells was measured by lysis in non-ionic detergent to release virus-encapsulated luciferase. Compared to DMSO-treated (control) cells, no significant difference was observed in luciferase activity in samples that were treated with EIPA, indicating that ZEBO-VLP binding to cells was unaffected (Fig. 4F). Finally, to directly visualize the effect of EIPA on virus uptake, Vero cells treated with DMSO or EIPA were incubated with gfpZEBOV-VLPs. Confocal microscopy revealed that there was a marked drop (3.5-fold) in gfpZEBO-VLP uptake in cells treated with EIPA as compared to that in DMSO-treated cells (Figs. 4G and H).
Vaccinia virus was shown to induce fluid phase uptake and exhibit colocalization with fluid phase markers such as dextran [36]. To see if a similar induction of fluid phase uptake was seen with ZEBOV, dextran and virus were incubated together on cells and examined by confocal microscopy. Starting within 10 min and continuing until at least 60 min post-binding, at any one time, approximately 20% of internalized VLPs overlapped with dextran (Fig. 4I). An additional 20–30% of the remaining VLPs were also found juxtaposed to vesicles containing dextran, indicating a close association with this compartment. Furthermore, while performing these experiments we observed that cells incubated with ZEBO-VLPs appeared to have more dextran containing vesicles than intact cells. Indeed, when studied in detail, a (2–3 fold) increase in the number of dextran-containing vesicles per cell was seen after incubation with ZEBOV as compared to cells incubated with VSV or medium alone (Fig. 4J). A similar outcome was seen for cells incubated with VLPs (not shown).
The above data indicated that cellular uptake of ZEBOV primarily occurs through a pathway that has characteristics of macropinocytosis. Another hallmark of macropinocytosis is its dependence on the activity of Pak1 [9]. Therefore, the role of Pak1 in entry of ZEBOV was investigated. First, we measured the effect of siRNA-mediated suppression of endogenous Pak1 and found that the infection of gfpZEBOV was significantly reduced in cells transfected with Pak1 siRNA (Fig. 5A). To confirm, cells transfected with plasmid encoding wild-type Pak1 or DN Pak1 were challenged with gfpZEBOV. The infection of gfpZEBOV was reduced >95% in cells expressing the DN Pak1 protein than in cells expressing the wild-type Pak1 protein (Fig. 5B). A similar effect of DN Pak1 was observed when ZEBO-VLPs were tested in the entry assay (not shown). The protein CtBP/BARS, is also known as important for macropinocytosis [15], [16] and substitutes for dynamin in promoting vesicle scission from the plasma membrane. Again, siRNA were used to suppress expression and the impact on infection measured. Two independent siRNA were able to suppress expression of CtBP/BARS by >70% and 80% respectively (Fig. 5C, left panels). Infection was also reduced by 50% and 90% respectively (Fig 5C, right). This observation indicated that the suppression of CtBP/BARS expression must cross a threshold before becoming limiting to ZEBOV infection but plays an important role. Both sets of data support roles for Pak1 and CtBP/BARS in ZEBOV infection.
Macropinocytosis is heavily actin-dependent. Actin is required for the formation of plasma membrane ruffles in macropinosome formation, as well as trafficking of macropinosomes into the cell [38]. Ligands that utilize macropinocytosis often promote changes in the cell actin dynamics by regulating various cellular proteins involved in controlling F-actin assembly and disassembly. Arp2 protein is an integral component of a multi-protein complex that serves as a nucleation site for de novo actin assembly. We observed a significant increase in the size of Arp2-containing complexes shortly after ZEBOV binding to cells (Fig. 5D, E). Analysis of the data indicated a 2-fold increase in the number of large (>0.25 µm2) Arp2-containing complexes (Fig. 5D). A similar outcome was seen with cells incubated with VLPs (not shown) and a significant proportion of VLPs were associated with Arp2 complexes (Fig. 5F). Further support for a role of actin in ZEBOV entry came from the observation that gfpZEBO-VLPs were associated with F-actin foci within the interior of the cell but this was not seen for VSV-VLPs (Fig. 5G). Similarly, the gfpZEBO-VLPs were also seen associated with vasodilator-stimulated phosphoprotein (VASP), an actin-associated protein that promotes actin nucleation (Fig. 5H). In each of these cases, VLPs and staining for each marker often did not completely overlap. Instead VLP and actin or VASP often were closely juxtaposed and is consistent with nucleation occurring around vesicles containing the VLP. These observations suggested that the virus actively promotes actin assembly and associates with actin-based structures to facilitate its uptake and/or trafficking.
The above findings indicated that ZEBOV is primarily internalized by a macropinocytosis-like pathway in Vero and HEK293 cells. However, the subsequent trafficking route to the site of penetration into the cytoplasm remained unknown. We found that fluorescently-labeled ZEBOV particles significantly co-localize with early endosomal antigen-1 (EEA1; OMIM:605070) shortly after incubation with cells (Fig. 6A). At any time up to 60 min after the start of incubation, more than 30% of VLPs were associated with this marker (Fig. 6B). This confirmed a role for endocytic uptake into cells and suggested that following internalization, ZEBOV is delivered to an EEA1-positive compartment, likely sorting endosomes. Typically, the cargo from EEA1-positive compartments is delivered to early endosomes followed by trafficking to late endosomes. These vesicles are characterized by the presence of Rab5 (OMIM:179512) and Rab7 (OMIM:602298) GTPases on the cytoplasmic face of the vesicle, respectively, which play a key role in regulating their trafficking. Consistent with a role for early and late endosomes and in contrast to the lack of effect of DN Eps15 and Cav-1 expression, GFP-tagged DN Rab5 or DN-Rab7 resulted in significant reduction (P<0.001 for each) in infection by gfpZEBOV (Fig. 6C). To determine if the effect was due inhibition of virus entry, VLP entry assays were performed. As compared to the negative control (GFP alone), wild-type Rab5 had no significant effect on entry of either ZEBO-VLP or VSV-VLPs, while there was >50% reduction in entry of both ZEBO-VLPs in cells transfected with either DN-Rab5 or DN-Rab7. The level of entry inhibition seen for ZEBO-VLP was similar to that of VSV-VLPs (Fig. 6C) and indicated that like VSV, ZEBOV is taken up by Rab5-dependent early and Rab7-dependent late endosomes. However, this does not mean that both virus types are present in the same vesicle population but that similar trafficking proteins are required at this stage of endocytosis. Currently, it is unknown whether ZEBOV envelope fusion occurs in late endosomes or further trafficking to a different compartment is needed.
Endocytosis offers an efficient way for viruses to cross the significant physical barrier imposed by the plasma membrane and to traverse the underlying cortical matrix. Viruses have also evolved to target distinct endocytic pathways that are capable of delivering the capsid into the cell cytoplasm at sites suitable to initiate replication and to avoid destructive compartments like the lysosome. Understanding the pathway of virus entry and deciphering the mechanism regulating it is important for understanding viral pathogenesis as virus entry into host cell is the first critical step in pathogenesis of infection.
While there is ample evidence that ZEBOV enters cells through endocytosis in a pH-dependent manner [6], [7], [20], the specific endocytic and trafficking pathways have not been clearly defined. Previous studies to elucidate the ZEBOV entry pathway have produced conflicting findings. Most of these studies relied on the use of retrovirus-based pseudotypes in which the Ebola virus GP is coated onto the surface of a retrovirus capsid containing a recombinant genome. The use of this system overcomes the need for high bio-containment but suffers from not having native virus morphology, GP density, and other biochemical characteristics. One early study on ZEBOV uptake using pseudotyped virus indicated caveolae as important [22] but later work indicated that caveolin activity was not required [39]. In contrast, a recent study concluded that clathrin-mediated endocytosis was the major entry pathway for ZEBOV [21]. While multiple approaches were used, including dominant-negative mutant expression and siRNA to specifically disrupt clathrin-mediated endocytosis, the key data was obtained using lentivirus-based retroviral pseudotypes. In comparison, previous work using wild type virus [20] implicated both clathrin and caveolar endocytosis in entry of ZEBOV. However, only pharmacological inhibitors were used in this study and drug specificity was not examined, making interpretation difficult. Indeed, the only evidence of clathrin involvement in infection was provided using chlorpromazine. Chlorpromazine is a useful drug and there is ample evidence indicating that it disrupts clathrin-coated pits, but it has recently been demonstrated to also interfere with biogenesis of large intracellular vesicles such as phagosomes and macropinosomes [24].
Here, by combining distinct and independent approaches we have performed a detailed analysis of each major endocytic pathway and have obtained, a clear and accurate picture of how ZEBOV enters the cell and identified important cellular proteins that are required. Careful assessment of specificity and functionality of each pathway was performed and correlated to infection and virus uptake. Replication-competent infectious ZEBOV, as well as ZEBO-VLPs (which are morphologically similar to infectious ZEBOV and contain the native matrix protein in addition to GP) were used to study the virus entry mechanism. Drugs were used to inhibit pathways but issues of specificity and pleiotropy were assessed by testing the function of each pathway after treatment. This was done by using independent markers such as transferrin, CTxB and high molecular weight dextran for CME, CavME and macropinocytic uptake respectively. We also assessed the association of fluorescent VLPs with each marker as well as markers of each endocytic compartment being examined. Furthermore, highly specific dominant-negative mutants and/or siRNAs were also used to corroborate the data obtained by pharmacological inhibitors. Importantly, throughout this work a sensitive contents-mixing virus entry assay was used in discriminating against blocks in virus entry versus blocks in downstream steps in the infection cycle. This is particularly important to do when using drugs that often affect multiple cellular functions. It is noteworthy that in each case, virus infection with wild type or the GFP-expressing ZEBOV correlated exactly with the outcomes of the VLP-based assays. This approach gives a highly detailed view of the mechanism of ZEBOV uptake into cells.
Unlike previous studies [20], [21], [22], we found no evidence for the involvement of either CME or CavME in ZEBOV entry and infection. However, there was strong association of fluorescently-labeled ZEBO-VLPs with lipid rafts, and a marked reduction of ZEBOV infection by MBCD or nystatin, as reported previously [32]. This signified that cholesterol-rich lipid raft domains are required for productive entry of the virus. However, cholesterol-rich membrane microdomains play important roles in many forms of endocytosis including caveolae-dependent, non-clathrin/non-caveolar pathways, and macropinocytosis [38], [40]. Our previous work indicated that entry of ZEBOV was dependent on signaling through PI3K and Rac1 [25], which are important regulators of macropinocytosis [38]. Work by others also showed that Rho GTPases play a role in ZEBOV uptake [23]. Each of these cellular signaling proteins are known to be important in macropinocytosis. Macropinocytosis is also distinguished from the other pathways principally by criteria that include actin-dependent structural changes in the plasma membrane, regulation by PI3K, PKC, Rho family GTPases [9], [13], [33], Na+/H+ exchangers, Pak1, actin, actin regulatory factors, involvement of CtBP/BARS [14], [38] as well as ligand-induced upregulation of fluid phase uptake and colocalization of the internalized ligand with fluid phase markers [14], [36], [37]. In our examination of ZEBOV entry mechanism, we found that EIPA, a potent and specific inhibitor of the Na+/H+ exchanger [34], [35], [36], [37] blocked ZEBOV infection and entry. Furthermore, ZEBOV caused significant induction of dextran uptake (a fluid phase marker) and the internalized virus particles colocalized with dextran. Pak1 regulates macropinocytosis by promoting actin remodeling and macropinosome closure through phosphorylation of proteins LIMK and CtBP/BARS, respectively [9], [16]. We found that suppression of both Pak1 and CtBP/BARS activity by siRNA or expression of a DN form of Pak1 reduced virus entry and infection.
Actin plays a central role in formation and trafficking of macropinosomes. Actin remodeling is a key event during macropinocytosis and is often triggered by stimuli that promote macropinocytosis. Arp2, among other actin regulatory proteins, has been implicated in macropinocytosis [9]. Arp2 also plays an important role in actin remodeling. It is an integral component of a large multi-protein complex that forms in response to stimuli that trigger actin assembly, and serves as a nucleation site for assembly of actin monomers to form F-actin [41]. We observed a significant increase in the size of the Arp2-containing complexes shortly after ZEBOV binding to cells, indicating stimulation of actin nucleation by the virus. The increase in Arp2 nucleation paralleled an increase in large dextran containing vesicles inside cells corresponding to macropinosomes. This activity appears to be associated with the ZEBOV glycoprotein as VLPs were also capable of inducing a similar increase in Arp2 nucleation and dextran uptake (data not shown). Additionally, we found marked association of fluorescently-labeled ZEBO-VLPs with F-actin foci, as well as with the Arp2-containing complexes and actin-regulatory protein, VASP, that resides in membrane ruffles and promotes actin foci formation. Together, these data provide evidence for a role of actin in ZEBOV entry and suggest that the virus can actively promote localized actin remodeling to facilitate its uptake through macropinocytosis or a similar mechanism.
Despite using multiple approaches, we found no evidence for a role of dynamin in ZEBOV entry. Dynamin is a large GTPase that is involved in scission of newly-formed endocytic vesicles at the plasma membrane [42], [43]. Dynamin-independent entry of ZEBOV further ruled out roles for clathrin or caveolae-mediated pathways as both require dynamin activity [38], [44], [45]. In contrast, the majority of studies suggest that macropinocytosis is independent of dynamin [9]. Recently a novel mechanism has been described for scission of shigatoxin-containing vesicles in which Arp2-dependent actin-triggered membrane reorganization directly leads to vesicle severance [46]. As indicated above, we observed a marked increase in the size of Arp2 complexes shortly after incubation with ZEBOV and a significant association of ZEBO-VLPs with these and F-actin foci but it is unclear if this resulted in membrane scission. In addition, several reports have indicated that C-terminal binding protein (CtBP/BARS), originally identified as a nuclear transcription factor, likely replaces dynamin in scission of nascent macropinosome from the plasma membrane [15], [16]. As discussed above, suppression of CtBP/BARS by siRNA reduced infection and is consistent with the requirement for macropinocytosis in ZEBOV infection.
Interestingly, ZEBOV VP40-based VLPs bearing VSV envelope glycoprotein were found to enter cells through clathrin-mediated endocytosis, as has been reported for the wild-type virus [27]. This suggested that the choice of the internalization pathway is primarily determined by envelope glycoprotein specificity. This is in contrast to a study in influenza virus, where profound differences were seen in the entry characteristics of early passage filamentous virus compared to the laboratory grown spherical isolates that tend to use clathrin-mediated endocytosis [47]. These data indicated a more pronounced role of virion morphology on the choice of endocytic pathway. The apparent reason for this discrepancy is not clear but may relate to the differences in biological characteristics of the viruses and/or cell types used in the two studies.
Overall, our data provide strong evidence that in HEK293T and Vero cells infection by ZEBOV occurs by a process that is closely related to macropinocytosis. We cannot say that entry occurs exclusively by this pathway, but that its disruption blocks the majority of infection and particle uptake. Our work also indicates that clathrin and/or caveolar endocytosis play at most, only a minor role in infection by wild type virus.
A few other viruses as well as bacteria, require macropinocytosis to establish infection [14]. Each uses different mechanisms to induce macropinocytosis. Vaccinia virus has been shown to trigger macropinocytosis by mimicking apoptotic bodies [36]. In contrast, Coxsackie virus and adenovirus activate macropinocytosis by binding to the cell surface proteins occludin and integrin αV, respectively [48], [49]. The mechanism by which ZEBOV triggers macropinocytosis is currently unknown but likely involves GP interaction with cell receptors. Axl (a receptor tyrosine kinase) and integrin βI have been suggested to act as virus receptors [50], [51]. Although, the role of Axl or integrin βI has not been studied in the context of macropinocytosis, there is evidence that several other receptor tyrosine kinases and integrins can trigger macropinocytosis [52], [53], [54]. Therefore, it will be important to analyze the role of Axl and/or integrin βI in this context.
After formation, macropinosomes traffic further into the cytoplasm and may acquire new markers and/or undergo heterotypic fusion with other vesicles of the classical endolysosomal pathway thereby successively transferring the cargo to more acidic compartments such as early and late endosomes [38], [55]. Consistent with this, we found that ZEBO-VLPs co-localized with EEA1-positive vesicles soon after binding [25]. Interestingly, the timing of colocalization of VLPs with EEA1 positive vesicles coincided with their appearance in dextran-containing macropinosomes (within 10 min after binding). Possible explanations may be that the macropinosomes acquire EEA1 shortly after formation or that they undergo prompt fusion with EEA1 positive vesicles. Our data also provided evidence that ZEBOV infection and entry was dependent on Rab5 and Rab7 function, indicating the involvement of early as well as late endosomes in ZEBOV uptake and infection. While a role of early endosomes in Ebola virus entry has not been previously reported, our finding that ZEBOV is trafficked to late endosomes is consistent with prior studies that showed inhibition of Ebola pseudovirion infection by dominant-negative Rab7 [56] and proteolytic processing of Ebola GP1 by late endosome-resident cathepsins [6], [7]. However, it is important to note that many distinct endocytic vesicles associate with Rab5 and Rab7 during maturation but differ by the ligands they carry [57]. This explains why transferrin, a marker of CME, was never seen associated with ZEBOV containing vesicles, even though both require Rab5 and Rab7 for endocytosis.
The intracellular trafficking of the macropinosome is not well understood and existing data provide evidence both for and against the involvement of classical endolysosomal pathway [58]. However, little mechanistic information is available with respect to virus entry by macropinocytosis. Prior to our work only one study analyzed trafficking in any detail, using vaccinia virus and found that virus particles did not colocalize with markers of classical endolysosomal pathway [59]. This difference is likely due to the fact that ZEBOV requires transport to an acidic compartment for membrane fusion while vaccinia virus, which is a pH-independent virus, may undergo nucleocapsid release prior to fusion of macropinosomes with more acidic compartments of the endolysosomal pathway. Our findings now add novel and valuable information regarding macropinosome trafficking mechanism in general and in the context of virus entry.
In conclusion, the evidence presented here demonstrates that ZEBOV utilizes a macropinocytosis-like pathway as the primary means of entry into HEK293T and Vero cells. Once taken up by endocytosis, virus trafficking occurs through early and then late endosomes; however, the exact site where envelope fusion and nucleocapsid release occur is unknown. We do not know if ZEBOV and other filoviruses follow the same pathway into other cell types, like macrophages, that are thought to be a primary target for infection. However, most cell types are capable of macropinocytosis and it is likely that the same or a similar pathway will be used. These findings are important as they not only provided a detailed understanding of ZEBOV entry mechanism, but also identified novel cellular factors that may provide new potential targets for therapies against this virus. It will be important to determine if other filoviruses share the same pathway. If so, it may be possible to develop broad-spectrum therapies that temporarily block this pathway in cells.
Human Embryonic Kidney HEK293T and Vero cells were maintained in Dulbecco's modified Eagle's (DMEM) medium supplemented with 10% fetal bovine serum (Gemini Bioproducts, GA), 1% non-essential amino acids (Sigma, MO) and 1% penicillin-streptomycin solution (Sigma, MO).
All pharmacological inhibitors were purchased from Calbiochem (San Diego, CA) or Sigma (St. Louis, MO). Stock solutions were prepared either in water, DMSO, or methanol, as per manufacturer's recommendation, and stored at −80°C in small aliquots. Alexafluor-labeled reagents including cholera toxin B subunit, transferrin, dextran (10,000 MW) and secondary antibodies were from Invitrogen (Eugene, OR). Specific antibodies against clathrin light chain, caveolin, dynamin, cholera toxin B, Arp2, CtBP/BARS, phospho-VASP and Pak1 were purchased from Santa Cruz Biotechnology, Inc. (Santa Cruz, CA) or Cell Signaling Technology (Beverly, MA). siRNA were from Qiagen (Valencia, CA) and transfections were performed using Dharmafect transfection reagent 1 according to the manufacturer's (Dharmacon, Lafayette, CO) instructions. 4–6 pmol of siRNA were used per transfection of cells in 0.1 ml of medium per well of a 96-well plate. Assays were performed 48 h after transfection.
All plasmids were prepared using Qiagen kits or by CsCl gradient centrifugation following standard procedures. The plasmid encoding VSV-G envelope glycoprotein (pLP-VSVG) was purchased from Invitrogen. Construction of the plasmid encoding Nef-luciferase fusion protein (pCDNA3-nef-luc) has been described previously [60]. Plasmids encoding ZEBOV matrix protein (VP40), ZEBOV envelope glycoproteins were kindly provided by Christopher Basler (Mount Sinai School of Medicine), Paul Bates (University of Pennsylvania), and Luis Mayorga (Universidad Nacional de Cuyo, Argentina) respectively. Plasmids expressing dominant-negative Eps15, caveolin-1, and dynamin-2 K44A have been described previously [19]. The Pak1 expression plasmids were obtained from Addgene (Cambridge, MA).
ZEBOV-VLPs were produced by co-transfecting HEK293T cells with plasmids encoding ZEBOV matrix (VP40) protein, ZEBOV envelope glycoproteins, and Nef-luciferase fusion protein using the calcium phosphate method. For VSV-VLP, plasmid encoding ZEBOV glycoproteins was replaced with one encoding VSV-G. Cell culture supernatant was collected 48 h after transfection and cell debris was cleared by centrifugation (1,200 rpm for 10 min at 4°C). Subsequently, VLPs were purified by centrifugation (25,000 rpm in SW28 rotor for 3.5 h at 4°C) through a 20% (w/v) sucrose cushion in PBS. The VLP pellet was resuspended in 0.01 volume of DMEM, aliquoted and stored at 4°C. Assays were performed within 2–3 days after purification of VLPs.
HEK293T cells were used for contents mixing assays to measure nucleocapsid release into the cell cytoplasm. The cells were removed from plates by trypsin treatment, pelleted by centrifugation and then resuspended in fresh medium. Cells (106 per assay point) were mixed with nef-luciferase containing VLPs in a volume of 0.2 ml and incubated at 37°C on a rotating platform for 3 h. Subsequently, the cells were washed 2–3 times with DMEM to remove the unbound VLPs and the final cell pellet was resuspended in 0.1 ml of luciferase assay buffer lacking detergent (Promega, WI). Luciferase activity was then measured using a Turner Design TD 20/20 luminometer and expressed as counts/sec.
To study drug activity on virus entry, cells were pre-treated with drug for 1 h, followed by incubation with VLPs in the continued presence of the drug. Virus entry was then measured as described above. For measuring the effect of ectopic gene expression, cells were transfected with the control plasmid or one encoding the protein of interest. Cells were then used for entry assays 36 h after transfection. Typical transfection efficiency was 50–70%.
Wild type ZEBOV (Mayinga strain) was provided by Michael Holbrook (UTMB, TX) and the recombinant virus encoding GFP (gfpZEBOV) was from Heinz Feldman (NIH, Rocky Mountain Laboratory, MT). The virus was cultivated on Vero-E6 cells by infection at an MOI of approximately 0.1. All infected cells expressed GFP approximately 24 h post-infection. Culture supernatants were collected after 7 d and clarified by centrifugation at 2000 x g for 15 min. Virus titer was determined by serial dilution on Vero-E6 cells. Cells were incubated with virus for 1 h and then overlaid with 0.8% tragacanth gum in culture medium. 10 d post-infection cells were fixed with formalin, and stained with crystal violet 10 d post-infection for plaque counting. All experiments with ZEBOV were performed under biosafety level 4 conditions in the Robert E. Shope BSL-4 Laboratory at UTMB.
Cells were pre-treated with inhibitors for 1 h and then incubated with gfpZEBOV at 37°C for 2 h (except in the case of MBCD and nystatin, where cells were washed to remove the inhibitors prior to incubation with the virus). Subsequently, the unbound virus particles were removed by washing with PBS, and cells incubated in fresh growth medium. Twenty-four h later, cells were washed and fixed with 10% formalin for 48 h. Images were taken by epifluorescence microscopy and infected foci counted. Counting was performed using the Cell Profiler software package [61]. The processing pipeline used by the software is available upon request.
HEK293T or Vero-E6 cells were cultivated overnight on chambered coverglass slides (Nunc, Rochester, NY) at a density of 50%. The following day, cells were incubated with GFP-tagged ZEBO-VLPs. Cells were then washed three times in DMEM and fixed in 4% fresh paraformaldehyde in PBS. After one wash in PBS residual paraformaldehyde was neutralized by addition of 0.1 M glycine buffer, pH 7.4 and cells were permeabilized using 0.1% Triton X-100 for 1 min at room temperature. For immunofluorescence, cells were incubated with the appropriate primary antibody, typically diluted 1∶200 in PBS. After washing in PBS, the cells were then incubated with the indicated secondary Alexafluor conjugated secondary antibody. Cells were imaged using a Nikon TE Eclipse inverted microscope with a 100× oil immersion lens or a Zeiss LSM 510 confocal microscope in the UTMB optical imaging core.
Cells were transfected with plasmids encoding GFP-tagged forms of the protein of interest. For work with Pak1, myc-tagged or GST-tagged expression constructs were also used. After 24 h, the cells were challenged with wild type ZEBOV. After an additional 48 h, the cells were fixed in formalin. Cells were then stained for ZEBOV VP40 using a rabbit polyclonal antiserum (Ricardo Carrion, Southwest Foundation for Biomedical Research, San Antonio, TX) followed by an Alexa633 secondary antibody. For Pak1 work, cells were also stained for myc-tag or GST-tag using the corresponding primary and a fluorescently labeled secondary antibody. Cell nuclei were also stained using DAPI (Invitrogen). Images were taken using an epifluorescence microscope and the intensity of GFP fluorescence and VP40 staining was evaluated on a cell per cell basis using the Cell profiler software package [61]. For this, cells were first identified by DAPI staining of the cell nuclei. Then cytoplasmic fluorescence intensity for GFP and VP40 staining was determined. The algorithm pipeline used for this part of the analysis is available from R. Davey upon request. The output data, which gives intensities on a scale of 0 to 1, was converted to a scale of 0–1024 using Excel. This data file was then converted to a text file and processed using A2FCS software, which is part of the MFI/FCS Verification Suite (Purdue University) and is available at http://www.cyto.purdue.edu/flowcyt/software/Catalog.htm. This conversion makes the data accessible to conventional FACS analysis software. The data were then analyzed using FlowJo v7.5 (http://www.flowjo.com). Gates were set to exclude cells that were not infected and not expressing the tagged protein, as determined in control experiments. These were done for normal cells infected by virus but not stained, cells not infected by virus but stained with VP40 specific antibody and cells expressing the tagged protein and not infected (stained with antibody against the tagged protein when used). To quantitate the infection dependency of ZEBOV on expression of each construct, the proportion of cells that were expressing each tagged protein construct and infected by ZEBOV was calculated as a fraction of the total cell population. While not used here, this analytical approach can be extended further by setting gates for low, moderate and high levels of ectopic gene expression and then correlating the outcome on infection.
To measure the effect of DN-Cav1 on 10A1 MLV infection, HEK293 cells were transfected with plasmid encoding GFP or GFP tagged DN-Cav1 protein. Thirty six hours after transfection cells were infected with 10A1 MLV pseudotype encoding truncated CD4 receptor (Miltenyi, Germany) as a marker for infection. 36 h after infection the cells were stained for CD4 expression with PE-labeled mouse anti-human CD4 antibody (BD Pharmingen Cat#555347). After 1 h the cells were washed in PBS and fixed in 4% paraformaldehyde. Cells were stained with DAPI to identify cell nuclei and were imaged by a Nikon TE eclipse microscope with an automated motorized stage. To analyze the effect of DN-Cav1 on infection, images were analyzed using Cell Profiler software (Broad Institute, Cambridge, MA) to detect total cells, cells expressing the expression construct and those infected by detection of CD4. Analysis was then performed as above.
|
10.1371/journal.pcbi.1006458 | Assessment of mutation probabilities of KRAS G12 missense mutants and their long-timescale dynamics by atomistic molecular simulations and Markov state modeling | A mutated KRAS protein is frequently observed in human cancers. Traditionally, the oncogenic properties of KRAS missense mutants at position 12 (G12X) have been considered as equal. Here, by assessing the probabilities of occurrence of all KRAS G12X mutations and KRAS dynamics we show that this assumption does not hold true. Instead, our findings revealed an outstanding mutational bias. We conducted a thorough mutational analysis of KRAS G12X mutations and assessed to what extent the observed mutation frequencies follow a random distribution. Unique tissue-specific frequencies are displayed with specific mutations, especially with G12R, which cannot be explained by random probabilities. To clarify the underlying causes for the nonrandom probabilities, we conducted extensive atomistic molecular dynamics simulations (170 μs) to study the differences of G12X mutations on a molecular level. The simulations revealed an allosteric hydrophobic signaling network in KRAS, and that protein dynamics is altered among the G12X mutants and as such differs from the wild-type and is mutation-specific. The shift in long-timescale conformational dynamics was confirmed with Markov state modeling. A G12X mutation was found to modify KRAS dynamics in an allosteric way, which is especially manifested in the switch regions that are responsible for the effector protein binding. The findings provide a basis to understand better the oncogenic properties of KRAS G12X mutants and the consequences of the observed nonrandom frequencies of specific G12X mutations.
| The oncogene KRAS is frequently mutated in various cancers. When the amino acid glycine 12 is mutated, KRAS protein acquires oncogenic properties that result in tumor cell-growth and cancer progression. These mutations prevail especially in the pancreatic ductal adenocarcinoma, which is a cancer with an exceptionally dismal prognosis. To date, there is a limited understanding of the different mutations at the position 12, also regarding whether the different mutations would have different consequences. These discrepancies could have major implications for the future drug therapies targeting KRAS mutant harboring tumors. In this study, we made a critical assessment of the observed frequency of KRAS G12X mutations and the underlying causes for these frequencies. We also assessed KRAS G12X mutant discrepancies on an atomistic level by utilizing state-of-the-art molecular dynamics simulations. We found that the dynamics of the mutants does not only differ from the wild-type protein, but there is also a profound difference among the different mutants. These results emphasize that the different KRAS G12X mutations are not equal, and thereby they suggest that the future research related to mutant KRAS biology should account for these observations.
| The small GTPase protein KRAS is a signal-transducing protein, which binds GDP in its inactive state and GTP in its active state [1]. The gene KRAS is frequently mutated in various human cancers. The mutation is most often, in about 86% of the cases [2], found at G12. In fact, every missense mutation at G12 (G12X) is oncogenic. The oncogenic properties associated with KRAS G12X mutation are characterized by the deficiency of the intrinsic GTPase activity and the insensitivity for GTPase-activating proteins (GAPs) [3,4]. These alterations lead to increased KRAS signaling, as there is more active GTP-bound protein present. Still, the mutant KRAS undergoes GDP–GTP cycling [5]. The basis of the specific G12X mutation frequencies has remained unclear, except for the G12C transversion mutation (c.34G>T) associated with smoking in lung cancer [6,7].
An interesting discrepancy among KRAS G12X mutants is observed in their intrinsic GTPase activity [8]. The G12A mutation exhibits the most hindered intrinsic hydrolysis (~1% compared to the wild-type), whereas the G12C mutation displays the least hindered activity (~72%). All G12X mutants, however, show insensitivity to GAPs that accelerate hydrolysis [8]. Importantly, not only RAS G12X mutants exhibit a discrepancy in GTP hydrolysis, but they also give rise to differences in the preferred signaling pathway (in terms of effector protein binding) [9,10]. This behavior was first observed in NSCLC cell lines [9], where KRAS G12D showed activation of PI3K and MEK signaling, while G12C and G12V mutants exhibited activated RalGDS-pathway and diminished growth factor-dependent Akt activation. Furthermore, an NMR study revealed different binding preferences for mutant HRAS G12V compared to wild-type HRAS, with various effector proteins [10]. Here, HRAS G12V showed reduced interactions with Raf and enhanced binding with RalGDS. However, given that the non-hydrolysable GTP-analog GNP was used in the study, the difference is not due to impaired hydrolysis. Similarly with HRAS, KRAS G12X mutants exhibit reduced affinity to Raf compared to wild-type [8]. The G12D, G12R, and G12V mutants display highly reduced affinity to Raf, while the affinity of G12A is only moderately reduced. Interestingly, the affinity of the G12C mutant is similar to that of wild-type.
To bind RAS, the effector proteins use a ubiquitin (UB)-like fold: a RAS-binding domain (RBD) or a RAS-association domain (RA) [11,12]. While KRAS has not been co-crystallized with any of its effector proteins, distinct effector proteins have been resolved in complex with HRAS: RalGDS (PDB ID: 4G0N) [13], Raf-1 (PDB ID: 1LFD) [14], PI3Kγ (PDB ID: 1HE8) [15], PLCε (PDB ID: 2CL5) [16], RASSF5 (PDB ID: 3DDC) [17] and AF-6 (PDB ID: 6AMB) [18]. These effector proteins bind to HRAS on top of its switch regions: switch-I (residues 30–40) and switch-II (residues 58–72), and the binding conformation of HRAS is almost identical in all of the complexes (S1A Fig). Given this, and since the G12X mutation is far from the binding interface (S1B Fig), Smith and Ikura [10] proposed that the discrepancies in the effector protein binding profiles of the mutants are due to altered switch dynamics. Overall, switch-I displays highly dynamic characteristics manifested as two different states when GTP is bound to RAS, and the distribution between these states is altered in mutants [19–22]. Given that the switch regions in HRAS and KRAS are identical (S1C Fig), their expected binding mode to their effectors is alike. A model of KRAS in complex with A-Raf-RBD tethered to a lipid-bilayer nanodisc suggested by NMR data agrees with this binding mode [23]. At the cellular level, the isoform specificity to effector proteins is primarily determined via membrane interactions [24], but the differences among RAS isoforms’ absolute effector protein binding affinities rise from allosteric effects [25]. It was observed that even a single point mutation in RAS (Q61L) has long-range effects on dynamics and alters effector protein interactions [13].
Previous molecular dynamics (MD) simulation studies of KRAS at microsecond timescales have mainly focused on the dynamical differences between the three wild-type RAS isoforms (HRAS, KRAS, NRAS) [26], differences among selected KRAS and HRAS mutants [27,28], the role of the hypervariable region (HVR) [29], KRAS’s membrane association or orientation [30–32], and KRAS oligomerization on the membrane [33]. The total simulation times of these studies were in the range of 1–8 μs, which is reasonable but likely not sufficient to unravel long-time dynamics associated with slow conformational changes. More importantly, there is a lack of comprehensive atomistic MD simulations of all KRAS G12X mutants with extensive simulation times, allowing a reliable analysis for the differences in structure and dynamical behavior between the wild-type and the mutants, especially in the effector protein binding interface.
What is the underlying cause for the broad range of different G12X mutations? How do these distinctly different mutations manifest themselves in the structure, dynamics, and function of KRAS? This knowledge is crucial to understand KRAS oncogenesis and to develop future therapies targeting mutant KRAS harboring tumors. Therefore, in the present study we first assessed to what extent G12X mutation frequencies are explained by mutation probability. Intriguingly, an outstanding mutational bias emerged from the data. We next employed state-of-the-art atomistic MD simulations (total simulation time 170 μs) to study the dynamical behavior of KRAS with its natural ligands (GDP, GTP) bound, both in the wild-type KRAS and with all existing oncogenic G12X mutations. The results provided compelling evidence that mutations alter the dynamics of KRAS, that the alteration is mutation specific, displays allosteric characteristics, and that the alteration is manifested especially in the effector protein-binding interface. Furthermore, our data suggest that the observed mutational bias and the oncogenic properties of the individual KRAS G12X mutants are caused, at least in part, by mutation-specific altered dynamics.
First, to perceive up-to-date data of KRAS G12X missense mutation frequencies, we compiled data from the Catalogue of Somatic Mutations in Cancer [2]. A total of 32,654 tumor samples identified with a KRAS G12X missense mutation were found from the database. For our analysis, we included only tissues that exhibited these mutations >10%. This status is displayed in eight tissue types, which in total comprised 31,251 positive samples (95% of all KRAS G12X mutations in the database). The large intestine (18,174), the lung (5,640), and the pancreas (5,528) were observed to have numerous positive samples, whereas the biliary tract, the endometrium, the ovary, the peritoneum, and the small intestine comprised altogether only 2,085 positive samples.
A point mutation in KRAS G12X may result in one of six possible missense mutations (Fig 1J). However, instead of being evenly distributed, these specific mutations display considerable variation (Fig 1A). Overall, G12D (42%), G12V (28%), and G12C (14%) mutations are very common, whereas G12A, G12R, and G12S are less popular. When the relative fractions of these mutations are considered in different tissues, they are readily observed to vary significantly (Fig 1B–1I) [2]. For instance, the G12R mutation is observed in the pancreas with a probability of 13%, while in the small intestine it appears in less than 2% of the cases. The predominating mutations are G12D and G12V, the lung being an exception with G12C standing as the most abundant mutation.
In a G12X missense mutation, the guanine (G) base in c.34G or c.35G is substituted to adenine (A), cytosine (C), or thymine (T) (Fig 1J). This base-substitution type exhibits variation (Fig 1K). G>A and G>T mutations (47.4% and 42.1%, respectively) occur very often, while the G>C mutation (10.5%) takes place more seldom. These occurrences display some variation in different tissues. Particularly the lung differs from other tissues with a higher G>T fraction and a diminished proportion of G>A mutations (P < 0.001). Meanwhile, in the pancreas the probability of the G>C mutation is increased (P < 0.001). Moreover, as all of the G12X mutations occur in the first or the second guanine of the codon (c.34G, c.35G) (Fig 1J), we ascertained if there is a mutational bias between these positions. Interestingly, 76.6% of the G12X mutations are c.35G>X mutations (G12A, G12D, G12V) and only 23.4% are c.34G>X mutations (G12C, G12R, G12S) (Fig 1L). In fact, all tissues, except for the lung (55.3%), display 77–90% of c.35G>X mutations. The positional mutation preference for c.35G>X seems to be the highest with a G>A mutation (>7x), whereas G>C or G>T exhibit nearly twofold preference, 1.75- and 2.03-fold, respectively (Fig 1M). A few exceptions in the c.35G>X preference, however, appear in specific tissues. In the pancreas, the G>C mutation occurs nine times more often in c.34 than in c.35 (Fig 1O). As for the G>T mutations, the lung is the only tissue where c.34 is preferred (>1.5x) (Fig 1P). All tissues, interestingly, exhibit over fivefold c.35 preference in G>A mutations (Fig 1N). Above all, the pancreas (>28x), the peritoneum (>37x) and the ovary (>41x) exhibit the most prominent preference for the G>A mutation in c.35.
We evaluated how random the occurrences of the specific G12X mutations are. To this end, we used the transition:transversion mutation ratio as a figure of merit, and compared this figure to a value of 2.3, which is the ratio for missense mutations observed in large-scale genomic analyses [34,35]. If the mutations would take place randomly, G12D and G12S mutations should be the most abundant mutations as they are transition mutations (S2 Fig). G12D mutation is consistent with this view, as it occurs very often in all tissues. Meanwhile, G12S is not consistent with this behavior at all, as it occurs in tumors, perhaps surprisingly, very rarely. Also, regardless of the tissue type, the G12V mutation is overexpressed compared to values expected based on the assumption of random occurrences. Concluding, the mutations’ probabilities of occurrences are not consistent with a transition:transversion mutation ratio based on a random process.
Since local DNA-sequences have clearly a major influence on the mutation probability, a sequence-dependent basis for the observed non-random mutations may exist. The TGGT sequence lacks a typical hotspot mutation region, such as a CpG site [36]. However, an adjacent GG region is a susceptible site for a mutation [37,38]. The oxidation of guanine by endogenous reactive oxygen species may also result in DNA mutation [39]. Both guanines, the 5’G and the 3’G in a GGT-sequence, are found to act as sites for frequent one-electron oxidation reactions, and they exhibit only a minor difference (0.05 eV) in their vertical ionization potential [37]. The oxidation can further transform guanine to 7,8-dihydro-8-oxoguanine (8oxoG), which promotes especially the G>T transversion mutation [40], and the G>T mutations take place on a regular basis (Fig 1K). Interestingly, studies of DNA-adduct formation by exogenous agents have resulted in adduct formation only in c.34G, and not in c.35G [41,42]. Finally, cigarette smoking promotes G12C mutations exhibited regularly in the lung tissue (Fig 1D and 1K) [7]. Concluding, there are several potential mechanisms able to alter the mutation profile of guanine, thereby leading to the data we discussed above.
To understand how G12X mutations affect KRAS functionality, we conducted a total of 170 μs (85 x 2 μs) atomistic MD simulations of wild-type KRAS and all G12X missense mutants, with GDP and GTP. Each individual system was replicated five to ten times starting from different initial conditions (S1 Table).
In the simulations, we observed no differences in the dynamics of the residue 12 (or in its vicinity), which appeared to be extremely stable. In contrast, the switch regions (switch-I and switch-II) exhibited highly dynamic behavior demonstrated by the root-mean-square fluctuation (RMSF) analysis, which revealed major fluctuations in the protein in these regions (S3 and S4 Figs). Nevertheless, there were no evident differences between the different systems, as generally the individual replicas displayed variation as much as the different systems. Only with GDP, the G12A and the G12S display a different RMSF profile in the switch-I region.
To gain better insight into the protein dynamics, we conducted principal component analysis (PCA) [43] with an objective to find the most significant large-scale motions of KRAS. PCA revealed that the greatest dynamic movements in the protein occur in the switch regions (Fig 2, S5A Fig). Furthermore, the most significant principal components 1–3 (PC1-3, see S5B Fig for contributions) highlight that there are strong differences between GDP-bound and GTP-bound systems. PC1 of the GTP-bound systems displayed movement only in the switch regions, whereas PC1 of the GDP-bound systems exhibited additional movement also in the α3-helix. PC2 of GDP-bound KRAS revealed the movement in the switch regions and also extensive motion in the α3-helix, the hairpin loop between the β2- and β3-sheets, and the P-loop (see S1D Fig). PC2 of GTP-bound KRAS in turn brought out the movements observed in GDP-bound systems, and further also the motion in the α1- and α4-helices, and in the turn near the SAK-motif. These observations indicate that the key to resolve the changes in protein dynamics is the γ-phosphate.
Notably, the α4-helix motion is only observed with GTP bound systems (PC2). This observation is in agreement with the experimental results by Mazhab-Jafari et al. [23]. They observed that the GDP-bound KRAS drives the protein in the “exposed” configuration on the membrane, where the α4-helix is located in close proximity to the membrane (PDB ID: 2MSC). This would indicate that the dynamical stability of the helix is important for this state.
In order to ascertain dissimilarities between the different systems, we next generated score plots for the principal components PC1-3 (Fig 3, S5C and S5D Fig). The results highlight dissimilarities between the wild-type KRAS and the mutants, as well as between GDP- and GTP-bound proteins. Interestingly, in all of these systems, only the G12R and G12S mutants with GDP appear to reside in the closed switch-I conformation, whereas all other systems eluded this conformation. Even more interestingly, both of these mutants evaded this conformation when they were bound to GTP. The fully open conformation of switch-I appears to be more accessible to the systems, especially with the G12D mutant with GDP. Moreover, wild-type in both GDP- and GTP-bound systems seems to have a unique state with a high-scoring value (+3 and +4) in PC1 and a low-scoring value (-2 and -0.5) in PC2. Taken together, the results show for all the mutants that the profile of their large-scale motions differs from wild-type regardless of the bound ligand, and that the profile is also unique to each mutant.
We extended the analysis by carrying out PCA for each system to illustrate the differences in their dynamics (S6–S9 Figs). The individual PCA analyses highlight not just variation in the switch region movements among the systems, but they also show that specific systems display more dynamical behavior in the α3-helix, hairpin loop between the β2- and β3-sheets, the α4-helix, the loop between β5-sheet and α4-helix, and in the SAK-motif (residues 145–147) regions. For example, in GTP-bound systems only the G12A, G12D, G12R, and G12V mutants exhibited movement in the α4-helix in their PC1 or PC2. Interestingly, these are also the systems that exhibit clearly diminished Raf affinity [8]. Also, the G12R mutant with GTP displayed notably reduced movement in the switch-II region in both PC1 and PC2.
Even though there are no additional direct interactions from the mutated side-chains of G12X, a mutation in this position inflicts a change to the dynamics in the distant sites of KRAS that were highlighted by the PCA analysis. To investigate this, and to identify possible interaction network routes in KRAS, we conducted an interaction network analysis [44]. Interestingly, we identified a hydrophobic hub network in KRAS that indeed connects the distant sites in the structure and is able to convey these effects in an allosteric way (Fig 4). Therefore, a change in KRAS dynamics in one of the hydrophobic hubs could traverse through this network even to the distant sites. This hub network is comprised of 11 hubs: V14, M72, F78, L79, F90, I100, V114, A146, A155, F156 and L159. One of these hubs, V14, is located in the P-loop, in the close proximity of G12X.
This hub interaction network is highly distorted in G12A and G12S mutants (S10 and S11 Figs). The distortion in these mutants is further not ligand dependent. For example, in the V14 hub, the G12A and G12S mutants lack the interaction to A81 (<2% vs. wild-type 26.9% and 39.7%, with GDP and GTP, respectively), and also display highly diminished interaction to A11 compared to the wild-type KRAS (S10A Fig). From the V114 hub, these mutants lack interactions to A155 and L79, but instead have a strengthened interaction to I142 and a totally new interaction to L113, which is not displayed by other systems (S11A Fig). From the hub A146, they lack the interactions to A18 and L19 (S11B Fig). From the hub A155, both lack the interactions to V114, L79, and I142, but instead they have an elevated interaction to F156, and the G12A has an additional interaction to V152 (S11C Fig).
In contrast to G12A and G12S, the other mutants (G12C, G12D, G12R, and G12V) seem to follow more closely the wild-type’s interaction patterns. However, selected interactions are shifted even with these mutants, although not that extensively as observed with G12A and G12S. For instance, in the hub M72 located in the switch-II region the interaction patterns are shifted with G12C, G12R, and G12V (S10B Fig). Interestingly, the GTP-bound G12D mutant displays almost identical interactions with the wild-type.
We also noticed that the frequency of the salt-bridge between the residues D154–R161 was altered in different systems (S11F Fig). Both of these residues are located in the α5-helix, in the close proximity of three hydrophobic hubs: A155, F156 and L159 (Fig 4). With the wild-type KRAS this salt-bridge is more stable with GDP (69.1%) than with GTP (46.4%). Meanwhile, again with the G12A and G12S mutants this salt-bridge is highly distorted (4.5%–20.6%), regardless of the bound ligand.
As discussed above, the PCA and the interaction network analysis suggest that the protein dynamics is altered among the systems, yet in some obscure manner. To gain better insight into the origin of these differences in wild-type and mutant GTP-bound systems (active KRAS), we analyzed the simulation data by constructing Markov state models (MSMs) [45,46] to explore the long-time statistical conformational dynamics of KRAS. The goal here was to identify the clusters of highly identical protein conformational states, here called metastable states, and to explore how the conformations of the wild-type and the mutant proteins are distributed between these metastable states. For the analysis, we selected the wild-type KRAS together with the most abundant mutants G12D and G12V, and the intriguing G12R mutant, which displays a highly variable distribution in the different tissues (Fig 1).
The MSM analysis identified seven metastable states represented schematically in Fig 5. Overall, all systems populate frequently two of the states: the states VI and VII (77–87%). The less populated metastable states I-IV are specifically represented among the systems. The metastable state I is quite unique for G12R (6%) and the state III for G12V (6%), whereas the other systems are mostly absent from these two states. The metastable state IV is only present in wild-type (3%) and in the G12V mutant (3%). The moderately populated metastable state V, where switch-II appears in a fixed conformation, is rarely observed with the G12V mutant (1%), whereas it is similarly represented among the other mutants and the wild-type (12–16%). In fact, the switch-II conformation appears to be closed in the effector protein complexes (S1A Fig). However, none of the observed metastable states corresponds to this specific switch-II binding end-point conformation.
The states can be further divided in four groups based on their switch conformations (Table 1). The states I and V as combined form the first group, where switch-I appears to be in a fully open conformation and switch-II in a fixed conformation. This group is frequently occupied by G12R (24%), whereas it is mostly absent from G12V (1%). The states VII and VI form individually the second and the third groups, in respective order. In the state VII, switch-II is in a fixed conformation and switch-I is more closed compared to the first group. This group is more frequently populated by G12R. In the state VI, switch-I is open and switch-II is in a mixed conformation between the fixed and perpendicular conformations. This state is clearly less populated by G12R compared to the other systems. The fourth group, where the switches appear in a perpendicular conformation, is frequent with G12V (12%), whereas the other mutants rarely visit this state (1%).
Of all the mutants, G12D displays the most similar metastable state population distribution compared to wild-type (Fig 5, Table 1). This is most evident in the most populated states (states IV, VI, and VII), yet G12D also differs from wild-type in the less populated states (I–IV). In contrast, the conformations of G12R are clearly shifted towards the fixed switch-II states, whereas G12V is shifted away from these states towards the perpendicular states. The results suggest that for G12V the shift among the states is due to the mutant’s lipophilic character, which may cause changes in solvent organization preventing specific switch-II conformations. Finally, it is exceptional that while wild-type does not populate the metastable states I and III at all, there are mutants (G12R, G12V) whose population in these states is significant (about 6%). This summarizes the main message: the conformation distribution of KRAS mutants includes conformations not occupied by wild-type, and these conformations are also mutation specific.
Although frequently observed in cancer, not only is the basis for the specific frequencies of KRAS G12X mutations poorly understood [47], but also the effects of these specific mutations on a molecular scale are not clear. To the best of our knowledge, this is the first study to assess KRAS G12X mutation probabilities, and to understand how they are associated with the observed mutation frequencies.
Generally, the mutation frequencies have an explanatory basis. For instance, chemical characteristics of c.35G explain the enrichment of the G12V mutations by oxidation. However, complex mutation distributions are displayed by the tissues, and we conclude that some of the observed frequencies cannot be explained simply by the mutation probability. For example, there is no clear explanation why, on average, a mutation occurs five times more probably in, e.g., c.35G than in c.34G. One plausible explanation is that the 3D-environment in the DNA-sequence may aid the c.35G mutations to evade DNA-repair mechanisms. In fact, the structures of DNA in complex with N-glycosylase/DNA lyase (OGG1), which is a base-excision repair enzyme for 8oxoG, exist in a catalytically active form for 8oxoG that is adjacent to guanine only in the 5’-position in -AGGT- sequences (S12 Fig) [40,48–51]. Correspondingly, this 5’G position in the KRAS sequence (-TGGT-) represents the c.34G position, thus suggesting that the c.34G position is more susceptible for DNA-repair. Nevertheless, this observation holds true only for the G>T transversion mutation. For the other mutations and their repair mechanisms, the positional bias needs to be investigated, especially for the G>A transition mutation, which holds the strongest bias in favor of c.35G mutations (>5x in all tissues). Furthermore, exceptions or an enhanced preference in c.35G for specific mutations in particular tissues were observed. For instance, it seems that either the advantage for G12D or the disadvantage for G12S, or both, exists in the pancreas, given that there is a 28-fold preference for G>A mutations for c.35G over c.34G. Similarly, the G12R mutation displays an advantage in the pancreas, while G12A is perhaps disadvantageous, given that there is a 9-fold preference for c.34G over c.35G in the mutation probability in the G>C mutations. Altogether, these data suggest that specific mutations are advantageous or disadvantageous depending on the cellular and tissue environments. Therefore, we hypothesize that the biochemical and biophysical differences among mutants, resulting in signal-transducing differences, may explain, at least in part, the observed mutational bias.
To gain insight into these observed discrepancies among the mutants on a molecular level, we carried out a comprehensive all-atom MD simulation study of all KRAS G12X missense mutants. We found that mutations have a profound effect on the dynamics of KRAS. In particular, we observed that the switches are highly dynamic. This conformational flexibility revealed through atomistic simulations is consistent with 31P-NMR spectroscopy studies of RAS proteins [21,52], while the published KRAS crystal structures do not unlock this behavior. Even in our extensive analysis of the long-timescale simulation data the differences in the dynamics were not readily visible. This is not surprising given that even though the binding affinity of a specific mutant toward an effector protein is increased or diminished, the ability to bind still exists [8]. This suggests that the changes in protein dynamics are quite subtle and difficult to quantify. In our work, we unlocked this issue through the analysis of the simulation data using PCA, interaction network analysis and MSMs that indeed revealed the differences, not only between the wild-type and the mutants, but also between the mutants.
In order to capture the subtle differences among the mutants, we kept our MD simulation systems realistic but sufficiently simple, enabling the extended simulation times close to 200 μs in total. Even though the HVR and the cell membrane were absent from our simulations, we recognize that these elements have a substantial influence on KRAS dynamics [53,54] and signaling [55]. We therefore cannot deduce whether some of the observed mutational effects attenuate or amplify through these factors. However, effects related to the cell membrane remain to be explored in follow-up studies.
Importantly, we identified the hydrophobic hub interaction network that is able to convey the shifts in KRAS dynamics throughout the whole structure in an allosteric manner. The crystal structures of KRAS G12X mutants display only minor differences, but the lack of structural differences does not exclude the allosteric effect of the mutation [56]. The shift in the dynamics by G12X is able to occur via the closest hydrophobic hub V14. As the wild-type KRAS has a flexible glycine residue in this position, a G12X mutation alters the dynamics of the neighboring residue A11 or the whole P-loop (including both A11 and V14). As this hydrophobic hub V14 is connected to a hydrophobic network, a local alteration in KRAS dynamics can be conveyed via the hydrophobic network to the other remote structural regions in KRAS in an allosteric manner.
Supporting the fact that V14 is an important hub in the KRAS hydrophobic interaction network, a mutation in this position, V14I, is found to be one of the responsible mutations for the Noonan syndrome [57,58] and may also predispose tumor development [59]. Whereas the V14I mutation does not change the GTPase activity of KRAS, it displays similar affinity to RAF1, as does also the G12V mutant [60]. Therefore, a mutation that has an influence in the dynamics of these hydrophobic hub interactions may have a dramatic influence in overall KRAS dynamics and thereby KRAS signaling.
The most altered interaction pattern within the hydrophobic interaction network in all the hubs is observed with the G12A and G12S mutants. Surprisingly, the other mutations (G12C, G12D, G12R and G12V) are not radically different compared to the wild-type, although some alterations in the hub interactions are evident. However, even though a mutant, such as G12D, displays the same interaction frequencies as the wild-type, the characteristics of the interactions may still differ, as the exact characteristics of these interactions, unfortunately, cannot be derived from this analysis, only their frequencies.
To highlight that the alteration in KRAS dynamics is also present with the mutants that display a minor shift in the hydrophobic hub interaction network compared to the wild-type, is the observed variability in the distribution among the metastable states of exceptional importance. The MSM confirmed the indirect effect of the mutation on the switch-region protein conformations and dynamics. As for MSMs one needs to have extended simulation data, we focused on the most important KRAS G12X mutants (G12D, G12R and G12V). In crystal structures these dynamic metastable states are not observed. This is due to the fact that in the structures the switch regions are disordered, if there are no crystal contacts to the switches. Based on the MSMs, the G12D mutant follows the dynamics of the wild-type more closely than of the G12R and G12V mutants. Intriguingly, this is in line with the findings of the interaction network analysis, where the G12D displayed the most similar profile with the wild-type. In particular, our MD results show that the effects of KRAS G12X mutants are mutation-specific, and suggest that the observed changes in protein conformations and dynamics may alter protein activity [61].
We consider that the difference in the mutant dynamics, for instance the G12V dynamics with its inability to reach the metastable states I and V (Fig 5), may reflect the differences observed in the RAS effector protein binding [8,10]. In fact, simple protein complexes assemble generally via a single pathway [62], and the observed metastable states may correspond to the first steps in the effector protein binding process. These states may be important for specific effector protein binding and pathway activation. However, based on the simulation data we were unable to distinguish if a putative effector protein(s) or a particular signaling pathway(s) is related to a specific metastable state. It needs to be clarified, if these states act as intermediate steps in the KRAS–effector protein association and play a role in the macromolecular recognition process, effecting the association kinetics of the complex formation [63]. Furthermore, multiple other aspects related to the altered dynamics may also affect KRAS mediated signaling. The altered dynamics may cause a conformational change of KRAS on the membrane, resulting in occluded conformation from a specific effector [23], alter the dimer formation [64], or affect KRAS nanoclustering [65]. Furthermore, the altered dynamics may affect the stability of a KRAS–effector complex itself that may lead to a more stable complex, resulting in binding with a longer lifetime, or conversely to a more unstable complex, resulting in faster dissociation. Altogether, this implies that the altered protein dynamics has an influence on the KRAS binding partner selection.
As a crucial factor for KRAS dimerization, the intermolecular D154–R161 salt-bridges between the dimers were recently identified [66]. The KRAS dimerization is a GTP-dependent process [67]. Here we observed that with the GTP-bound wild-type, there is a shift in the dynamics of the putative α4-α5 dimer interface region, manifested by the more unstable intramolecular D154–R161 salt-bridge. This suggests that the more stable intramolecular salt-bridge within the monomer residues could hinder the dimer formation in GDP-bound KRAS, whereas due to the change in the dynamics on the site in GTP-bound KRAS, the more unstable intramolecular salt-bridge could promote the formation of intermolecular salt-bridges among these residues, and thus dimer formation. This intramolecular salt-bridge is again, regardless of the bound ligand, heavily distorted with the G12A and G12S mutants that have a major effect on the hydrophobic interaction networks.
In general, the mutational frequency data combined with the observation from the simulations suggest that at least in the pancreatic cancer, where a KRAS mutation is a key initiator [68], a major shift in KRAS dynamics is not tolerated. This fact is manifested by the low frequencies of G12A and G12S mutants in the pancreas. Moreover, the distorted dynamics could also offer an explanation why the G12S mutant is rarely observed even though it is a transition mutation.
It has been clearly shown that the KRAS G12X mutation has an effect on the intrinsic GTPase activity and that it causes insensitivity for GAPs. However, it seems that the interpretation of the mutation effect on the oncoprotein’s behavior has been oversimplified. First, in specific tissues G12X mutation frequencies exhibit an inexplicable individual bias. Furthermore, the mutation inflicts individual changes in the protein dynamics, affecting the allosteric communication network that conveys the shift in dynamics to the remote sites within KRAS. Finally, this shift in protein dynamics may lead to modulated KRAS mediated signal transduction. We therefore suggest that altered dynamics among KRAS G12X mutants may promote the observed non-random frequencies in specific tissues. In order to establish successful therapies against mutant KRAS-harboring tumors, these discrepancies between the G12X mutants need to be reconsidered thoroughly. Concluding, KRAS G12X mutants are not equal, they are unique.
The simulations were conducted using the GROMACS package v. 4.6 with the (all-atom) OPLS-AA force field [69–72]. For the simulations, a high resolution (1.24 Å) truncated (169/188 residues) GDP-bound wild-type KRAS structure (PDB ID: 4OBE) [73] was selected, where most of the HVR is absent (see S1 Fig). Mutant KRAS structures were generated from the wild-type structure using Maestro [74]. For GTP-systems the GDP was replaced with GTP. As a model for GDP and GTP ligands, the default OPLS-AA parameter set was used and the geometry optimization for GDP and GTP was conducted with Gaussian [75], using the Hartree-Fock method and the 6–31+G** basis set. The partial charges for both ligands were derived from the electrostatic potential by performing a RESP fitting procedure with R.E.D. Tools IV [76,77]. Co-crystallized water molecules from the crystal structure were hold intact, with an exception of GTP-systems with one water molecule, which occupied the γ-phosphate binding position to Mg2+. Water molecules were described with the TIP3P model [78]. In each system, the protein was solvated in a cubic box (edges at least 1.3 nm from the protein), and the system was neutralized using a physiological ion concentration (140 mM) of K+ and Cl- ions. After energy minimization, the system preparation was done in four stages to obtain properly equilibrated and different initial structures for replica simulations (see S1 Table for details). The simulations were performed with periodic boundary conditions in the NpT ensemble. The V-rescale and Parrinello-Rahman methods were used for temperature (310 K) and pressure (1 atm) coupling, respectively [79,80]. The default 2 fs time step was used for integration of equations of motion. To preserve the length of all bonds, the LINCS algorithm was used [81]. For Lennard-Jones interactions and the real-space part of the particle mesh Ewald electrostatics, a cutoff of 1.0 nm was used [82]. Each system was simulated for 2 μs with 5–10 independent replicas, such that the individual system simulation time was 10–20 μs and the total simulation time was 170 μs.
The RMSF were calculated using GROMACS tools [69]. Principal component analysis (PCA) was conducted for the backbone atoms by the GROMACS covariance analysis tools. For PCA, we discarded the first 300 ns and used only the last 1.7 μs of each simulated system, to remove the potential bias of the starting GDP crystal structure from the results. To reduce noise from the flexible terminal regions, we excluded from the analysis the first three residues from the N-terminal and the last five residues from the C-terminal. The PCA structures (Fig 2, S5–S9 Figs) were obtained utilizing the GROMACS tool gmx_anaeig, and visualized with PyMOL 1.8 [83] using a pymol-script Modevectors [84].
The analysis of KRAS interaction networks (the hydrophobic clusters and the salt-bridges) was conducted with PyInteraph [44] and visualized with PyMOL [83]. This analysis was conducted for full trajectories and all residues were included in the analysis.
Markov state model generation was conducted with PyEMMA 2, following the general recommendations [46]. As an input, we used distances between the residues 12, 32, 34, 36, 48, 59, 62, 64, 67, 105, 122, 126, and 138 from the simulation trajectories. We selected these residues based on their functional importance in KRAS (location in the interaction surface with the effector proteins), the results of PCA (dynamical importance), or both. Furthermore, a slow linear subspace from this input was estimated by TICA [85], as TICA highlights the slowest motions from simulations and is highly suitable for generation of a MSM [86], using 40 ns as a lag-time, and two dimensions. Furthermore, the output of TICA was clustered using the k-means clustering, and the discretized trajectories from the clustering analysis were used to generate the BayesianMSM. The number of clusters in k-means were set as √N, as recommended in [46]. The microstates were grouped in seven metastable states by the Perron-cluster cluster analysis (PCCA++) method [87], based on the spectral analysis (S13C Fig). The generated models were validated by two methods. First, we calculated the resulting timescales and found that the timescales were constant in the used 40 ns lag-time (S13–S15 Figs). Furthermore, we conducted the Chapman-Kolmogorov test, which displayed that the model followed the expected estimates. The occupations of individual mutants in each metastable state (Fig 5) were computed from their individual Markov models.
The KRAS G12X mutation data was collected from COSMIC database v.79 (http://cancer.sanger.ac.uk/cosmic/) [2]. In our assessment of the KRAS G12X mutation probability, we included only single nucleotide substitutions. This choice was made based on the fact that more complex mutations and their probabilities (e.g., adjacent double substitutions) are not predictable with the existing knowledge. Fisher’s exact test was used to analyze the specific differences in mutation frequencies.
|
10.1371/journal.pbio.0060158 | Neuronal Correlates of the Set-Size Effect in Monkey Lateral Intraparietal Area | It has long been known that the brain is limited in the amount of sensory information that it can process at any given time. A well-known form of capacity limitation in vision is the set-size effect, whereby the time needed to find a target increases in the presence of distractors. The set-size effect implies that inputs from multiple objects interfere with each other, but the loci and mechanisms of this interference are unknown. Here we show that the set-size effect has a neural correlate in competitive visuo-visual interactions in the lateral intraparietal area, an area related to spatial attention and eye movements. Monkeys performed a covert visual search task in which they discriminated the orientation of a visual target surrounded by distractors. Neurons encoded target location, but responses associated with both target and distractors declined as a function of distractor number (set size). Firing rates associated with the target in the receptive field correlated with reaction time both within and across set sizes. The findings suggest that competitive visuo-visual interactions in areas related to spatial attention contribute to capacity limitations in visual searches.
| It is well known that the brain is limited in the amount of sensory information that it can process at any given time. During an everyday task such as finding an object in a cluttered environment (known as visual search), observers take longer to find a target as the number of distractors increases. This well-known phenomenon implies that inputs from distractors interfere with the brain's ability to perceive the target at some stage or stages of neural processing. However, the loci and mechanisms of this interference are unknown. Visual information is processed in feature-selective areas that encode the physical properties of stimuli and in higher-order areas that convey information about behavioral significance and help direct attention to individual stimuli. Here we studied a higher-order parietal area related to attention and eye movements. We found that parietal neurons selectively track the location of a search target during a difficult visual search task. However, neuronal firing rates decreased as distractors were added to the display, and the decrease in the target-related response correlated with the set-size-related increase in reaction time. This suggests that distractors trigger competitive visuo-visual interactions that limit the brain's ability to find and focus on a task-relevant target.
| It has long been known that the visual system is limited in its ability to process multiple simultaneous inputs. Psychophysical evidence for visual capacity limitations comes from the observation that irrelevant distractors impair the ability to detect or discriminate a task-relevant target. Distractor interference is of several types. Distractors positioned close to the target can impair target visibility (i.e., the ability to detect the target and distinguish its features, generating the phenomena of lateral masking and crowding) [1,2]. The critical separation for crowding—approximately half the retinal eccentricity—well exceeds visual acuity limits, suggesting that this form of interference arises in higher-order visual processing stages [1]. The “set-size effect” is another form of interference that operates over larger distances in search tasks in which the target does not pop out from the display [3–6]. The set-size effect does not affect target discriminability but increases the time needed to find a target in the presence of a large number of distractors.
Although the set-size effect has been widely documented, its neural substrates have remained unknown. In extrastriate cortical areas, blood-oxygen-level-dependent functional MRI activation decreases as more stimuli are added to a display, suggesting that neuronal populations representing individual inputs engage in mutually suppressive (competitive) interactions [7]. Single-neuron recordings have confirmed the presence of competitive interactions in cortical visual areas V2 and V4, and have shown that attention biases the competition in favor of the attended stimulus while suppressing the effect of distractors [8]. However, these studies have not related neuronal competitive interactions to specific forms of visuo-visual interactions.
Because the set-size effect does not impair discriminability per se, it is thought to reflect a form of attentional, rather than visual, interference. Here we tested this idea by examining how set-size affects search-related neural activity in the lateral intraparietal area (LIP), an area important for spatial attention and eye movements [9]. Monkeys performed a covert visual search task in which they discriminated the orientation of a target surrounded by variable numbers of distractors without shifting gaze to the target. As expected, LIP neurons responded more if the target appeared than if a distractor appeared in their receptive field (RF), thus reliably encoding target location. However, firing rates associated with both the target and the distractors decreased with an increasing number of distractors (set size), reflecting the operation of competitive visual interactions. The set-size-related decline in target responses correlated with performance accuracy and reaction time. The findings suggest that the set-size effect is explained, at least in part, by long-range competitive interactions that limit the strength of signals related to spatial attention.
Two monkeys performed a covert visual search task in which they discriminated the orientation of a visual target surrounded by a variable number of distractors (Figure 1). A trial began when monkeys shifted gaze to a fixation point located at the center of a stable circular array of 2, 4, or 6 figure-8 placeholders (left panels). The array was positioned so that, when monkeys achieved central fixation, one placeholder (the “RF stimulus”) entered upon a constant location in the center of the RF of the recorded neuron. After a 500 ms delay two randomly selected line segments were removed from each placeholder, revealing a search display with 2, 4, or 6 unique shapes. One of the shapes, a right- or left-facing letter “E” appearing at an unpredictable location, was the search target while the others were distractors. Without breaking central fixation monkeys reported the orientation of the target by releasing grasp of a bar held in the right or in the left hand. We refer to the fixation epoch prior to presentation of the search display as the presearch epoch and to the interval starting with removal of line segments and ending with the bar release as the reaction time or search epoch.
To examine the effect of set size, we used interleaved trial blocks in which the stable array contained 2, 4, or 6 elements (Figure 1, top, middle, and bottom rows). Increasing set size was associated with higher reaction times and lower accuracy (Figure 2). The set-size effect on reaction times, estimated using linear regression (Materials and Methods section, Equation 1), was significant in 70% of sessions with an average slope of 10.2 ± 1.1 ms/item (Figure 2A; 13.2 ms/item for the significant subset; both p < 0.01 relative to 0). Fitting the population data (Figure 2B) yielded a very similar slope of 10.6 ms/item (confidence interval (CI) [5.8, 15.5]; intercept, 425 ms; regression, p < 10−5; R2 = 0.11). Compared with correct trials, error trials had higher reaction times but a comparable set-size effect (intercept, 459 ms; p < 0.05 relative to correct trials; slope 14.7 ms/item; CI [10.2, 17.8]; regression, p < 0.05; R2 = 0.09). Fitting the accuracy values (Figure 2C) yielded a slope of −2.2%/item (CI [–2.9, −1.5]; intercept, 100.5; regression, p < 10−5; R2 = 0.33). Thus, each additional distractor in the display caused an increase in reaction time of ∼10 ms and a decrease in accuracy of ∼2.2%.
A distinguishing feature of our task is that it required covert attention and a nontargeting motor report (a grasp release) but precluded oriented movement of either eye or limb toward the search target. However, it was possible that even while they maintained central fixation monkeys attempted to shift gaze toward the target. To examine this possibility we measured average eye position in consecutive 100 ms time bins during the search period (0–400 ms after search onset) as well as the end points of the first saccade made within 300 ms after the bar release (when the search array remained on the screen but the fixation point was removed). All eye position measures were uniformly distributed relative to target location (Rayleigh test for directedness of circular distributions, n = 1,710, 3,312, and 4,698 trials for set-sizes 2, 4, and 6; p > 0.6 for all measures). Thus we found no direct evidence that monkeys tended to shift gaze toward the search target during or after a trial.
LIP neurons are known to encode target location during visual search, whether search is accompanied by saccades [10–12] or is performed covertly, as in the present study [13]. Accordingly, the neurons that we describe here had robust target location selectivity during the active phase of search (Figure 3). In addition, their firing rates declined as a function of set size.
Figure 3A shows the responses of a representative neuron, and Figure 3B the average responses of the 50 neurons tested at all set sizes. Responses are segregated according to set size (red, green, and blue traces) and according to whether the target or a distractor was in the RF (solid versus dashed traces). During the presearch epoch (left panel, −200 to 0 ms) the visual array was uniform, and the RF visual stimulation was constant across set sizes. Nevertheless, firing rates declined as set size increased from 2 to 4 to 6. Once the placeholders changed shape (time 0 in left panel), neurons showed a small transient response to the visual offset at ∼50 ms latency (see also Text S1, note 1), followed by a robust signal of target location, whereby responses became much stronger if the target was in the RF than if a distractor was in the RF. Both target and distractor responses were lower at higher set sizes, but this neural set-size effect diminished by the time of the bar release (right panels).
To estimate the magnitude and time course of the set-size effect we fitted firing rates as a linear function of set size using linear regression (Materials and Methods section, Equation 2). We conducted this analysis separately for the presearch epoch (100 ms prior to search onset; Figure 4A and 4B) and in consecutive time bins spanning the search epoch (Figure 4C).
Figure 4A shows the results of the presearch analysis for the example neuron in Figure 3A. Because target location was unpredictable, all trials regardless of target location were pooled for this analysis. The neuron showed a significant set-size effect with a slope of −6.2 spikes/s/item (CI [–7.9, −4.6]; regression, p < 10−11; R2 = 0.31). Across the sample (Figure 4B), 56% of neurons had slopes significantly smaller than 0 with an overall mean of −2.0 ± 0.46 spikes/s/item (−4.0 spikes/s/item for the significant subset; p < 0.0001 relative to 0; n = 50). Thus, neurons showed a decrease in firing rates of ∼2 spikes/s, on average, for each item added to the display. This effect was present from the beginning of fixation (i.e., from the time when the stable placeholder entered into the RF by virtue of the monkeys' eye movements) (Figure S1).
To follow the evolution of the set-size effect during the search epoch we repeated the regression analysis in consecutive 50 ms time bins, this time segregating trials according to whether a target or a distractor appeared in the RF. Figure 4C shows the average slope as a function of time, in data aligned on search onset (left) and bar release (right), for target and distractor trials (circles versus triangles). In the first 200 ms of search the set-size effect remained comparable to that in the presearch epoch (p > 0.2 relative to presearch bins, for each time bin and trial type between 0 and 200 ms after search onset). However, the set-size effect declined markedly thereafter (i.e., slopes increased toward 0), and the average slope became statistically indistinguishable from 0 (open symbols) by 250 ms after search onset for both target and distractor trials (each p > 0.05 relative to 0). When the data were aligned on bar release (right) a small residual set-size effect was seen for distractor but not for target trials (all distractor slopes, p < 0.03; target slopes, p > 0.73 relative to 0). However, no significant differences were found between target and distractor slopes in any time bin (paired t-tests, p > 0.1). A two-way analysis of variance (ANOVA) with bin and trial type (target or distractor in RF) as factors confirmed that there was a highly significant effect of time (p < 10−10) but no effect of trial type or interaction between time bin and trial type (p > 0.1). The fraction of neurons showing significant slopes reached a peak of ∼55% (60% for the target, 50% for the distractors) between 100 and 150 ms after search onset and dropped to 15% (12% for target, 18% for distractor) in the last 50 ms before bar release. Thus, set-size effects were comparable whether a target or a distractor was in the RF and diminished gradually from the presearch epoch to the time of the bar release. A similar pattern was found in the larger subsets of neurons tested at only two of the three set sizes (Figure S2).
Because LIP neurons strongly distinguish between a target and a distractor in the RF, it is important to determine how set size affected neuronal selectivity for target location. We measured target location selectivity using receiver operating characteristic (ROC) analysis, which estimates the probability that an ideal observer can determine whether a target or a distractor is in the RF based on the distribution of firing rates associated with each (see Materials and Methods section). A ROC index of 0.5 indicates no selectivity, while indices above 0.5 indicate preference for the target over distractors in the RF.
The finding that firing rate versus set size slopes were similar for target- and distractor-related responses suggests that increasing set size reduced firing rates uniformly and thus did not change the difference between target and distractor responses. Indeed, as shown in Figure 5, both the time course and the asymptotic (peak) levels of the ROC values were unaffected by set size. The population ROC value (center panel) became significantly greater than 0.5 at a similar time across set sizes (p < 0.05; n = 50; 110–120, 90–100, and 130–140 ms for set sizes 2, 4, and 6). Likewise, the distributions of target discrimination times in individual neurons (see Materials and Methods section) showed no effect of set size nor significant differences between set sizes (one-way ANOVA followed by multiple comparisons; median times of 160, 150, and 150 ms for set-sizes 2, 4, and 6). The asymptotic ROC values (measured between 200 and 300 ms after search onset) were also not affected by set size (one-way ANOVA, p > 0.1). Thus, increasing set size reduced task-related firing rates but did not reflect the magnitude or time course of target–distractor selectivity.
To examine the relationship between the LIP response and performance accuracy we analyzed responses on error trials in which monkeys released the wrong bar (Figure 6). Because of the relatively low error rates few neurons had a sufficient number of trials in all trial categories at each set size. Therefore we turned to a pairwise analysis in which we separately compared correct and error trials in subsets of neurons that contributed a sufficient number of error trials at set-size 6 and set-size 2 (Figure 6A, n = 55 neurons) and at set-size 6 and set-size 4 (Figure 6B, n = 53 neurons). Although a significant effect of set size was present in both correct and error trials (black asterisks indicate p < 0.05 in 100 ms time bins), selectivity for target location (difference between solid and dashed traces) was entirely absent on error trials (colored asterisks indicate p < 0.05 in 100 ms time bins for the corresponding set size). As discussed in relation to Figure 2, manual latencies in error trials were longer than those in correct trials (for the present subsets of data, 2 versus 6, intercept, 440 ms, slope, 13.8 ms/item; 4 versus 6, intercept, 460 ms, slope, 18 ms/item; all p < 0.05 relative to correct trials). However, neuronal target location selectivity was absent up to the time of the bar release (right panels), ruling out the possibility that discrimination may have occurred later on error trials, commensurate with the longer reaction times. These findings suggest that at least some errors reflected failures in target selection, which were associated with a lack of target location selectivity in the LIP.
Two prior studies have reported that saccade reaction times correlated with the onset of significant neuronal discrimination between target and distractor in the RF [12,14]. However, as shown in Figure 5, in our data neither the time of neuronal discrimination between target and distractors nor the asymptotic level of discrimination varied across set size despite clear effects on reaction time. Indeed, we found no significant correlation between the change in ROC onset times and the corresponding change in reaction time across set sizes (2 versus 4, 2 versus 6, and 4 versus 6; all r < 0.02; p > 0.3).
In contrast with the constancy in the ROC signal, however, we found that the firing rates associated with the target itself did reliably correlate with reaction time (see also [15,16]). To examine this correlation within a set size, we separated trials into subgroups in which reaction time was shorter (thick traces) or longer (thin traces) than the median for each cell (Figure 7A). Target-related responses had a stronger and faster rise in trials with short reaction times than those with long reaction times, while distractor-related activity showed a much smaller dependence on reaction time. We confirmed these observations by computing trial-by-trial correlations between firing rates and reaction time as a function of time during the trial. To compute the correlation across the population (Figure 7B) we first normalized firing rates and reaction times by subtracting the average in each neuron's dataset. When the target was in the RF population correlation coefficients became significantly negative (indicating that higher firing rates were associated with shorter reaction times) starting 100–200 ms after search onset. In contrast, distractor-related responses were largely uncorrelated with reaction time except for a trend toward a positive correlation, which reached significance only for set-size 2 late in the trial (200–300 ms). The middle and right panels show the distribution of coefficients for target and distractor responses in individual neurons (computed without normalization) between 200 and 300 ms after search onset. Coefficients for the target response were shifted toward negative values with medians of −0.16, −0.24, and −0.21 for set sizes 2, 4, and 6 (all p < 10−5 relative to 0). Coefficients for distractor responses tended to be positive but had smaller absolute values (0.13, 0.09, 0; p < 0.05 for set-sizes 2 and 4). While these analyses included only trials in which the target was in or opposite the RF, we obtained similar results when we included all distractor trials. Thus, variability in the neural response to the target, but much less in that to the distractor, was correlated with variability in reaction time.
To see to what extent variability in the target response could account for the set-size effect in reaction time, we fitted reaction times as linear functions of target-related firing rates including set size as covariant (Figure 8; see Materials and Methods section). The analysis yielded slopes of −0.51 ms/spikes/s 100–200 ms after search onset, and −0.49 ms/spikes/s 200–300 ms after search onset, showing that reaction times increased by about 1 ms for each 2 spikes/s drop in neural response. Although modest, these slopes were highly significant (each p < 10−20 relative to 0). Correlation coefficients were comparable to those in Figure 7 (100–200 ms, −0.21, −0.17, and −0.15 for set-sizes 2, 4, and 6; 200–300 ms, −0.21, −0.17, and −0.15; all p < 0.05). The analysis of covariance also showed that intercepts differed significantly across set sizes. Intercepts (measured at 0 spikes/s) at set sizes 2, 4, and 6 were −12, 13, and 37 ms for 100–200 ms, and −14, 14, and 35 ms for 200–300 ms (each p < 0.05 for effect of set size). Thus, the analysis revealed two components of the behavioral set-size effect: one, captured by the intercept, was independent of LIP responses while a second, captured by the slope, was significantly correlated with LIP target-evoked responses.
In our task visual stimuli were placed at relatively large distances (medians of 15.0° and 10.7° at set sizes 4 and 6; see Materials and Methods section) that exceeded the distances associated with masking or crowding and may also exceed the span of some of the neurons' RF. We wondered if set-size effects represented interactions only within a RF (i.e., if they arose only in neurons that had more than one stimulus in the RF) or whether they represent interactions beyond the border of the classical RF. To address this question we examined whether set-size effects were related to the RF profile as estimated from the memory-saccade task (see Materials and Methods section).
Figure 9 shows the average normalized visual response on the memory-saccade task at the locations used in set size 4 (Figure 9A) and set size 6 (Figure 9B), aligned to the center of each neuron's RF (0°). Because LIP RFs can be asymmetric we sorted the data so that the two locations flanking the RF center were grouped according to their relative response strength (i.e., flankers associated with the stronger and weaker of the two values were averaged separately). Normalized responses were calculated for each neuron by subtracting the baseline firing rate and dividing by the peak response (always in the center of the RF, 0°). Finally, we segregated neurons according to whether firing rates in the 200 ms before search onset were significantly larger, smaller, or equivalent to those at set-size 2 (set-size 4, n = 11, 34, and 29, respectively; set-size 6, n = 9, 31, and 14).
If competitive interactions were limited to the span of the RF, then neurons with significant set-size effects should have significantly stronger excitatory responses at the stronger flanking location than neurons without a set-size effect. However, this was not the case. Average responses at 90° and 60° flanking locations (set sizes 4 and 6) were not statistically different from baseline (both p > 0.1, one-way ANOVA), showing that for most neurons the nearest flanking stimuli fell outside the visual RF [17]. Moreover, response magnitude at flanking locations did not differ between neurons that did or did not have a set-size effect. We also found no consistent tendency for neurons to show inhibitory surrounds near the borders of the RF, as there was no significant dip below baseline at the weaker of the two flanking locations. Activity at the weaker locations was also not related to the set-size effect (p > 0.1 for set-size effect, one-way ANOVA). These findings show that competitive effects in the LIP are not straightforwardly predicted by inhibitory surrounds near the excitatory RF and can extend beyond the confines of the classical RF.
Previously we have shown that during a similar covert search task LIP responses were modified by the active limb, with some neurons having stronger responses to the target if the monkey released the right bar and others preferring left bar release [13]. In the present dataset, approximately one-third of neurons showed limb effects. However, we found that the magnitude and time course of the set-size effect as well as the correlation with reaction time were equivalent in neurons with and without limb sensitivity and, for the former group, were equivalent for responses with the preferred and nonpreferred limbs (Figures S3 and S4). Thus the set-size effect did not depend on limb selectivity.
A second question is whether the set-size-related decline in activity may have been related to reward probability, which declined together with the monkeys' accuracy [18,19]. To examine this possibility we computed correlation coefficients between session-by-session firing rates and success rate (Figure 10A). We found no correlations either within a set size (Figure 10A; coefficients of −0.09, 0.06, and −0.08 for set sizes 2, 4, and 6; all p > 0.58) or in computing the differences across set sizes (Figure 10B; set size 2 versus 4, r = −0.12, p = 0.06; set size 2 versus 6, r = 0.08, p = 0.55). We also considered the possibility that monkeys estimated reward probability from local sequences of 10–20 trials rather than globally across long trial blocks [19]. However, correlation coefficients between local measures of firing rate and reward probability (measured in sliding windows of 20 trials) were statistically significant in only 2% of neurons, less than the 5% expected by chance. This precludes the possibility that the set-size effects were due to reward expectation.
Because increasing set size also increases the uncertainty of target location, an important question is whether set-size effects reflected location uncertainty (the diminished probability that the target would appear at any one location) rather than the number of stimuli per se [19,20]. While previous studies reported probability effects in saccade-based tasks [19,20], it is not clear whether monkeys use probability information to adjust the distribution of covert attention. To examine this question we trained monkeys in blocks of trials in which set size was constant (n = 6) but target location probability varied between 100% (the target appeared at a single, constant location) and 16.7% (the target appeared with equal probability at each location as in the standard condition). We ran these conditions for 20 sessions for each monkey using long blocks of, on average, 148 trials for each condition (range, 96–358 trials). Despite extensive testing we found no significant differences in either reaction time or response accuracy between the constant and the variable conditions, either in individual sessions or in the pooled data (for reaction time, p > 0.1, n = 6,003 and 5,994 trials in the constant and variable conditions; for accuracy, p > 0.2, n = 40 sessions). We also found no differences between reaction time and accuracy in the early and late portions of the blocks (first 25% of trials, constant location, 464 ± 32 ms, 85 ± 6% correct; variable location, 459 ± 71 ms, 83 ± 6% correct; last 25% of trials, constant location, 455 ± 74 ms, 83 ± 5% correct; variable location, 467 ± 56 ms, 83 ± 6% correct). These results suggest that monkeys did not take note of changes in location probability even when these changes were very large (100% to 16.7%). This makes it unlikely that they detected the more subtle changes in probability among set sizes 2, 4, and 6 (50% versus 25% versus 16.7%).
Figure 10C shows responses in a subset of 10 neurons tested at set-size 2 (red traces, 50% probability) and at set-size 6 in 100% and 16.7% probability conditions (blue dashed and blue solid traces). Responses at set-size 2 were much higher than those at set-size 6 but did not differ between probability conditions within set-size 6. A one-way ANOVA in each 100 ms time bin from 200 ms before to 300 ms after search onset revealed highly significant differences between set-size 2 and set-size 6 and either the 100% or the 16.7 % probability condition (p < 10−6 in each bin) but no significant effect of location probability within set-size 6 (p > 0.2 in each time bin). Thus neural responses, like behavioral performance, were related to the number of display elements independent of variations in target location probability.
The visual system does not process visual inputs as isolated entities but fashions integrated representations in which individual objects interact in multiple ways. Although it is known that visual responses in dorsal and ventral extrastriate areas are shaped by competitive visuo-visual interactions, the precise functional consequences of these interactions are not known [7,8,21]. Here we report that competitive interactions also operate in the LIP, a parietal area associated with attention and eye movements, implying that these interactions limit not only the fidelity of visual feature information but also the efficacy of top-down signals of spatial attention. Moreover, competitive interactions in the LIP correlate with the behaviorally measured set-size effect. We discuss our results in the light of prior studies of the LIP and extrastriate cortex and of the neural mechanisms of visual search.
The LIP has been proposed to encode a priority representation of the visual world, a sparse topographic representation in which only objects that are likely to be attended are strongly represented. Two principal factors are known to activate LIP neurons—the automatic orienting of attention toward a salient but task-irrelevant stimulus [9] and the voluntary selection of a saccade target [11,12,14,15,22–24]. Bisley and Goldberg have shown that presaccadic sustained activity in LIP is related to the deployment of covert attention that precedes an overt saccade [22]. Here we go a step further in linking the LIP with covert attention independently of saccades: in our task neurons were strongly active even though monkeys were explicitly trained to withhold saccades throughout the task. It remains in principle possible that monkeys formed covert plans to make a saccade toward the attended target (although we, like others [25], failed to find direct behavioral evidence for this idea). However, the correlations between LIP activity and performance of the covert search itself strongly implicate this area in covert target selection independent of eye movements. This conclusion is supported by the findings of Wardak et al. that reversible inactivation of the LIP impairs performance on visual search tasks whether these are performed with free or fixed gaze [26,27].
While quantitative models have speculated on the contributions of the LIP to overt saccade decisions [28], accounting for its contributions to covert attention is considerably more challenging. A common working hypothesis is that visual-oculomotor areas such as the LIP and the frontal eye field provide topographic, top-down feedback to feature-selective visual areas including V4, the middle temporal area, and inferior temporal cortex, which boosts visual responses to the attended object [29–31]. The biased competition theory proposes that attentional feedback is especially important in environments containing multiple distractors, where feedback biases neuronal competition in favor of the attended object, allowing neurons representing this object to “win” the competition and filter out the effects of distractors [32].
Our findings are consistent with the idea that LIP plays a specific role in selecting targets and overcoming distractor interference in cluttered visual environments. In our task, as in saccade-based search tasks, neurons selectively encode the location of the search target, and their responses correlate with the efficiency of target selection [11,12,14]. In contrast, responses to nontargets has been shown to correlate with the distracting effects of these objects on performance [22,24,33]. Here we show a novel mechanism of distractor interference: adding distractors to the display suppresses LIP activity through competitive visual interactions, producing a neuronal set-size effect that correlates with the effect of set size on performance. Together with the finding that deficits following LIP inactivation are larger at higher set sizes [26], the findings suggest that the LIP plays a special role in overcoming distractor interference in complex environments.
In light of these considerations, the dissipation of the set-size effect during the reaction time may represent an active process through which the brain suppresses distractor interference. It may be argued that the decline of the set-size effect reflected mere disengagement of the LIP from the task toward the end of the reaction time, as at this time there was a general decline (though not a complete disappearance) in target location selectivity (Figure 3). This possibility is unlikely, however, as high location selectivity was not, in and of itself, necessary for seeing a set-size effect: robust effects were found in the presearch epoch and on error trials, when firing rates were low and there was no location selectivity. Thus, it is more likely that the dissipation of distractor effects reflected an active search-related process. One such process could be selection of the search target. Target-related responses peaked between 100 and 200 ms after search onset, slightly before the time when the set-size effect was filtered out at the population level (200–250 ms). It is therefore possible that the elevated target-related activity suppressed distractor competition, consistent with a biased competition model [8,32]. In addition, feedback about limb motor planning, which reaches the LIP [13], may have helped render responses stereotyped and independent of set size [16].
Because the LIP receives strong input from extrastriate cortical areas, one must consider to what extent the competitive interactions that we report reflect properties of this bottom-up input. However, while competitive effects in extrastriate cortex are based on visual features, those in the LIP are based on spatial location. In areas V2 and V4, the middle temporal area, and the middle superior temporal area, visual competition is triggered when two stimuli are presented in close proximity within an individual RF so that one stimulus has a preferred feature (e.g., orientation or motion direction) while the other is nonpreferred [8,21,34]. These results support a model in which competition (mutual inhibition) arises between neurons with overlapping RFs but different preferred features [8]. In the LIP, in contrast, competitive effects arise among physically identical stimuli (the placeholders during the search epoch) and are triggered even when the competing stimuli are outside the classical RF. Thus competition in the LIP engages neurons that have nonoverlapping RFs but similar (or no) feature selectivity. These considerations suggest that our findings are more closely related with location-based competitive interactions in the superior colliculus and the frontal eye field [35,36] and reflect the internal organization of all three structures in topographic, nonfeature-selective representations. Thus visual clutter appears to affect multiple levels of representation through both space- and feature-based competition.
Increasing set size in our task also increased the uncertainty about target location or, conversely, lowered the probability that the target appears at any given location. However, our data suggest that under the present conditions monkeys did not explicitly compute and represent location probability. In a control condition in which set size was kept constant but target location probability varied between 100% and 16.7%, we found no effect of location probability on either behavior or neural responses. We note, however, that some models represent the effect of distractors as a broadening of the underlying noise distribution (i.e., distractor-related firing rate distribution) [37]. This type of increase in noise, which may be interpreted as an implicit representation of uncertainty, may indeed be an appropriate mathematical description of our data.
The lack of a probability effect in our task may appear puzzling given prior reports that monkeys are sensitive to manipulations of location or reward probability in saccade-based tasks [18–20]. We suggest, however, that this result is explained by the precise task conditions that we used. Whereas previous studies manipulated the probability of an overt, rewarded saccade, here we used location probability to bias attention, a variable that is by definition covert and cannot be directly rewarded (in our task, reward was linked to a manual response). In addition, the visual conditions in our task provided little incentive to guide attention to the likely target location based on probability information. Because the target was suprathreshold and was present until the manual response, monkeys could simply ignore the prior targets and find the target when it appeared with little loss of accuracy. Indeed, probability effects on covert attention were previously found in monkeys by using brief, near-threshold targets [38] but not in a detection task with suprathreshold stimuli [39]. Thus, our findings do not preclude the possibility that LIP neurons reflect location probability in conditions in which probability is computed and used; they show, however, that neurons are strongly affected by competitive visual interactions independently of target location probability.
The target-related activity in the LIP correlated with both performance accuracy and reaction time. Target location selectivity was entirely absent in error trials, suggesting that many errors reflected failures in locating the target (possibly along with failure in target discrimination or response selection).
In addition, target-related firing rates showed trial-by-trial covariation with reaction time so that higher firing rates were associated with shorter reaction times both within and across set sizes. While the correlation coefficients that we find are modest (−0.2 to −0.3 in Figures 7 and 8), existing evidence suggests that such weak correlations are to be expected in cortical association areas. Because correlations between neural activity and motor output increase along the sensory–motor continuum [40], an area such as the LIP, which represents a nonmotor processing stage, would a priori be expected to covary only weakly with trial-by-trial reaction time. The task that we used is complex and is likely to have engaged multiple areas in addition to the LIP, including extrastriate visual areas and areas related to limb motor planning, procedural memory, motivation, and reward evaluation, all of which have high trial-to-trial variability and weak interneuronal correlations [41–43]. In this regard it is remarkable that the correlations that we report are slightly larger than the average correlation coefficient of −0.09 reported in a saccade-based task [44]. Computational models show that reliable information may be extracted from ensembles of as few as 10–100 task-related neurons with highly variable, weakly correlated firing [42,45], suggesting that our findings reflect significant contributions of the LIP to covert search.
A puzzling aspect of our data is the finding that set size lowered firing rates but did not modify neuronal selectivity for target location (the discrimination between target and distractors in the RF) as indexed by the ROC analysis (Figure 5). We found that increasing set size reduced target- and distractor-related activity by similar amounts, leaving the dynamics of target location selectivity constant across set sizes (Figures 2–5). This appears to be at odds with two prior reports that found a consistent relationship between the time of onset of the ROC and saccade reaction time during visual search [12,14]. It should be noted, however, that variations in ROC dynamics in these studies may have been driven primarily by variations in target-related firing rates, as was the case in the within set-size analysis in Figure 7 (see also [16,44]). Thus, while both the ROC signal and the target response itself can covary with reaction times, the latter may show a more consistent relation across task conditions. Resolving this question will require more detailed understanding how activity is read out from the entire LIP map under different task conditions.
Two adult rhesus monkeys (Macaca mulatta) weighing 8–10 kg (one male and one female) were tested with standard behavioral and neurophysiological techniques. All methods were approved by the Animal Care and Use Committees of Columbia University and New York State Psychiatric Institute as complying with the guidelines within the Public Health Service Guide for the Care and Use of Laboratory Animals. Experiments used standard behavioral and physiological procedures described in detail elsewhere [13,24].
Electrode penetrations were aimed at the posterior half of the lateral bank of the intraparietal sulcus as guided by structural MRI. Upon isolation each neuron was first tested with the memory-saccade task on which, after the monkey fixated a central point, a small target annulus (1° diameter) was flashed for 100 ms. After a 1000–1250 ms delay the fixation point was extinguished, and monkeys were rewarded for making a saccade to the remembered location of the target within 100–500 ms. Neural responses were tested at 8–12 locations circularly distributed at a constant eccentricity around fixation, including the location estimated to be the center of the RF (that eliciting the strongest visual response).
The search task was conducted in randomly interleaved blocks of trials. An array containing 2, 4, or 6 figure-8 placeholders remained on the screen from the beginning to the end of a block, including intertrial intervals. A trial (and data collection) began with presentation of a fixation point (a 0.5° red square) at the center of the array. After monkeys grabbed two response bars and maintained central fixation for a 500 ms period, the search display was presented by removing two line segments from each placeholder. The line segments to be removed were selected randomly with the constraints that (1) a single item became the search target (a right- or left-facing letter “E”), (2) each of the remaining shapes was continuously connected, and (3) no shape was presented at more than one location in one trial. The location and orientation of the target were selected randomly with uniform probability and independently of each other. Monkeys were rewarded for continuously maintaining fixation within a window of 2° × 2° and indicating the orientation of the target by releasing the right bar for the right-facing E or the left bar for the left-facing E within 100-1000 ms after removal of the line segments. A correct response was followed by removal of the fixation point and delivery of reward 250 ms later. Incorrect responses (including fixation breaks and early, late, or inaccurate bar releases) were followed by removal of the fixation point without reward delivery. The target array remained on the screen for an additional 500 ms, allowing us to collect eye movement data following bar release. All trials terminated with restoration of the missing line segments, reinstating the placeholder display.
The radius of the placeholder array and its rotation around the center were varied for each training and recording session. During neural recording care was taken that one placeholder fell in the center of the neuron's RF as determined with the memory-guided saccade task. Stimuli were scaled with eccentricity and ranged from 1.5° to 3.0° in height and from 1.0° to 2.0° in width. Median RF eccentricity was 10.6° (range, 2.6–14.3°). Median separations between adjacent array elements at set sizes 2, 4, and 6 were, respectively, 20°, 15.0°, and 10.7° (ranges, 4.0–28.6°, 3.7–20.2°, and 4.6–14.3°, respectively). The ratios between interstimulus distance and eccentricity were 2.0, 1.4, and 1.0 at set sizes 2, 4, and 6, far exceeding the critical ratio of 0.5 that defines the critical distance for crowding [1].
Comparisons across samples were made with t-tests or paired t-tests and one-way and two-way ANOVAs as specified below and evaluated at p < 0.05. Regressions were calculated with weighted least-squares algorithms [46]. The 95% CI of the slope and intercept were calculated and used for statistical testing. In addition, we verified the results of the parametric tests using one-way nonparametric ANOVA (the Kruskal–Wallis test) and two-way nonparametric ANOVA (the Friedman test). In all cases the results were equivalent, and, for simplicity, we adopted the convention of reporting only the outcomes of the parametric statistics in the text. In addition, we used ROC analysis, which is a nonparametric measure of the separation between two firing rate distributions and hence the likelihood that an ideal observer can distinguish between the two [47]. Results are shown as mean ± standard error unless otherwise stated.
Data were collected from 107 neurons that had significant spatial selectivity during both memory-saccade and covert search tasks. However, the bulk of the analysis concentrates on 50 of those neurons (24 in monkey 1) that were tested at all three set sizes. Data from additional subsets tested at set-sizes 2 and 4 (n = 73), set-sizes 2 and 6 (n = 55), and set-sizes 4 and 6 (n = 53) were used for the error trial analysis and in additional analyses shown in Text S1.
Approximately 8% of all trials were discarded from the analysis, as they terminated in fixation breaks or in short- or long-latency bar release. Reaction times were measured as the time from presentation of the search display (measured by means of a light-sensitive diode mounted in the upper left corner of the screen) to the time of the bar release (measured by a transistor–transistor logic pulse emitted upon the onset or termination of contact with the bar). Accuracy was measured as the fraction of correct out of the total number of correct and incorrect bar releases. Reaction times (RTs) were analyzed separately for correct and erroneous responses. To assess sensitivity to set size, reaction times were fit with the regression model
where SS is the set size (2, 4, or 6) and ε is random error distributed as multivariate normal (Figure 2A and 2B). The slope b1 is an estimate of sensitivity to set size in milliseconds per item. This analysis was carried out on a neuron-by-neuron basis, where each data point represented one trial, and across the population, where each data point represents the average RT for a single session. Accuracy (percent correct) was fit using the population data.
All analyses were conducted on raw (unsmoothed) spike counts. Firing rates on the memory-saccade task were measured in the baseline (200 ms before target presentation), visual (50–250 ms after target onset), delay (400–900 ms after target onset), and presaccadic epochs (200 ms before saccade onset). A neuron was tested on the search task only if it showed significant spatial tuning during the visual, delay, or presaccadic epochs (p < 0.05, one-way ANOVA). Nearly all neurons (96%) had significant spatial tuning during the delay period. Visual responses on the memory-saccade task were highly correlated, across locations, with those during the search task (r = 0.94), showing that neurons preserved a constant spatial RF in both tasks.
Several analyses were performed for the covert search task. To measure the sensitivity of firing rates to set size we fit firing rates to the linear model
where FR is the trial-by-trial firing rate in the time bin indicated in the text. The coefficient b1 represents sensitivity to set size in units of spikes per second per item. Fits were obtained separately for trials in which the target or a distractor was in the RF. For the time-course analysis (Figure 4B) our choice of time bin (50 ms nonoverlapping windows) represented the best compromise between the need for temporal resolution and the need to use larger bins for more reliable estimation of firing rates and regressions.
To analyze the relationship between firing rates and reaction time, we first computed the Spearman correlation coefficient between firing rate on a trial-by-trial basis within each neuron (Figure 7B). To compute the coefficient across the population we pooled all trials after normalizing each neuron's data by subtracting the average in the appropriate time bin (Figure 7B, left panel). In a second step (Figure 8), we fit the data using analysis of covariance (ANCOVA), which simultaneously fits separate regression lines of the form
to data from set-sizes 2, 4, and 6. The slope parameter b1 indicates the sensitivity in units of milliseconds per spike per second, whereas any difference in the intercept indicates the component of RT that depends on set size independently of LIP firing rates. The ANCOVA was computed on normalized data pooled across all neurons. Reaction times were normalized by subtracting the average reaction time across all three set sizes. Firing rates were normalized by subtracting the average neuronal response across all time bins, set sizes, and target/distractor trials.
ROC indices comparing selectivity for target versus distractor in the RF were calculated for each neuron in 10 ms bins aligned on search display onset [47]. We found that ROC analysis was relatively robust to firing rate variations resulting from small bin sizes, allowing analysis with higher temporal resolution. Confidence intervals were obtained by a permutation test with 1,000 repetitions, and a value was deemed significant if its 95% confidence interval did not include 0.5. The onset of significant selectivity was defined as the start of the first four consecutive bins with ROC values significantly different than 0.5 [13].
Each of the above analyses was evaluated for each monkey separately. In no instance did we find significant differences between monkeys, and thus the pooled data are presented throughout the paper. Data are also pooled across right and left bar release; analysis of data segregated according to bar release is included in Text S1.
|
10.1371/journal.pcbi.1005825 | Insight into glucocorticoid receptor signalling through interactome model analysis | Glucocorticoid hormones (GCs) are used to treat a variety of diseases because of their potent anti-inflammatory effect and their ability to induce apoptosis in lymphoid malignancies through the glucocorticoid receptor (GR). Despite ongoing research, high glucocorticoid efficacy and widespread usage in medicine, resistance, disease relapse and toxicity remain factors that need addressing. Understanding the mechanisms of glucocorticoid signalling and how resistance may arise is highly important towards improving therapy. To gain insight into this we undertook a systems biology approach, aiming to generate a Boolean model of the glucocorticoid receptor protein interaction network that encapsulates functional relationships between the GR, its target genes or genes that target GR, and the interactions between the genes that interact with the GR. This model named GEB052 consists of 52 nodes representing genes or proteins, the model input (GC) and model outputs (cell death and inflammation), connected by 241 logical interactions of activation or inhibition. 323 changes in the relationships between model constituents following in silico knockouts were uncovered, and steady-state analysis followed by cell-based microarray genome-wide model validation led to an average of 57% correct predictions, which was taken further by assessment of model predictions against patient microarray data. Lastly, semi-quantitative model analysis via microarray data superimposed onto the model with a score flow algorithm has also been performed, which demonstrated significantly higher correct prediction ratios (average of 80%), and the model has been assessed as a predictive clinical tool using published patient microarray data. In summary we present an in silico simulation of the glucocorticoid receptor interaction network, linked to downstream biological processes that can be analysed to uncover relationships between GR and its interactants. Ultimately the model provides a platform for future development both by directing laboratory research and allowing for incorporation of further components, encapsulating more interactions/genes involved in glucocorticoid receptor signalling.
| Here we present modelling of the glucocorticoid receptor (GR) signalling network. The GR is the effector for a class of drugs known as corticosteroids, which are widely used in medicine for their anti-inflammatory effects and ability to induce apoptosis in leukaemic cells. However, side effects, treatment-related toxicity and glucocorticoid resistance remain and therefore increased understanding of the glucocorticoid receptor mechanism of action may improve therapeutic outcomes. The GEB052 model presented herein has been used to generate predictions for how the network is altered between glucocorticoid-sensitive and glucocorticoid-resistant scenarios, and these predictions have been verified using published gene expression data from established cell lines (for both qualitative and semi-quantitative analysis). The model has also been preliminarily assessed as a predictive clinical tool by correlating model predictions with clinical outcomes of thirteen leukaemia patients. Thus, the GEB052 model demonstrates successful modelling to understand GR function. GEB052 provides accurate predictions and has indicated potential routes through which glucocorticoid resistance may arise. The work presented herein thus demonstrates a proof-of-principle of this modelling approach to furthering GR research, and provides insight into potential mechanisms of corticosteroids resistance.
| Glucocorticoids (GCs) steroid hormones released from the adrenal cortex as part of the stress response play an important role in a variety of bodily processes such as inflammation, immunity, and numerous metabolic processes [1–3]. Their varied effects allow for their clinical application in numerous diseases, particularly for their potent anti-inflammatory and immunosuppressive effects to treat diseases such as arthritis [4, 5]. GCs are also prescribed for the treatment of lymphoid cancers, as they selectively induce cell death in leukocytes [6–8], highlighting the tissue specificity of their action and the need for further research into GC signalling.
GCs exert their effects through the glucocorticoid receptor (GR) which is an intracellular cytoplasmic receptor which, in the absence of a ligand, is part of a complex with chaperones such as heat-shock protein 90 [9]. Following ligand binding, GR dissociates from this complex and translocates to the nucleus where it regulates the expression of its target genes as an active transcription factor [10, 11]. Numerous factors control GR activity, including phosphorylation status [12], targeting to protein degradation pathways [13] and interaction with cofactors [14].
Clinically, synthetic GCs such as dexamethasone are used due to their higher potency and stability. Whilst GCs have achieved significant therapeutic outcomes, resistance to treatment and side-effects both remain an issue. Defective GR expression, Bcl-2 overexpression, and other aberrant signalling may contribute to glucocorticoid resistance [6, 15, 16]. Increased knowledge into the details of GR signalling may allow for the development of novel therapeutics and identification of resistance factors. Although high-throughput methodologies have provided insight into GR signalling [6], there remains a need to properly integrate large datasets in a cohesive manner.
Systems biology aims to accurately represent biological phenomena by constructing integrative models of molecular components and their interactions. Some models are quantitatively precise and require measurement of biological kinetic data, though they are often of a smaller scale, aiming to precisely model a particular subset of interactions. This approach has been applied to glucocorticoid research in numerous ways, such as the development of models of GR/c-Jun/Erg (Ets-related gene) crosstalk [17]. The models of GR/c-Jun/Erg confirmed known interaction phenomena but also identified Erg as a putative marker for glucocorticoid resistance [17]. Although such models provide useful insight, they are both time-consuming and resource-expensive to create due to the required biological data. Boolean modelling on the contrary allows for the generation of large-scale models that provide a qualitative overview of the behaviour of an entire network [18, 19]. In these cases, interactions and molecular levels are simplified to ON or OFF binary values, removing the need to know exact rate and kinetic equations thus reducing computational demand [20].
We have previously demonstrated that Boolean modelling may be successfully applied to cancer research through generating the PKT206 model of the p53 interactome [18] which has revealed novel mechanisms of p53 signalling and how this may be disrupted following loss of p53 function. Correct prediction rates reached 71% for the model, signifying the strength of this approach [18]. An expanded p53 interactome was later developed to more accurately model the signalling phenomena [19].
To overcome the qualitative nature of the Boolean modelling approach algorithms utilising microarray and/or ChIP-seq data have been developed such as the signal transduction score flow algorithm (STSFA) which analyses Boolean models in a semi-quantitative manner [21]. This algorithm has been applied to the original PKT206 model [22], which demonstrated improved predictive power over the original model analysis. Thus, application of this or similar algorithms represents a way to improve model accuracy through its semi-quantitative nature.
The aim of this research was to develop a Boolean model for the GR interaction network similar to the p53 interactomes [18, 19]. The model (GEB052: Glucocorticoid receptor model by Emyr Bakker, consisting of 52 nodes) contains 241 interactions. Nodes represent genes/proteins or inputs (glucocorticoid)/outputs (cell death and inflammation). CellNetAnalyzer [23] has been used for in silico analysis. Boolean model performance was assessed via comparison to microarray data [18] which demonstrated up to 60.4% of predictions depending on microarray data used for validation (average 57%) as correct, whilst STSFA analysis indicated a correct prediction rate of 80.1%. Using microarray data from thirteen leukaemia patients the model has been assessed as a predictive clinical tool. This report demonstrates the applicability of this modelling approach to nuclear receptor research, with the overarching aim being to eventually create models in a tissue-type, disease-specific and patient-centred manner.
The GEB052 model was built via a similar workflow to the PKT206 model [18] (Fig 1). STRING (Search Tool for the Retrieval of Interacting Genes/Proteins) is a database that provides information on functional associations between proteins, and thus this database represented a starting point for the interactions to be included within the model [24, 25]. Nodes within this model represent genes (or their associated proteins) and inputs/outputs such as the glucocorticoid and cell death and inflammation respectively. Model edges represent activation or inhibition relationships between model constituents.
To ensure consistency and cohesiveness of the model for the primary layer (proteins interacting with the GR), proteins interacting in a highly indirect manner (i.e. through multiple steps and proteins) were excluded during curation. The curation evidence used for cofactors would indicate either the stimulatory or inhibitory effect of that cofactor on the GR, or a report demonstrating that the cofactor in question was a GR coactivator or corepressor. For the curation of the second layer (interactions between the proteins within the primary layer) the “intermediary rule” was applied. This rule covered cases for which literature curation indicated that despite STRING listing a direct link between Protein 1 and Protein 2 (both of which interact with the GR individually), this regulation actually occurred through an intermediary protein (i.e. Protein 1 -> Intermediary Protein -> Protein 2). In these cases, if the intermediary protein was present within the primary layer then the reactions would be listed as proceeding through the intermediary protein (i.e. Protein 1 -> Intermediary Protein -> Protein 2), provided no additional evidence of a direct relationship of Protein 1 -> Protein 2 was observed. In cases where the intermediary protein did not exist within the primary layer, the reaction instead was put as a direct Protein 1 -> Protein 2 to reduce redundancy.
In numerous cases, multiple proteins were combined as one node within the model. This was due to either the proteins forming a heterodimer or proteins from the same family being grouped together. These nodes and their constituents can be seen in the S1 Text file.
Following completion of the second layer, the model was connected to cell death and inflammation as two outputs through Gene Ontology. The full curation tables for the model (detailing the mode of interaction and at least one PubMed ID linking to a paper verifying the interaction) for the primary layer, second layer, and link to outputs can be seen in the S1 Text file.
The GEB052 model (Fig 2) consists of 52 nodes (proteins, inputs, outputs) connected by 241 logical interactions of activation or inhibition. Although the visualisation shown in Fig 2 is useful for providing an overview, it can be difficult to follow individual reactions in this detailed overview. As a complement to the full visualisation shown above, an interaction matrix (generated in CNA) is shown in Fig 3.
Feedback loops within biological networks are essential to maintain network integrity [18]. The GEB052 model contained 64 two step (i.e. Protein 1 -> Protein 2 -> Protein 1) loops, 26 of which (40.6%) involved the GR. This thus highlights the obvious centrality and importance of the GR within the network. Only two-step feedback loops are considered for examination as feedback loops may otherwise consist of numerous steps which would complicate analysis [18].
In addition to feedback loop assessment, the degree (node connectivity) distribution was assessed (Fig 4). Excluding cell death and inflammation, six nodes demonstrated a very high level of connectivity (twenty or more edges). On the far right of Fig 4 is the GR, with a degree of 83. In addition to the GR, other nodes showing a very high degree include: AP-1 (36 edges); CREBBP/EP300 (20 edges); IL6 (21 edges); STAT3 (20 edges) and TP53 (22 edges). Other nodes exhibiting a high degree (ten or more edges) include: CREB1 (12 edges); HDAC1 (14 interactions); HSP90 (11 edges); IL10 (15 edges); NFKB (16 edges); SMAD3 (13 edges) and SUMO (10 edges). Other nodes (n = 37) exhibited a lower degree, possessing less than ten edges. Table 1 summarises the degree range observed within the model.
Understanding the node connectivity within the model was crucial to the choice of which nodes would be selected for in silico knockout analysis, as previous studies have focussed on in silico knockouts for only the most highly connected nodes [18, 19].
CNA is capable of generating a dependency matrix which, by taking into account all of the signalling pathways present within the model, is able to determine the overall relationships from one node to another. Six types of dependencies are available: no effect; ambivalent (stimulatory and inhibitory influence); weak inhibitor; weak activator; strong inhibitor and strong activator. Fig 5 shows the visualised dependency matrix for the full GEB052 model. It is apparent from examination of Fig 5 that the majority of dependencies are ambivalent factors. This observation correlates with the large number of feedback loops, as the highly integrated signalling within the model can lead to multiple signalling paths between model nodes, both positive and negative. As ambivalent dependencies are those most likely to change following in silico knockouts, the high number of ambivalent dependencies represents a good starting point for downstream analysis.
In total, 2704 (52*52) dependencies were observed within the GEB052 model: 896 of these were of no effect; 1710 were ambivalent; 33 were weak inhibitors; 63 were weak activators; 2 were strong activators and there were no strong inhibitors. To characterise how relationships are altered after perturbation to the model (mimicking potential mutations in vivo), each of the highly connected nodes (≥10 interactions, excepting model outputs) was deleted from the model and a dependency matrix generated, with the results shown in Table 2.
For each knockout scenario, a total of 2601 (51*51) dependencies was observed and as expected due its centrality within the model, the removal of the GR had the most significant effects on the dependencies (Fig 6). The majority of dependency alterations were from ambivalent factors to no effect, which is consistent with the high connectivity of the GR resulting in many nodes signalling through it to affect others. Thus, removal of this intermediary node results in a loss of signalling between model constituents.
In addition to the changes from ambivalent to no effect, numerous changes to and from other dependencies were observed across the numerous knockout scenarios. Across all knockout scenarios for the GEB052 model a total of 1249 dependency alterations was observed, which is reflective of the significant number of relationship changes that occur when network elements are lost. Even if changes from ambivalent factors to no effects are not considered (as there is no net change in positive or negative regulation) 323 predictions of dependency alterations (to or from activators or inhibitors) were seen. Although these all may exert physiological effects when translated from in silico to in vivo, it is anticipated that strong activators or strong inhibitors are the dependencies most likely to show an effect. Therefore there is a necessary focus on changes to or from strong inhibitors to strong activators, as has been performed previously for interactome modelling [19, 22].
For example, removal of the GR (which mimics GR mutation in vivo), resulted in the emergence of one strong inhibitor and one additional strong activator when compared to the wild type model (Table 2). In the unperturbed model DAP3 was ambivalent towards cell death whereas in the absence of the GR it became strong activator. In addition, STAT5B in the wild type model was ambivalent towards cell death whereas removal of the GR led to the dependency changing to strong inhibition. These predictions if confirmed by literature searches and laboratory-based experiments may have important clinical implications.
To assess the accuracy of the model published literature was surveyed to investigate whether the model predictions in dependency alterations have been previously observed in experimental research. The predictions that could not be verified by literature searching, were marked as a “Potentially Novel Prediction” [18] and the results are detailed in Table 3.
Literature validation requires that the KO node, Node A and Node B are all mentioned; for instance, for row four of the above table, the paper would have to mention HDAC1 silencing or inhibition, which leads to DAXX activating SUMO. Otherwise, effects could be non-specific and not wholly consistent with model prediction. During the initial literature validation papers mentioning all three nodes could not be found.
However, some preliminary evidence has been gathered. The model predicted that in the absence of HSP90, PRKDC would be strongly activated by NCOA6 and by itself (likely via feedback loops). Corroborating this to some extent is one report which investigated the relationship between HSP90 and PRKDC (catalytic subunit of DNA-PK) and found that the use of the HSP90 inhibitor geldanamycin markedly enhanced TRAIL-induced DNA-PK [26]. However, this result was complicated as the same paper also showed that DNA-PK is a client of HSP90, which was required for full DNA-PK activation [26]. Thus, although the effector node (i.e. Node A) is not mentioned, it is promising that the overall outcome may correlate with model prediction.
Similar to the above, the model predicted that SUMO expression would be significantly higher following the loss of HDAC1. It has been shown that HDAC inhibition increases sumoylation in general, however the effect in one instance was mediated primarily through HDAC2 [27]. In addition, it has been demonstrated that HDAC1 inhibits sumoylation of a target protein therefore loss of HDAC1 would increase its sumoylation and thus the abundance of SUMO protein [28]. Again, this is consistent with model predictions, however the effector Node A (in this case DAXX or SUMO) has not been mentioned in this report.
Some model predictions were incorrect. The model predicted that loss of HDAC1 would lead to increased expression of DAXX; however, research has shown the opposite, with HDAC inhibitors leading to a decreased expression of DAXX [29]. But again, this paper does not specifically mention DAXX or SUMO as the effector node, so it is only a preliminary assessment of model accuracy.
Although analysis of individual relationships via dependency matrices may provide insight into altered signalling, logical steady state analysis (LSSA) assesses the entirety of the model under different scenarios. The basal state for all model nodes is undetermined (NaN). Given a set of input values (i.e. GC = 1 for a glucocorticoid-sensitive simulation, or GC = 1, GR = 0 for a glucocorticoid-resistant simulation) LSSA will proceed to calculate the state (1/ON, NaN/undetermined or 0/OFF) of every downstream node within the model. Lastly, two LSSA scenarios may be compared to generate an Emod value for each node which predicts the overall state change of the node between the two scenarios (1 = upregulated, 0 = no change, -1 = downregulated). Fig 7 provides a visual representation of the LSSA results for both the glucocorticoid-sensitive and glucocorticoid-resistant simulations, whilst Table 4 below summarises the LSSA results as well as the Emod value for each node.
More determined (ON or OFF) nodes were seen in the glucocorticoid-resistant simulation, however this is balanced by the fact that a significantly higher number of nodes (23.1%) were OFF in the glucocorticoid-resistant simulation, which may reflect a loss of overall functionality within the network. The Emod values (upregulated, no change, or downregulated) for nodes are summarised in Table 5.
Model predictions may be verified by literature searching in terms of experimental identification of upregulation or downregulation between glucocorticoid-sensitive and glucocorticoid-resistant cells. For instance, it has been shown that GLUL is downregulated in glucocorticoid-resistant cells [30]. However, a more practical approach to validating model LSSA predictions is the use of microarray data, detailed below.
Microarray data from glucocorticoid-sensitive and glucocorticoid-resistant cells were obtained from the Gene Expression Omnibus database and six comparisons were performed (as detailed in the Methods). The Emod values obtained were compared to Eexp values created via comparison of a glucocorticoid-sensitive and glucocorticoid-resistant microarray. Comparison of these values gives the number of correct, small error and large error predictions within the model. The S1 Text file contains tables that show the Emod and Eexp values in addition to their comparison for each microarray validation performed. Table 6 below summarises the overall number of correct/small error/large error predictions across all comparisons.
As shown in Table 6, the GEB052 model generated accurate predictions across all scenarios (ranging from 54.2% to 60.4%, with an average of 56.6%). Given there are three possible outcomes (correct, small error and large error) a random model would achieve an expected correct prediction rate of 33.3%. The correct predictions from the six comparisons when compared to what a random model would achieve leads to a p-value <0.01, providing further evidence of the predictive capacity and potential of the GEB052 model.
In addition to the cell-based microarray data described above, the GEB052 model has been assessed as a predictive clinical tool (based on LSSA results) using microarray data from thirteen leukaemia patients (see the Methods section). The output of this result is shown in Fig 8, and the model’s LSSA results perform less well for analysis of individual patient data as an average correct prediction rate of 42% was observed, with 55% small error and 3% large error. Thus, although 55% of predictions were small error, the fact that large errors are still less than 5% is a promising indicator of the potential of the model. The fixed-state nature of LSSA (having only three discrete values) is a limitation on the analytical output which may partially explain this outcome, and thus a more quantitative analysis was taken next.
The GEB052 model has also undergone analysis using a semi-quantitative signal transduction score flow algorithm (STSFA) that superimposes ChIP-seq and/or microarray data onto a model to analyse the network with numerical data. The same comparisons for genome-wide validation was again utilised here (see Methods). The S1 Text file contains tables that show the calculation and output for each individual STSFA analysis and a summary is provided in Table 7.
As shown in Table 7, STSFA analysis achieved significantly higher correct prediction rates than for discrete LSSA predictions (compare to Table 6). An average of 80.1% correct predictions was observed, with an average of 18.9% small error and 1.0% large error (and three out of six simulations exhibiting no large errors). The correct prediction rates for LSSA against STSFA have been graphed and compared via an unpaired t-test (Fig 9), which shows the enhanced predictive power that the semi-quantitative STSFA analysis offers. Due to this, assessment of the clinical potential of the GEB052 model with STSFA analysis was performed, as detailed below.
Using microarray data from thirteen leukaemia patients (see Methods), the GEB052 model was analysed with the STSFA and the relative activation/inhibition of cell death was calculated for each patient. Patients were divided into two groups (twelve patients alive at risk assessment or one deceased at risk assessment) and the average +/- SEM (Fig 10). GEB052 model predictions indicated that the patient who died before risk assessment would have cell death more negatively regulated than those who were alive at risk assessment. Given that glucocorticoids are a chemotherapeutic drug for leukaemia, the cell death node in the model translates to death of the cancer cells in vivo. Thus, the model predicted that the patient who died before risk assessment would have cell death more negatively regulated; meaning that more cancer cells survive, in turn suggesting a worse prognosis. Thus, these preliminary model predictions correlate with clinical outcomes for the patients.
The widespread therapeutic use of glucocorticoids for many different diseases leads to a need to identify causes of therapy failure and glucocorticoid resistance. Systems biology offers the possibility of integrating the detailed knowledge of GR signalling to generate models that can be used to gain insight into how the network functions following a loss of GR function.
Computational research methodologies have been previously applied to GR research using approaches such as virtual ligand screening [32], development of models to quantitatively model specific signalling events [17] or the creation of models that aim to simulate glucocorticoid receptor control of both directly-regulated and indirectly-regulated genes [33]. Each of these approaches have provided insight to GR signalling, however to date a Boolean interactome model of the glucocorticoid receptor has not been developed. To generate the GEB052 model, the STRING database was to provide a basis for the interactions to be included. Following the generation of all the model links between the proteins interacting with the GR, model outputs in the form of cell death and inflammation were added via Gene Ontology and manual curation.
Interactome modelling has previously been applied to cancer research, such as the development of the original PKT206 p53 interactome and the later expanded PMH260 interactome [18, 19]. These models, in addition to the application of the STSFA to the PKT206 model [22] all showed good predictive ratios and thus a similar approach was undertaken here to model GR signalling.
The GEB052 model consists of 52 nodes connected by 241 logical interactions, and has 64 two-step feedback loops within the model. Comparatively, the PKT206 model had only 30 two-step feedback loops, whilst the expanded PMH260 model had only 34 feedback loops [18, 19]. The identification of 64 two-step feedback loops within the GEB052 model is particularly interesting as the GEB052 model is significantly smaller than PKT206 or PMH260, containing only 52 nodes compared to 206 or 260. Thus, despite the model being significantly smaller the network appears to be much more integrated and interconnected, which may explain the fact that the majority of nodes were unchanged between the sensitive and resistant LSSA simulation results, as well as potentially explaining the fewer number of changes to strong activators or inhibitors following dependency matrix generation in KO scenarios; the PKT206 model identified 63 changes to/from strong activators or inhibitors, whilst only ten were seen in the GEB052 model. This is due to the fact that feedback loops have been previously identified as important in the robustness of a network [18]. Although this initial analysis has focussed only on strong activators or inhibitors, 323 changes to or from activators or inhibitors were identified across all knockout scenarios, and thus examination of these dependency alterations would represent a source of future work.
The validation of model LSSA results through cell-based microarray data indicated an average of 56.60% correct predictions, 41.67% small error and 1.74% large error. The PKT206 interactome model displayed a correct prediction range from 52–71% [18]. The correct prediction range for the GEB052 model was lower (54.17% to 60.42%), less large errors were seen in the GEB052 model validation; two out of six comparisons yielded no large errors, whilst the other four led to a large error range of only 2.08% to 4.17%. It is important to note that the expanded PMH260 interactome displayed less large errors than the original PKT206 model, and therefore model expansion represents an additional source of future work for the GEB052 model.
Validation of model LSSA results with patient microarray data yielded lower correct prediction rates (an average of 42%). However, it is still promising that large error predictions comprised the minority of prediction outcomes, suggesting some potential of the model. Furthermore, working on the assumption that a random model would achieve a correct prediction rate of 33.3%, then a 42% correct average from thirteen sets of data is statistically significantly higher (p<0.0001).This lower correct prediction rate could be attributed to a variety of factors, including the relatively small size of the model, as well as the complexity of translating findings from a simulation of a small gene regulatory network to a whole organism level. Additionally, as specified in the introduction, effects of GCs are very cell-type specific; the GR may differentially modulate genes depending on the type of tissue. Thus, an additional way to further develop the GR model is through the incorporation of tissue-specific interactions and the development of cell-specific forms of the model, although the relevant literature required for this is currently incomplete.
Consistent with previous research [22] the STSFA demonstrated a statistically significantly higher level of correct prediction rates (80.1% for STSFA compared to 56.6% for LSSA). Curiously, large error predictions for STSFA appeared only in microarray data from B-ALL and not T-ALL, which indicates that the model may predict T-ALL to a better standard than B-ALL. Furthermore, the enhanced predictive power of STSFA (likely due to its semi-quantitative nature) provided a justification for its use in clinical assessment. By using microarray data from thirteen leukaemia patients (taken before chemotherapy treatment, and thus analytical outcomes represent true predictions) Fig 10 was generated. The fact that the model predicted that the patient who died before risk assessment would have cell death being more inhibited (equating to death of the cancer cells) is promising, as it provides a potential link between model prediction and clinical outcome. However, this analysis is admittedly preliminary due to the fact that there are a small number of patients (13) and one group had only one patient, whilst the other had twelve. Thus, although promising, further assessment with a larger patient cohort is needed.
Ultimately, the GEB052 model construction, validation, and clinical assessment represent a proof of principle of the applicability of this approach to glucocorticoid receptor research. The GEB052 model under Boolean analysis provides good predictive ratios for cell-based microarray data, and application of the semi-quantitative STSFA to the model demonstrated even higher correct predictive rates. Lastly, the use of the GEB052 model under STSFA analysis has also shown promise at the clinical level using microarray data from thirteen leukaemia patients. Key points for future development include model expansion and incorporation of tissue-specific reactions. In addition, it is recognised that there are multiple isoforms of the GR, each of which can have different effects on downstream nodes, and in fact interactions between different GR isoforms can be a determinant of its activity [34]. Thus, it may also be useful to develop models of different GR isoforms to better represent physiological occurrences. Regardless of future directions, the GEB052 model represents a promising starting point and potential clinical tool given its predictive ratios and the correlation of its STSFA output with patient clinical outcomes. Application of individual patient data to the model could thus be a stepping stone towards personalised therapy.
STRING (Search Tool for the Retrieval of Interacting Genes/Proteins, v9.1 at the time of curation) was used as the database of known and predicted protein interactions [24]. Extraction and filtering of data was performed in a similar manner as described previously [18]. The “protein.actions.v9.1.txt.gz” file was downloaded from STRING and all high confidence (≥ 0.7) interactions for the glucocorticoid receptor were then extracted. TSC22D3 (GILZ, glucocorticoid-induced leucine zipper) and EP300 were also included due to their known importance in GR signalling or similarity to CREBBP respectively.
Manual curation of STRING data was then undertaken via extensive literature searches of the two putative interacting proteins. STRING includes various interaction modes such as “activation”, “inhibition” and “binding”. In all cases, manual curation was undertaken to confirm STRING records, and also to uncover any functional relationships between the two genes that were not included in STRING. Manual curation was essential as the nature of the STRING database (such as being based on text mining) results in the possibility of incorrect interactions being retained in the database. It has previously been shown that multiple types of errors can occur such as incorrect gene name recognition [18].
After manual curation of all the interactions with the GR (the “primary layer”), all high-confidence interactions for the proteins that were shown to interact with GR were extracted. This list was then filtered to retain interactions only between the proteins which appeared in the primary layer. Additional curation was then undertaken in order to verify STRING data (the “second layer”), and thus after this a closed two-layer model was produced. All curations of predicted interactions were double-curated to improve model reliability.
The GO database [35] was used to provide biological outputs for the model. Following completion of the second layer, GO terms/annotations were collected for each node of the network, pooled together and ranked by the most common, and then the most common terms related to biological outputs were chosen. This lead to several groups of GO terms: cell death; inflammation; immune response; metabolism; development; cell growth and proliferation. For this first version of the model only cell death and inflammation were chosen as outputs due to their relevance in glucocorticoid therapy. For all model links to outputs, manual double-curation was again undertaken to verify interactions.
Model visualisation was undertaken through the use of Cytoscape, an open-source software for data visualisation [36]. Curated interaction records were imported into the program and visualised after adjusting parameters. Application of the STSFA (below) was also conducted in Cytoscape.
CellNetAnalyzer (CNA, v2017.1c) is a MATLAB toolbox which allows for the analysis of gene-regulatory models based on the topology of the interaction network. Interactions between nodes of the network are represented through hypergraphs which can allow for interaction combinations such as OR functions or the use of AND functions, both of which allow for more accurate representation of true biological reactions (such as several proteins forming a complex to activate or inhibit a target) [23]. CNA was used to construct Boolean signal flow networks. At present, the model presented herein does not contain AND reactions; in cases where interactions converge to the same node the combination follows OR logic by default. Inhibitory reactions are represented by a NOT modifier. These logics have been described in detail by Klamt and colleagues [23].
Several types of analysis are available through CNA, such as the generation of an interaction matrix (which summarises the participation of each node in every reaction), logical steady state analysis (LSSA) and the generation of dependency matrices. By defining (i.e. ON or OFF) the state of nodes (particularly input nodes) of the model, LSSA will calculate the steady state of network nodes downstream of the input based on the interactions within the model. Three node states are possible under LSSA: 1 (ON), 0 (OFF) or NaN (undetermined). A node may be assigned NaN if multiple states are possible; this may be caused by input conditions being insufficient to determine all node states, or through feedback loops leading to multiple steady states and oscillatory behaviour [18, 37].
The second main approach used in CellNetAnalyzer is the generation of dependency matrices. A dependency matrix provides a visual and numerical representation of the overall relationships between the nodes of the network, taking into account all of the interactions within the model (thus allowing indirect functional relationships to be considered). Six different types of dependencies are possible based on the relationship between nodes in the interaction:
Comparison of the dependency matrices from the full model to a modified (i.e. KO model) can unveil modified relationships and signalling. Because model KOs simulate in vivo loss-of-function mutations, these matrix comparisons provide predictions for how cells will behave. These predictions may then be verified in the laboratory to assess model predictive power and model accuracy [18, 19].
Comparisons between two sets of LSSA results (such as a GC-sensitive scenario against a GC-resistant scenario) were also carried out as previously described [18], which allows for the assessment of node upregulation or downregulation between two scenarios.
In brief, LSSA calculates the state (inactivated (0), undetermined (NaN) or activated (1)) of nodes within the network following a set of input value(s). For Scenario 1 (i.e. a GC-sensitive simulation), node i state was defined as S(i)1 which has a value of NaN, 0, or 1. For Scenario 2 (i.e. a GC-resistant simulation), node i state was defined as S(i)2, which may also have a value of NaN, 0, or 1. Lastly the value Emod was used to define the predicted change in node state from Scenario 1 to Scenario 2, where 1 means the node is upregulated, -1 means the node is downregulated and 0 means the node state is unchanged:
Emod=-1ifS(i)1=1andS(i)2=0Emod=-1ifS(i)1=1andS(i)2=NaNEmod=-1ifS(i)1=NaNandS(i)2=0
Emod=0ifS(i)1=1andS(i)2=1Emod=0ifS(i)1=0andS(i)2=0Emod=0ifS(i)1=NaNandS(i)2=NaN
Emod=1ifS(i)1=0andS(i)2=1Emod=1ifS(i)1=NaNandS(i)2=1Emod=1ifS(i)1=0andS(i)2=NaN
Consistent with previous publications [18, 19] the predictions generated by the GEB052 model were assessed against microarray data from GC-resistant and GC-sensitive cells. Twelve microarrays were obtained from the Gene Expression Omnibus (GEO) database, and the following six comparisons were utilised as shown in Table 8.
Differential expression analysis was performed using the dynamic threshold method used by Tian and colleagues and Hussain and colleagues [18, 19] and the expression change (Eexp) value was calculated, where 1 equates to upregulation, 0 to no change, and -1 to downregulation. Using the GC-resistant array as the target scenario, and the GC-sensitive array as the source scenario, fold changes for all microarray probe IDs were generated between the target and source scenarios. The Log10 for all fold changes was calculated, and a dynamic threshold based on the average Log10 fold change + the standard deviation (upper limit) and the average Log10 fold change–the standard deviation (lower limit).
For each gene present in the model, the median value for all probe IDs relevant to the gene was calculated for both the source and target scenario, in addition to the fold change of the median values. For model nodes which represented the combination of multiple genes (i.e. the AP-1 node which represents FOS and JUN) the median value for all probe IDs for all of its constituents was used. Log10 values of these fold changes were calculated and compared to the dynamic threshold; if higher than the upper limit, the gene was determined as upregulated (Eexp = 1), whilst if the value was lower than the lower limit the gene was determined as downregulated (Eexp = -1) and if its value lay between the lower and upper limits then the gene was determined as unchanged (Eexp = 0).
To evaluate model performance, the absolute value of Emod−Eexp was calculated, which could take three possible values: 0 (no difference between Emod and Eexp; model prediction was correct); 1 (small difference between Emod and Eexp; small error prediction) and 2 (large difference between Emod and Eexp; large error prediction meaning that the model predicted the opposite of what occurred in cells).
Table 9 shows the microarray data used for clinical validation of LSSA results.
To compare model LSSA results with clinical data from patients, thirteen microarrays (detailed in Table 9) from leukaemia patients (taken following treatment with prednisolone) were obtained from the GEO database [31]. For each individual patient, Log10 RMA values for all probe IDs were calculated and a dynamic threshold based on the average +/- standard deviation was generated. The median Log10 values for all the probe IDs for genes within the model were then compared to the threshold: if the value was higher than the upper limit, the gene was considered as upregulated; if the value was lower than the lower limit, the gene was considered as downregulated and if the value lay between the lower and upper limits then the gene was unchanged. These values were then compared to model LSSA results of a GC-sensitive simulation, where 1 is equivalent to upregulated, 0 to downregulated and NaN to unchanged.
The STSFA plugin for Cytoscape [21] was used to apply the STSFA to the model. As with previous studies [19] Log2 RMA values were scaled up by a factor of 100 and superimposed onto the model using the pathway scoring application. A limitation of the STSFA is that it apparently cannot handle directly ambivalent relationships; that is, if Node A both directly activates and inhibits Node B, the STSFA cannot accurately handle this. To correct for this, all directly ambivalent relationships were removed prior to the application of the STSFA. Mathematically, this is not unreasonable as even if the direct ambivalent interactions were considered, the overall regulation would be zero as it would theoretically be positively and negatively affected by equal amounts. The same twelve microarray datasets listed in Table 8 were used for STSFA analysis.
STSFA results from the GC-sensitive and GC-resistant arrays were used to generate an Emod value, whilst the Eexp values for each Comparison were the same as for the cell-based microarray genome-wide model validation. To generate the Emod values, fold changes between the node scores of the resistant output and sensitive output were generated, followed by the Log10 of the fold changes. From the Log10 fold changes for each node a dynamic threshold based on the average +/- standard deviation was generated, and nodes were considered as upregulated if their score was higher than the upper limit, downregulated if their score was lower than the lower limit, and unchanged if the score lay between the two. These Emod values were compared to the Eexp values to assess model accuracy in the same way as the cell-based microarray genome-wide model validation.
To assess the potential of the GEB052 model as a predictive clinical tool, microarray data from thirteen leukaemia patients (taken before patients were treated) was obtained from the GEO database, following its deposit after the original study that generated the data [31] (Table 10).
For each patient the STSFA was used to superimpose their microarray data onto the model. The STSFA assigns node score to every node within the model, in addition to calculating weights for each of the edges (indicating the “strength” of the regulation). Patients were split into two groups (alive at risk assessment or dead at risk assessment). The total incoming edge weights to cell death (one output of the GEB052 model) was calculated for each patient, and an average made for each group, in addition to calculating the SEM. This analysis thus correlated model predictions with clinical outcomes.
To assess whether the correct prediction rates of the model were statistically significant, the WolframAlpha computational knowledge engine (http://www.wolframalpha.com/) was used in conjunction with the search term “Probability of [X] success in [Y] trials, chance of success is [Z]”. In these cases [X] equates to the number of correct predictions, [Y] to the total number of predictions and [Z] to the chance of success (one in three, as there are three possible outcomes).
To determine if any nodes were systematically incorrect across the comparisons shown in Table 8, the absolute values of Emod−Eexp were totalled for each node and all comparisons. As previously stated, the absolute value of Emod−Eexp can take three possible values: 0 (correct), 1 (small error) and 2 (large error). A threshold of four (indicating that the node had small errors in more than 50% of comparisons, or had two large errors) was chosen to determine a node as incorrect. The incorrect node determination for cell-based microarray validation is shown in the S1 Text file.
|
10.1371/journal.pcbi.1000579 | A High-Throughput Screening Approach to Discovering Good Forms of Biologically Inspired Visual Representation | While many models of biological object recognition share a common set of “broad-stroke” properties, the performance of any one model depends strongly on the choice of parameters in a particular instantiation of that model—e.g., the number of units per layer, the size of pooling kernels, exponents in normalization operations, etc. Since the number of such parameters (explicit or implicit) is typically large and the computational cost of evaluating one particular parameter set is high, the space of possible model instantiations goes largely unexplored. Thus, when a model fails to approach the abilities of biological visual systems, we are left uncertain whether this failure is because we are missing a fundamental idea or because the correct “parts” have not been tuned correctly, assembled at sufficient scale, or provided with enough training. Here, we present a high-throughput approach to the exploration of such parameter sets, leveraging recent advances in stream processing hardware (high-end NVIDIA graphic cards and the PlayStation 3's IBM Cell Processor). In analogy to high-throughput screening approaches in molecular biology and genetics, we explored thousands of potential network architectures and parameter instantiations, screening those that show promising object recognition performance for further analysis. We show that this approach can yield significant, reproducible gains in performance across an array of basic object recognition tasks, consistently outperforming a variety of state-of-the-art purpose-built vision systems from the literature. As the scale of available computational power continues to expand, we argue that this approach has the potential to greatly accelerate progress in both artificial vision and our understanding of the computational underpinning of biological vision.
| One of the primary obstacles to understanding the computational underpinnings of biological vision is its sheer scale—the visual system is a massively parallel computer, comprised of billions of elements. While this scale has historically been beyond the reach of even the fastest super-computing systems, recent advances in commodity graphics processors (such as those found in the PlayStation 3 and high-end NVIDIA graphics cards) have made unprecedented computational resources broadly available. Here, we describe a high-throughput approach that harnesses the power of modern graphics hardware to search a vast space of large-scale, biologically inspired candidate models of the visual system. The best of these models, drawn from thousands of candidates, outperformed a variety of state-of-the-art vision systems across a range of object and face recognition tasks. We argue that these experiments point a new way forward, both in the creation of machine vision systems and in providing insights into the computational underpinnings of biological vision.
| The study of biological vision and the creation of artificial vision systems are naturally intertwined—exploration of the neuronal substrates of visual processing provides clues and inspiration for artificial systems, and artificial systems, in turn, serve as important generators of new ideas and working hypotheses. The results of this synergy have been powerful: in addition to providing important theoretical frameworks for empirical investigations (e.g. [1]–[6]), biologically-inspired models are routinely among the highest-performing artificial vision systems in practical tests of object and face recognition [7]–[12].
However, while neuroscience has provided inspiration for some of the “broad-stroke” properties of the visual system, much is still unknown. Even for those qualitative properties that most biologically-inspired models share, experimental data currently provide little constraint on their key parameters. As a result, even the most faithfully biomimetic vision models necessarily represent just one of many possible realizations of a collection of computational ideas.
Truly evaluating the set of biologically-inspired computational ideas is difficult, since the performance of a model depends strongly on its particular instantiation–the size of the pooling kernels, the number of units per layer, exponents in normalization operations, etc. Because the number of such parameters (explicit or implicit) is typically large, and the computational cost of evaluating one particular model is high, it is difficult to adequately explore the space of possible model instantiations. At the same time, there is no guarantee that even the “correct” set of principles will work when instantiated on a small scale (in terms of dimensionality, amount of training, etc.). Thus, when a model fails to approach the abilities of biological visual systems, we cannot tell if this is because the ideas are wrong, or they are simply not put together correctly or on a large enough scale.
As a result of these factors, the availability of computational resources plays a critical role in shaping what kinds of computational investigations are possible. Traditionally, this bound has grown according to Moore's Law [13], however, recently, advances in highly-parallel graphics processing hardware (such as high-end NVIDIA graphics cards, and the PlayStation 3's IBM Cell processor) have disrupted this status quo for some classes of computational problems. In particular, this new class of modern graphics processing hardware has enabled over hundred-fold speed-ups in some of the key computations that most biologically-inspired visual models share in common. As is already occurring in other scientific fields [14],[15], the large quantitative performance improvements offered by this new class of hardware hold the potential to effect qualitative changes in how science is done.
In the present work, we take advantage of these recent advances in graphics processing hardware [16],[17] to more expansively explore the range of biologically-inspired models–including models of larger, more realistic scale. In analogy to high-throughput screening approaches in molecular biology and genetics, we generated and trained thousands of potential network architectures and parameter instantiations, and we “screened” the visual representations produced by these models using tasks that engage the core problem of object recognition–tolerance to image variation [10]–[12],[18],[19]. From these candidate models, the most promising were selected for further analysis.
We show that this large-scale screening approach can yield significant, reproducible gains in performance in a variety of basic object recognitions tasks and that it holds the promise of offering insight into which computational ideas are most important for achieving this performance. Critically, such insights can then be fed back into the design of candidate models (constraining the search space and suggesting additional model features), further guiding evolutionary progress. As the scale of available computational power continues to expand, high-throughput exploration of ideas in computational vision holds great potential both for accelerating progress in artificial vision, and for generating new, experimentally-testable hypotheses for the study of biological vision.
In order to generate a large number of candidate model instantiations, it is necessary to parameterize the family of all possible models that will be considered. A schematic of the overall architecture of this model family, and some of its parameters, is shown in Figure 2. The parameterization of this family of models was designed to be as inclusive as possible–that is, the set of model operations and parameters was chosen so that the family of possible models would encompass (as special cases) many of the biologically-inspired models already described in the extant literature (e.g. [1]–[4],[7],[9]). For instance, the full model includes an optional “trace” term, which allows learning behavior akin to that described in previous work (e.g. [4], [20]–[22]). While some of the variation within this family of possible models might best be described as variation in parameter tuning within a fixed model architecture, many parameters produce significant architectural changes in the model (e.g. number of filters in each layer). The primary purpose of this report is to present an overarching approach to high-throughput screening. While precise choices of parameters and parameter ranges are clearly important, one could change which parameters were explored, and over what ranges, without disrupting the integrity of the overarching approach. An exhaustive description of specific model parameters used here is included in the Supplemental Text S1, and is briefly described next.
Model parameters were organized into four basic groups. The first group of parameters controlled structural properties of the system, such as the number of filters in each layer and their sizes. The second group of parameters controlled the properties of nonlinearities within each layer, such as divisive normalization coeffients and activation functions. The third group of parameters controlled how the models learned filter weights in response to video inputs during an Unsupervised Learning Phase (this class includes parameters such as learning rate, trace factors, etc.; see Phase 2: Unsupervised Learning below). A final set of parameters controlled details of how the resulting representation vectors are classified during screening and validation (e.g. parameters of dimensionality reduction, classification parameters, etc.). For the purposes of the work presented here, this class of classification-related parameters was held constant for all analyses below. Briefly, the output values of the final model layer corresponding to each test example image were “unrolled” into a vector, their dimensionality was reduced using Principal Component Analysis (PCA) keeping as many dimensions as there were data points in the training set, and labeled examples were used to train a linear Support Vector Machine (SVM).
Each model consisted of three layers, with each layer consisting of a “stack” of between 16 and 256 linear filters that were applied at each position to a region of the layer below. At each stage, the output of each unit was normalized by the activity of its neighbors within a parametrically-defined radius. Unit outputs were also subject to parameterized threshold and saturation functions, and the output of a given layer could be spatially resampled before being given to the next layer as input. Filter kernels within each stack within each layer were initialized to random starting values, and learned their weights during the Unsupervised Learning Phase (see below, see Supplemental Text S1). Briefly, during this phase, under parametric control, a “winning” filter or filters were selected for each input patch, and the kernel of these filters was adapted to more closely resemble that patch, achieving a form of online non-parametric density estimation. Building upon recent findings from visual neuroscience [18],[23],[24], unsupervised learning could also be biased by temporal factors, such that filters that “won” in previous frames were biased to win again (see Supplemental Text S1 for details).
It should be noted that while the parameter set describing the model family is large, it is not without constraints. While our model family includes a wide variety of feed-forward architectures with local intrinsic processing (normalization), we have not yet included long-range feedback mechanisms (e.g. layer to layer). While such mechanisms may very well turn out to be critically important for achieving the performance of natural visual systems, the intent of the current work is to present a framework to approach the problem. Other parameters and mechanisms could be added to this framework, without loss of generality. Indeed, the addition of new mechanisms and refinement of existing ones is a major area for future research (see Discussion).
While details of the implementation of our model class are not essential to the theoretical implications of our approach, attention must nonetheless be paid to speed in order to ensure the practical tractability, since the models used here are large (i.e. they have many units), and because the space of possible models is enormous. Fortunately, the computations underlying our particular family of candidate models are intrinsically parallel at a number of levels. In addition to coarse-grain parallelism at the level of individual model instantiations (e.g. multiple models can be evaluated at the same time) and video frames (e.g. feedforward processing can be done in parallel on multiple frames at once), there is a high degree of fine-grained parallelism in the processing of each individual frame. For instance, when a filter kernel is applied to an image, the same filter is applied to many regions of the image, and many filters are applied to each region of the image, and these operations are largely independent. The large number of arithmetic operations per region of image also results in high arithmetic intensity (numbers of arithmetic operations per memory fetch), which is desirable for high-performance computing, since memory accesses are typically several orders of magnitude less efficient than arithmetic operations (when arithmetic intensity is high, caching of fetched results leads to better utilization of a processor's compute resources). These considerations are especially important for making use of modern graphics hardware (such as the Cell processor and GPUs) where many processors are available. Highly-optimized implementations of core operations (e.g. linear filtering, local normalization) were created for both the IBM Cell Processor (PlayStation 3), and for NVIDIA graphics processing units (GPUs) using the Tesla Architecture and the CUDA programming model [25]. These implementations achieve highly significant speed-ups relative to conventional CPU-based implementations (see Figure 1 and Supplemental Figure S1). High-level “outer loop” coordination of these highly optimized operations was accomplished using the Python programming language (e.g. using PyCUDA [26]), allowing for a favorable balance between ease of programming and raw speed (see Supplemental Text S2). In principle, all of the analyses presented here could have been performed using traditional computational hardware; however, the cost (in terms of time and/or money) of doing so with current CPU hardware is prohibitive.
Figure 1 shows the relative speedup and performance/cost of each implementation (IBM Cell on Sony's PlayStation 3 and several NVIDIA GPUs) relative to traditional MATLAB and multi-threaded C code for the linear filtering operation (more details such as the raw floating point performance can be found in the Supplemental Figure S1). This operation is not only a key component of the candidate model family (see below) but it's also the most computationally demanding, reaching up to 94% of the total processing time (for the PlayStation 3 implementation), depending on model parameters (average fraction is 28%). The use of commodity graphics hardware affords orders-of-magnitude increases in performance. In particular, it should be noted that the data presented in this work took approximately one week to generate using our PlayStation 3-based implementation (222x speedup with one system) on a cluster of 23 machines. We estimate that producing the same results at the same cost using a conventional MATLAB implementation would have taken more than two years (see Figure S1).
Our approach is to sample a large number of model instantiations, using a well-chosen “screening” task to find promising architectures and parameter ranges within the model family. Our approach to this search was divided into four phases (see Figure 3): Candidate Model Generation, Unsupervised Learning, Screening, and Validation/Analysis of high-performing models.
As a first exploration of our high-throughput approach, we generated 7,500 model instantiations, in three groups of 2,500, with each group corresponding to a different class of unsupervised learning videos (“petri dishes”; see Methods). During the Screening Phase, we used the “Cars vs. Planes” object discrimination task [11] to assess the performance of each model, and the most promising five models from each set of 2,500 models was submitted to further analysis. The raw computation required to generate, train and screen these 7,500 models was completed in approximately one week, using 23 PlayStation 3 systems [41]. Results for models trained with the “Law and Order” petri dish during the Unsupervised Learning Phase are shown in Figure 6A. As expected, the population of randomly-generated models exhibited a broad distribution of performance on the screening task, ranging from chance performance (50%) to better than 80% correct. Figure 6B shows the performance of the best five models drawn from the pool of 2,500 models in the “Law and Order” petri dish. These models consistently outperformed the V1-like model baseline (Figure 7), and this performance was roughly maintained even when the model was retrained with a different video set (e.g. a different clip from Law and Order), or with a different random initialization of the filter kernel weights (Figure 6C).
Since these top models were selected for their high performance on the screening task, it is perhaps not surprising that they all show a high level of performance on that task. To determine whether the performance of these models generalized to other test sets, a series of Validation tests were performed. Specifically, we tested the best five models from each Unsupervised Learning petri dish on four test sets: two rendered object sets, one rendered face set, and a modified subset of the MultiPIE face recognition image set (see Validation Phase in Methods). Performance across each of these validation sets is shown in Figure 7 (black bars). While the exact ordering of model performance varied somewhat from validation set to validation set, the models selected during the Screening Phase performed well across the range of validation tasks.
The top five models found by our high-throughput screening procedure generally outperformed state-of-the-art models from the literature (see Methods) across all sets, with the best model found by the high-throughput search uniformly yielding the highest performance across all validation sets. Even greater performance was achieved by a simple summing of the SVM kernels from the top five models (red bar, Figure 7). Of note, the nearest contender from the set of state-of-the-art models is another biologically-inspired model [7],[8].
Interestingly, a large performance advantage between our high-throughput-derived models and state-of-the-art models was observed for the MultiPIE hybrid set, even though this is arguably the most different from the task used for screening, since it is composed from natural images (photographs), rather than synthetic (rendered) ones. It should be noted that several of the state-of-the-art models, including the sparse C2 features (“SLF” in Figure 7), which was consistently the nearest competitor to our models, used filters that were individually tailored to each validation test–i.e. the representation used for “Boats vs. Planes” was optimized for that set, and was different from the representation used for the MultiPIE Hybrid set. This is in contrast to our models, which learned their filters from a completely unrelated video data set (Law and Order) and were screened using an unrelated task (“Cars vs. Planes”, see Methods). While even better performance could no doubt be obtained by screening with a subset taken from each individual validation test, the generalizability of performance across a range of different tasks argues that our approach may be uncovering features and representations that are broadly useful. Such generality is in keeping with the models' biological inspiration, since biological visual representations must be flexible enough to represent a massive diversity of objects in order to be useful.
Results for the 2,500 models in each of the other two “petri dishes” (i.e. models trained with alternate video sets during unsupervised learning) were appreciably similar, and are shown in Supplemental Figures S7 and S8, using the same display conventions set forth in Figures 6 and 7.
We have demonstrated a high-throughput framework, within which a massive number of candidate vision models can be generated, screened, and analyzed. Models found in this way were found to consistently outperform an experimentally-motivated baseline model (a V1-like model; [10]–[12]), and the representations of visual space instantiated by these models were found to be useful generally across a variety of object recognition tasks. The best of these models and the blend of the five best models were both found to consistently outperform a variety of state-of-the-art machine vision systems for all of the test sets explored here, even without any additional optimization.
This work builds on a long tradition of machine vision systems inspired by biology (e.g. [1]–[4],[7],[9]). However, while this past work has generated impressive progress towards building artificial visual systems, it has explored only a few examples drawn from the larger space of biologically-inspired models. While the task of exploring the full space of possible model instantiations remains daunting (even within the relatively restricted “first-order” class of models explored here), our results suggest that even a relatively simple, brute-force high-throughput search strategy is effective in identifying promising models for further study. In the parameter space used here, we found that a handful of model instantiations performed substantially better than the rest, with these “good” models occurring at a rate of approximately one in five-hundred. The relative rarity of these models underscores the importance of performing large-scale experiments with many model instantiations, since these models would be easy to miss in a “one-off” mode of exploration. Importantly, these rare, high-performing models performed well across a range of object recognition tasks, indicating that our approach does not simply optimize for a given task, but can uncover visual representations of general utility.
Though not conceptually critical to our approach, modern graphics hardware played an essential role in making our experiments possible. In approximately one week, we were able to test 7,500 model instantiations, which would have taken approximately two years using a conventional (e.g. MATLAB-based) approach. While it is certainly possible to use better-optimized CPU-based implementations, GPU hardware provides large increases in attainable computational power (see Figure 1 and Supplemental Figure S1).
An important theme in this work is the use of parametrically controlled objects as a way of guiding progress. While we are ultimately interested in building systems that tolerate image variation in real-world settings, such sets are difficult to create, and many popular currently-available “natural” object sets have been shown to lack realistic amounts of variation [10]–[12]. Our results show that it is possible to design a small synthetic set to screen and select models that generalize well across various visual classification tasks, suggesting that parametric sets can capture the essence of the invariant object recognition problem. Another critical advantage of the parametric screening approach presented here is that task difficulty can be increased on demand–that is, as models are found that succeed for a given level of image variation, the level of variation (and therefore the level of task difficulty), can be “ratcheted up” as well, maintaining evolutionary “pressure” towards better and better models.
While we have used a variety of synthetic (rendered) object image sets, images need not be synthetic to meet the requirements of our approach. The modified subset of the MultiPIE set used here (“MultiPIE Hybrid”, Figure 5) is an example of how parametric variation can also be achieved using carefully controlled photography.
While our approach has yielded a first crop of promising biologically-inspired visual representations, it is another, larger task to understand how these models work, and why they are better than other alternatives. While such insights are beyond the scope of the present paper, our framework provides a number of promising avenues for further understanding.
One obvious direction is to directly analyze the parameter values of the best models in order to understand which parameters are critical for performance. Figure 8 shows distributions of parameter values for four arbitrarily chosen parameters. While in no way conclusive, there are hints that some particular parameter values may be more important for performance than others (for quantitative analysis of the relationship between model parameters and performance, see Supplemental Text S3, Figures S9 and S10). The speed with which large collections of models can be evaluated opens up the possibility of running large-scale experiments where given parameters are held fixed, or varied systematically. Insights derived from such experiments can then be fed back into the next round of high-throughput search, either by adjusting the parameter search space or by fundamentally adjusting the algorithm itself. Such iterative refinement is an active area of research in our group.
The search procedure presented here has already uncovered promising visual representations, however, it represents just the simplest first step one might take in conducting a large-scale search. For the sake of minimizing conceptual complexity, and maximizing the diversity of models analyzed, we chose to use random, brute-force search strategy. However, a rich set of search algorithms exist for potentially increasingly the efficiency with which this search is done (e.g. genetic algorithms [42], simulated annealing [43], and particle swarm techniques [44]). Interestingly, our brute-force search found strong models with relatively high probability, suggesting that, while these models would be hard to find by “manual” trial-and-error, they are not especially rare in the context of our high-throughput search.
While better search algorithms will no doubt find better instances from the model class used here, an important future direction is to refine the parameter-ranges searched and to refine the algorithms themselves. While the model class described here is large, the class of all models that would count as “biologically-inspired” is even larger. A critical component of future work will be to adjust existing mechanisms to achieve better performance, and to add new mechanisms (including more complex features such as long-range feedback projections). Importantly, the high-throughput search framework presented here provides a coherent means to find and compare models and algorithms, without being unduly led astray by weak sampling of the potential parameter space.
Another area of future work is the application of high-throughput screening to new problem domains. While we have here searched for visual representations that are good for object recognition, our approach could also be applied to a variety of other related problems, such as object tracking, texture recognition, gesture recognition, feature-based stereo-matching, etc. Indeed, to the extent that natural visual representations are flexibly able to solve all of these tasks, we might likewise hope to mine artificial representations that are useful in a wide range of tasks.
Finally, as the scale of available computational resources steadily increases, our approach naturally scales as well, allowing more numerous, larger, and more complex models to be examined. This will give us both the ability to generate more powerful machine vision systems, and to generate models that better match the scale of natural systems, providing more direct footing for comparison and hypothesis generation. Such scaling holds great potential to accelerate both artificial vision research, as well as our understanding of the computational underpinnings of biological vision.
|
10.1371/journal.pbio.0050072 | Modification of a Hydrophobic Layer by a Point Mutation in Syntaxin 1A Regulates the Rate of Synaptic Vesicle Fusion | Both constitutive secretion and Ca2+-regulated exocytosis require the assembly of the soluble N-ethylmaleimide–sensitive factor attachment protein receptor (SNARE) complexes. At present, little is known about how the SNARE complexes mediating these two distinct pathways differ in structure. Using the Drosophila neuromuscular synapse as a model, we show that a mutation modifying a hydrophobic layer in syntaxin 1A regulates the rate of vesicle fusion. Syntaxin 1A molecules share a highly conserved threonine in the C-terminal +7 layer near the transmembrane domain. Mutation of this threonine to isoleucine results in a structural change that more closely resembles those found in syntaxins ascribed to the constitutive secretory pathway. Flies carrying the I254 mutant protein have increased levels of SNARE complexes and dramatically enhanced rate of both constitutive and evoked vesicle fusion. In contrast, overexpression of the T254 wild-type protein in neurons reduces vesicle fusion only in the I254 mutant background. These results are consistent with molecular dynamics simulations of the SNARE core complex, suggesting that T254 serves as an internal brake to dampen SNARE zippering and impede vesicle fusion, whereas I254 favors fusion by enhancing intermolecular interaction within the SNARE core complex.
| Most living cells constantly renew their membrane compositions and frequently communicate with neighboring cells by delivering cargo molecules from small vesicles. A key step in cargo delivery requires the fusion of the vesicle membrane with the target membrane mediated by SNARE proteins. In most cellular compartments, fusion occurs constitutively, requiring little participation of other molecules. In other cellular compartments, such as synapses in the nervous system, vesicle fusion is predominantly triggered by intracellular calcium ions. At present, constitutive and regulated fusion modes are not well understood.
In this study, we found that a mutant SNARE protein, syntaxin at the synapse, contained a building block commonly conserved for syntaxins functioning along constitutive secretory pathways. Further, our modeling predicted that the mutant syntaxin could form a tightly packed SNARE bundle closely resembling that found in the endosome, but differing from the relatively loosely packed bundle found at the wild-type synapse. Our experimental data support the hypothesis that the mutant syntaxin lowered the energy barrier for vesicle fusion by tightening the SNARE bundle. These findings reveal a novel, intrinsic structural feature of the SNARE complex that regulates vesicle fusion rate at different cellular compartments.
| Soluble N-ethylmaleimide–sensitive factor (NSF) attachment protein receptor (SNARE) proteins are thought to mediate vesicle fusion in all eukaryotes [1–4]. In nerve terminals, there are two target-SNAREs (t-SNAREs, also called Q-SNAREs), syntaxin 1A and synaptosome-associated protein-25 kDa (SNAP-25) on the plasma membrane, and one vesicle-associated SNARE (v-SNARE, also called R-SNARE), synaptobrevin 2 on synaptic vesicles [2]. Prior to exocytosis, the t- and v-SNAREs are thought to form a trans complex composed of a four-stranded helical bundle with one helix each from syntaxin and synaptobrevin and two helices contributed by SNAP-25 [5–9] (Figure 1A). As vesicles undergo fusion, the SNARE complex rearranges from a trans to a cis configuration such that all the SNARE proteins are localized to one membrane. The cis complex is then thought to be rapidly disrupted by the ATPase NSF [5,10–12], allowing the v-SNARE to be recycled into synaptic vesicles [13]. Although the specific mechanism of vesicle fusion is still in debate, it is now widely accepted that the formation of this four-helix bundle is essential for the fusion of the vesicle phospholipid bilayer with the plasma membrane phospholipid bilayer [3].
Vesicle fusion can be constitutive or triggered by calcium ion (Ca2+) [14]. In the latter case, the putative Ca2+ sensor synaptotagmin I plays a critical role [2,3]. Constitutive vesicle fusion differs from regulated secretion in that it is relatively less dependent on intracellular Ca2+. This has been demonstrated in reconstituted secretory cells [15] and at synapses, including mammalian [16,17] and invertebrate nerve terminals [18]. In these preparations, removal of extracellular Ca2+ or reduction of intraterminal [Ca2+] by Ca2+ chelators does not stop spontaneous vesicle fusion. At the Drosophila larval neuromuscular junction (NMJ), Ca2+-free saline containing ethylene glycol tetraacetic acid (EGTA) does not alter the rate of spontaneous release [19]. These observations collectively suggest that spontaneous vesicle fusion can occur even when intracellular [Ca2+] is reduced. This implies that a mechanism exists to overcome the energy barrier for vesicle fusion at low-Ca2+ conditions. Because SNARE complexes also mediate vesicle fusion along the constitutive secretory pathway [1,20], it is conceivable that this mechanism lies within the different structural and/or biochemical properties of SNARE complexes used for constitutive secretion and Ca2+-regulated exocytosis.
The synapse offers an ideal site to test this hypothesis because both forms of secretion co-exist and the SNARE proteins involved in the process are well studied. Furthermore, vesicle fusion can be readily detected at single-vesicle levels using electrophysiology [14]. In this study, we focused on a point mutation, T254I in syntaxin 1A, located at the +7 layer of the SNARE core complex [6], and its role in SNARE complex assembly and synaptic transmission in Drosophila. In an earlier study [21], it was demonstrated that this mutation (syx3–69) completely abolished the assembly of the SNARE complex at the restrictive temperature. Consequently, synaptic transmission was fully blocked and the fly paralyzed. Along with previous genetic deletion or mutation studies [22–24], these results provided important in vivo evidence that SNARE complex assembly was essential for synaptic vesicle fusion. However, our re-investigation of the syx3–69 mutant shows that the T254I mutation blocks neither the assembly of the SNARE complex nor synaptic transmission at the restrictive temperature. Instead, we find that the T254I mutation promotes the formation of the SNARE complex as well as vesicle fusion at permissive temperatures. These findings are consistent with a molecular model of the SNARE complex, suggesting that the T254I mutation causes a structural change of the +7 layer so that the mutant layer more closely resembles those found along constitutive secretory pathways. By enhancing the hydrophobic core of the molecule in the vicinity of layer +7, towards the C-terminal transmembrane helix, the mutant SNARE complex favors “constitutive-like” vesicle secretion either by increasing intermolecular interactions among the SNARE bundles or by stimulating vesicle docking and/or priming. These results suggest an evolutionarily conserved mechanism intrinsic to the structure of SNARE complexes that could act as a molecular switch to regulate the rate of vesicle fusion.
Syntaxin 1A is a critical component of the SNARE complex and is thought to be essential for synaptic vesicle fusion [1–3,24]. A previous study showed that mutation of threonine (T) to isoleucine (I) at position 254 in the Drosophila syntaxin 1A was sufficient to abolish the assembly of SNARE complexes at restrictive temperatures [21]. However, this conclusion is questionable if we take into consideration the conservation and divergence of residues at this position among different syntaxins. Our sequence analysis shows that, with the exception of syntaxin 4, most syntaxins found at the plasma membrane have a highly conserved T254 residue at the +7 layer (Figure 1B). Notably, the T254-containing syntaxins, such as syntaxin 1, 2, and 3, are typically used for regulated vesicle fusion at either synapses or neurosecretory cells in a diverse range of animal species [25–27]. In contrast, syntaxins involved in most constitutive secretion pathways in both animals and plants have one of the following amino acids at their equivalent positions: isoleucine, leucine (L), or valine (V) (Figure 1B; see also Figure S1). Valine, leucine, and isoleucine are similar in that they are hydrophobic, branch-chained amino acids. Therefore, this substitution of the residue at position 254 among syntaxins in the constitutive pathways is highly conserved throughout evolution. There are a few exceptions to this generalization. The yeast plasma membrane syntaxin orthologs have a threonine at the equivalent position (Figures 1B and S1). Furthermore, T254-containing syntaxins could also function in non-synaptic secretions, such as syntaxin 2 in postsynaptic membrane trafficking [26] and Drosophila syntaxin 1A in cuticle secretion [23,28]. Nonetheless, the overall feature emerging from our analysis is that syntaxins with conserved isoleucine at the +7 layer appear to be selectively involved in regulated secretion at synapses or neurosecretory cells.
It is particularly interesting to note that the T254I mutation found in the syx3–69 mutant approximates a reversion to a residue of wild-type syntaxins found in the constitutive secretory pathway. Notably, syntaxin 5 isoforms place an isoleucine at the site equivalent to position 254. Syntaxin 5 clearly functions in mammal cis Golgi networks at normal body temperature, similar to the temperature at which the syx3–69 mutant is reported to lose the ability to form SNARE complexes [21]. This prompted us to reconsider whether the T254I mutant syntaxin 1A indeed ceases to function at restrictive temperatures. To this end, we thoroughly re-examined the behavior, synaptic transmission, and SNARE complex formation of the syx3–69 mutant fly at elevated temperatures. Our tests showed that syx3–69 mutant flies were rapidly paralyzed at 38 °C and recovered within 3 min when returned to permissive temperature after a 20-min period of paralysis (Figure 2A and 2B). The paralysis and recovery rates were identical to those shown previously [21]. However, different from the previous observations, we noted that the syx3–69 mutant fly was paralyzed, but not motionless: the flies constantly shook their legs and abdomens during the period of paralysis at 38 °C (compare Video S1 with Video S2). We used a wild-type fly (Canton-S [CS]) and two other temperature-sensitive paralytic flies, Shibirets1 (Shits1) and paralyticts1 (parats1) as controls. As expected, Shits1 and parats1 flies were completely paralyzed due either to a block of synaptic vesicle recycling [29] or a failure of action potential propagation [30], respectively, and did not exhibit the shaking seen in the syx3–69 mutant. Upon returning to room temperature, Shits1, parats1, and syx3–69 flies all resumed their normal activities (Video S3).
These behavioral observations suggest that synaptic transmission persists in syx3–69 flies at the restrictive temperature. To further test this idea, we examined leg movement upon the activation of the giant fiber pathway in adult flies [31]. We stimulated the giant fiber neurons located in the head and observed the movement of fly legs (the body and wings were anchored with wax on a slide). Repetitive and phase-locked leg shaking was readily observed in syx3–69 flies at both the permissive temperature (20 °C; unpublished data) and restrictive temperature (38 °C) following each stimulus of the giant fiber neurons (see Video S4). In contrast, Shits1 flies moved their legs in response to each stimulus only at the permissive temperature (20 °C; unpublished data), but not at the restrictive temperature (Video S5). Figure 2C (rightmost panels) summarizes the spontaneous and electrical stimulation–evoked leg movement in syx3–69 flies and the lack of such movement in Shits1 flies at restrictive temperatures.
The persistence of synaptically evoked leg movements at the restrictive temperature suggests that synaptic transmission remains intact through multiple synapses (an electrical synapse and two chemical synapses) along the giant fiber pathway [31]. To directly measure synaptic transmission, we next recorded the synaptic response of the dorsal longitudinal indirect flight muscles (DLMs) from syx3–69 flies maintained at 38 °C. Our results show that evoked synaptic transmission and the resulting action potential persisted at 38 °C (n = 6; Figure 2D). During the course of these experiments, we noted that intracellular electrodes were often dislodged from DLMs only from syx3–69 flies, and there was a high incidence of spontaneous action potentials in the mutant DLMs (Figure 2D, inset). In contrast, Shits1 flies completely lost synaptic transmission upon activation of the giant fiber neuron at restrictive temperatures [32] (unpublished data). Hence, synaptic transmission is not blocked at restrictive temperatures in syx3–69 flies. As shown below, it is likely that paralysis of the syx3–69 mutant is caused by excessive or uncoordinated release of transmitter, rather than a complete block of exocytosis as suggested previously [21].
Consistent with the observation that synaptic transmission persists along the giant fiber pathway, light-induced “on” and “off” transient potentials of electroretinograms (ERGs) were not blocked by exposure of the syx3–69 fly to the restrictive temperature (Figure 3). These transients are thought to reflect synaptic transmission from photoreceptors to downstream interneurons in the retina [33]. The control fly, Shits1, lost its transient potentials at 33 °C, consistent with a depletion of the vesicle pool [21,29,32] (Figure 3B). However, the findings from the syx3–69 fly differ from those reported earlier [21], which showed that the restrictive temperature reversibly blocked these transients. In our experiments, we carefully monitored the temperature of the syx3–69 fly by placing a temperature probe adjacent to the experimental fly. Additionally, we mounted another syx3–69 fly beside the experimental fly so that we could observe the paralysis during the exposure at 38 °C and the recovery afterward. In a total of eight experiments, we never saw a loss of these transient potentials. In fact, our results showed that the “on” transient potential was slightly increased in amplitude at 38 °C (see Figure 3C). Additionally, we also observed spontaneous and light-induced high-frequency “bursting” activities typically indicative of enhanced neuronal activity in both the wild-type and the syx3–69 flies (see arrowheads in Figure 3A and 3C; see also [34]). Hyperactivity of the thoracic ganglion was also observed independently by Dr. Bruno van Swinderen's laboratory when syx3–69 flies were exposed to the restrictive temperature (B. van Swinderen, personal communication).
Taken together, both our behavioral tests and electrophysiological analyses support the notion that synaptic transmission is not blocked in syx3–69 mutants at restrictive temperatures. These results further suggest that the formation of the SNARE complex is not abolished in syx3–69 mutants at the restrictive temperature. To test this hypothesis, we measured the level of SNARE complexes using the methods described previously [11,21]. We first established the “linear” range that allows optimal detection of changes in the sodium dodecyl sulfate (SDS)-resistant SNARE complex (Figure S2) and then measured the level of SNARE complexes in syx3–69 mutants. Our results showed that the amount of the 7S SNARE complex and high molecular weight SNARE multimers (or oligomers) remained at wild-type levels in syx3–69 mutant flies at the restrictive temperature (Figure 4A and 4B). The syx3–69 mutant fly was exposed to 38 °C for 20 min prior to rapid freezing with liquid nitrogen and extraction of the SDS-resistant SNARE complex, as described previously [21]. In 50 separate experiments, we consistently observed the SNARE complex. This result has also been independently noted in Dr. Leo Pallanck's laboratory (L. Pallanck, personal communication). In a number of experiments, we also included the comatose (comt) mutant in our Western analysis and detected a consistent accumulation of the SNARE complex (Figure 4C), which is thought to be caused by dysfunction of NSF at restrictive temperatures [10–12]. Taken together, all our observations show that the T254I mutation in syntaxin 1A does not block SNARE complex formation nor does it block synaptic transmission at restrictive temperatures. Because our results differ markedly from those reported earlier [21], we sought to confirm whether the mutant fly we studied indeed carried the T254I mutation as shown in the syx3–69 mutant. Sequencing confirmed that there is a single base change from ACC to ATC in the open reading frame of syntaxin 1A (see Figure S3). Furthermore, we were able to rescue the paralysis (unpublished data) and electrophysiological defects by neuronal expression of the wild-type syntaxin 1A in the syx3–69 mutant background (see below). These results leave little doubt that the phenotype we study here is specifically caused by the T254I mutation in syx3–69 mutant flies.
To account for the hyperactivity observed in syx3–69 flies, we next examined whether the T254I mutation in syntaxin 1A has any effect on SNARE assembly and synaptic function at permissive temperatures. Upon examination of the available crystal structures of SNARE core complexes [6], we found that many of the central layers are tightly packed with hydrophobic residues contained within the four helical bundles. An example of this tight packing in the +1 layer of the synaptic SNARE core complex is illustrated in Figure 5A and 5B. The interactions of Leu57 and Ile178 from SNAP-25, Ile230 from syntaxin 1A, and Leu60 from synaptobrevin form square-planar geometry typical of the leucine zipper motif. In contrast, the +7 layer containing the wild-type syntaxin 1A is relatively loosely packed due to the presence of a conserved polar threonine residue at position 251 (equivalent to position 254 in Drosophila syntaxin 1A) [6,7,21], which packs against more hydrophobic partners. Results of examination of homologous neuronal SNARE syntaxin proteins implied a similar loosely packed configuration in this layer [7]. Interestingly, the homologous layer of the endosomal SNARE X-ray structure (1GL2) [20] shows more reliance on hydrophobic, branched-chain amino acids, than the synaptic SNARE (Figure 5C). The resulting interaction may contribute more hydrophobic stability of the zippered endosomal complex relative to the wild-type synaptic SNARE complex. We therefore propose that the tightened +7 layer in the SNARE complex containing the T254I mutant syntaxin 1A may mimic the function of the endosomal complex.
Our modeling results do not support the observation that the T254I mutation debilitates SNARE complex assembly, as previously reported [21]. To further verify this, we conducted molecular dynamics simulations of the SNARE complex in a water bath at 300 K for 5 ns using GROMACS [35,36]. After equilibration was achieved, there was no gross difference between the interactions of wild-type SNARE components and the mutant SNARE components. Also, the wild-type SNARE structure shows some degree of “fraying” at the termini of the complex [6]. Although this fraying effect is probably not physiologically relevant, per se, it does illustrate the looser packing of residues at the periphery of the wild-type complex. In our simulations, one would expect a destabilization of the termini (increase in fraying) if this mutation were indeed unstable; however, none was observed. Our simulation shows that the T254I mutation does not destabilize the complex, nor does it obviously increase the fraying at the terminus relative to wild-type.
Based on this structural analysis and modeling, we predict that the T254I mutation facilitates the formation or stability of the SNARE complex by enhancing intermolecular hydrophobic interactions among the four SNARE α-helices. Because this layer is near the C-terminal of the SNARE core complex, a tighter zippering of the SNARE complex may make fusion more probable by lowering the energy barrier for fusion and thereby partially abrogate the Ca2+-dependence of exocytosis. The mutant protein could also promote vesicle fusion by enhancing vesicle docking/priming. In other words, the T254I mutation may increase the rate of spontaneous release, turning the synapse into a constitutive secretion site. Alternatively, the T254I mutation could stabilize the cis SNARE complex such that it impedes vesicle recycling and ultimately reduces exocytosis upon repetitive nerve stimulation.
To test these structural predictions, we first investigated the biochemistry of SNARE complex assembly in the syx3–69 mutant at room temperature. Unlike the results obtained at the restrictive temperature (Figure 4), our measurements showed that the average amount of the SDS-resistant 7S SNARE complex was significantly increased in the syx3–69 mutant compared to that in the wild type (CS) at 22 °C (n = 50, p < 0.05) (Figure 6A and 6B). Similarly, the level of SNARE multimers was also significantly increased in the mutant (n = 9, p < 0.05). These results show that the level of SNARE complexes is increased in the syx3–69 mutant. Concerned that the SDS-resistant SNARE complex is unique to neuronal SNAREs [37,38], we next used an alternative method to immunoprecipitate the SNARE complex from fly-head extracts using a polyclonal antibody against one of the SNARE components, SNAP-25 [39,40]. Our results showed that the SNAP-25 antibody readily and specifically precipitated syntaxin 1A and synaptobrevin, but not tubulin (Figure 6C). When normalized to the amount of proteins precipitated from the head extracts in the wild-type flies, the levels of syntaxin 1A, synaptobrevin, and SNAP-25 were all increased slightly (Figure 6D; n = 4). Even though these changes are not statistically significant, the trend is consistent with those observed for the SDS-resistant complexes.
The level of the SDS-resistant SNARE complex has been shown to correlate well with the level of exocytosis [39,41,42]. We next tested whether this increase in the rate of SNARE complex assembly had any physiological effects on synaptic vesicle fusion. We recorded action potential–independent and constitutive (or spontaneous) miniature excitatory postsynaptic potentials (mEPSPs or minis) from third instar larval body-wall muscles innervated by motoneurons [43,44]. These mEPSPs are caused by constitutive secretion of glutamate from the nerve terminal [43]. Surprisingly, we found that the frequency of constitutive release was dramatically increased some 7-fold in the mutant (n = 9) compared to the wild type (n = 8; p < 0.001) (Figure 7A and 7B). The average mini amplitude was similar in both the syx3–69 mutant (n = 11) and the wild-type larvae (n = 8; p > 0.1) (Figure 7C), suggesting that quanta and postsynaptic receptors likely remain normal. Immunocytochemical studies of glutamate receptors failed to show detectable differences between the mutant and the wild type (unpublished data). This mini recording was conducted in saline containing 0.8 mM Ca2+ and 1 μm TTX, which was also used for evoked synaptic potentials (below). The resting potential was not different between these two genotypes (−69.7 ± 1.2 mV, n = 8, for the wild type, and −69.4 ± 0.9 mV, n = 9, for the mutant; p > 0.5). In these and all other larval recordings shown in this study, the muscle input resistance (between 5–9 MΩ) did not differ between the wild-type and the mutant larvae.
To test whether this increase in mini frequency depends on extracellular Ca2+, we recorded minis in a Ca2+-free saline. The unusually high rate of spontaneous release remained in the syx3–69 mutant in the absence of extracellular Ca2+ (n = 8), but significantly higher than that in the wild type (n = 8; p < 0.001) (Figure 7B). The resting potential was not different (−69.75 ± 1.18 mV, n = 8, for the wild type; −69.75 ± 0.85 mV, n = 8, for the mutant; p > 0.5). The lack of effects by Ca2+ removal on mini frequency is consistent with an earlier report showing that Ca2+-free saline plus EGTA did not alter mini frequency at the Drosophila larval NMJ [19]. Furthermore, mini frequency remained 13-fold higher in syx3–69 mutants compared to the wild type in Ca2+-free saline containing the membrane-permeable Ca2+ chelator EGTA-AM (n = 4). These results indicate that the T254I mutant syntaxin 1A couples the formation of SNARE complexes with constitutive vesicle fusion even when the intracellular [Ca2+] is greatly reduced.
An increase in SNARE complex assembly could enhance Ca2+-evoked exocytosis. On the other hand, the dramatic increase in the rate of constitutive vesicle fusion could deplete the vesicle pool and reduce Ca2+-evoked release. To distinguish these possibilities, we recorded action potential–evoked excitatory postsynaptic potentials (EPSPs) from muscles bathed in 0.8 mM Ca2+ saline. We observed that the amplitude of evoked EPSPs was also significantly increased to 37 mV (n = 11) in the syx3–69 mutant from 25 mV (n = 9) in the wild type (p < 0.005; Figure 7D and 7E). Because the average mini amplitude was not significantly different between the mutant and the wild type (p > 0.1), this increase in EPSP amplitude most likely reflected an enhancement in presynaptic release. Factoring in the respective average mini amplitude in these flies and after correction of EPSP amplitudes for nonlinear summation [45,46], there was a 2-fold increase in quantal content from 40.0 (n = 8) in the wild type to 79.9 (n = 11) in the mutant (Figure 7F; p < 0.001). As with the mini measurement, there was no difference in resting potentials of the muscle fiber between the wild type and the mutant. These results indicate that the average number of synaptic vesicles undergoing exocytosis induced by an action potential is significantly increased in the syx3–69 mutant.
Another possibility predicted by our structural modeling is that the T254I mutation may slow vesicle recycling by stabilizing post-fusion cis SNARE complexes. To test this hypothesis, we repetitively stimulated the motor nerve at 10 Hz for a prolonged period (5 min). We adjusted the extracellular [Ca2+] such that the initial EPSP amplitude was similar between the wild-type control (at 1.5 mM Ca2+) and the syx3–69 mutant (at 1 mM Ca2+) (Figure 8A and 8B). At these [Ca2+], the resting potential was −76.5 ± 2.6 mV (n = 4) and −75.6 ± 1.6 mV (n = 6) for the wild type and the mutant, respectively. The basal release was 52.8 ± 1.0 mV (n = 4) and 49.5 ± 1.6 mV (n = 6) for the wild type and the syx3–69 mutant (p > 0.1), respectively. As previously shown, there was an initial, rapid decline in the amplitude of EPSPs after the onset of the moderate stimulation at 10 Hz [47]. The EPSP amplitude then reached a steady-state level approximately 60%–65% of single-pulse–induced EPSPs (Figure 8C). Under such stimulation conditions, the steady-state release level is thought to reflect the balance between vesicle recycling and exocytosis [47]. There were no statistical differences in the rate of EPSP decline or the steady-state levels between the wild-type and the syx3–69 mutant larvae (p > 0.05). EPSPs recovered at a similar rate after the 5-min stimulation (Figure 8C). These results suggest that the T254I mutation does not have a detectable effect on synaptic vesicle recycling.
At Drosophila NMJs, the readily releasable pool (RRP) of synaptic vesicles is estimated to be 230, which can be rapidly depleted within a few stimuli [47]. We examined the RRP by measuring the relative amplitude of the first ten EPSPs after the onset of the 10-Hz stimulation. Our results showed no significant difference between the wild-type (n = 4) and the syx3–69 mutant flies (n = 5; Figure 8D). Thus, the RRP of synaptic vesicles is not reduced in the syx3–69 mutant despite the extraordinarily high rate of spontaneous fusion rate. Taking into consideration the high rate of spontaneous vesicle fusion, it is reasonable to assume that vesicle docking is, in effect, increased in syx3–69 mutants.
Oligomerization of the 7S SNARE complex into high molecular weight complexes is proposed to be essential for vesicle fusion [39]. This implies that multiple SNARE complexes are required to promote vesicle fusion. The precise number of SNARE complexes required for vesicle fusion is unknown, but is estimated to be between three and 15 pairs [3,48]. Hence, one could envision a scenario in which the I254 mutant syntaxin 1A exerts a dominant positive effect on synaptic vesicle fusion in heterozygous mutant flies (i.e., flies that also have one copy of the wild-type syntaxin 1A) by acting as part of the multimeric SNARE complex (Figure 9A). To test this hypothesis, we generated heterozygous syx3–69/+ larvae. The resting potential of the muscle fiber in syx3–69/+ larvae was −70.9 mV (n = 7) at 0.8 mM Ca2+, which is not significantly different from those in the wild type (+/+; −69.8 mV) and the syx3–69/syx3–69 homozygous mutant (−69.4 mV) under the same [Ca2+] (p > 0.3). The frequency of spontaneous fusion (6.4 Hz; n = 7) was significantly higher than that in the wild type (2.65 Hz, p < 0.001), but much lower than that in the homozygote (19.67 Hz, p < 0.001). This observation is consistent with the working model that I254-containing syntaxin 1A has a dominant positive effect on vesicle fusion.
We next recorded evoked release in heterozygotes and showed that the amplitude of evoked EPSPs (39 mV; n = 9) was similar to that in the homozygote (37.3 mV; p > 1), but significantly higher than that in the wild type (25.3 mV, p < 0.001) (Figure 9B–9E). These results indicate that the T254I mutant syntaxin 1A also has a dominant positive effect on Ca2+-triggered vesicle fusion. However, it is clear that dilution of the mutant SNARE complex by the wild-type syntaxin 1A does not reduce evoked release, as it does to constitutive secretion. This observation is inconsistent with the possibility that the T254I mutant syntaxin 1A enhances Ca2+ influx, as one would expect a greater increase in evoked release in the homozygote. A likely explanation we suggest is that the T254I mutant syntaxin 1A stimulates the formation of SNARE complexes in a dominant fashion. For a given level of SNARE complexes, the energy barrier for fusion correlates negatively with the amount of the T254I mutant syntaxin 1A. Although this energy barrier is increased for constitutive fusion in the heterozygote compared to the homozygote, this barrier should be overcome easily by the rise of intraterminal Ca2+. Our measurement of the SDS-resistant complex confirmed that the amount of SNARE complex was similar between homozygotes and heterozygotes (Figure S4).
This model further predicts that the increase in evoked release should occur in both homozygotes and heterozygotes at both low and high [Ca2+]. To this end, we recorded EPSPs at two additional Ca2+ concentrations (1 mM and 0.4 mM). These Ca2+ concentrations did not alter resting potentials (unpublished data), but they did affect transmitter release (Figure 9F). At 1 mM Ca2+, the average EPSP amplitude was similar in the heterozygote (syx/+, 48.6 mV, n = 7) and homozygote (syx/syx, 49.5 mV, n = 6), but was consistently larger than the wild type (+/+, 42.7 mV, n = 8; p < 0.01). At 0.4 mM Ca2+, the amplitude of EPSPs in the wild type was quite small (3.5 mV, n = 9). In comparison, the amplitude of EPSPs was significantly larger in both heterozygotes (13.5 mV, n = 8) and homozygotes (15.4 mV, n = 9) of syx3–69 (p < 0.001). The average amplitude of EPSPs was then compared with those seen at 0.8 mM [Ca2+] (Figure 9F). The relatively smaller increase of EPSP amplitude at increasingly higher [Ca2+] reflects the ceiling effect due to non-linear summation. Nonetheless, these results show that evoked release is dramatically enhanced in I254-containing flies at a wide spectrum of extracellular [Ca2+].
The possibility remains that the dominant positive effects we have seen in the heterozygote could result from a second site mutation elsewhere rather than the T254I mutation in the syntaxin locus. To address this concern, we generated transheterozygous flies (syx3–69/syxΔ229) in which the syx3–69 mutant chromosome was placed in trans to a null syntaxin mutation (syxΔ229) [23]. In syx3–69/syxΔ229 mutants, the mini frequency was 22.3 Hz (n = 9), which is significantly higher than that in the wild-type larvae (3.2 Hz, n = 8; p < 0.0001) (Figure 10A). At 0.8 mM Ca2+, the evoked EPSP amplitude was also significantly increased to 37.9 mV (n = 9) from 29.5 mV (n = 8) in the wild-type larvae (p < 0.01) (Figure 10B). The resting potential of the mutant animal (−72.4 mV) was similar to that (−73.8 mV) in the wild-type larvae. These results are highly similar to those found in the syx3–69/syx3–69 homozygote. Along with the molecular evidence presented earlier, these results provide further genetic and electrophysiological evidence that the effects we have observed in the syx3–69 mutant is specifically caused by the T254I mutation in the syntaxin gene.
To further demonstrate indeed the “neutralizing” effect on the T254I mutant syntaxin 1A is mediated by the wild-type syntaxin 1A in the heterozygote, we then performed a genetic rescue experiment in which we selectively expressed the wild-type syntaxin 1A gene in postmitotic neurons using the Gal4-UAS binary system [49,50]. When the wild-type syntaxin 1A gene was expressed in the wild-type background (C155 Gal4/+; UAS-Syx 1A/+), it had no significant effect on either the constitutive secretion rate or the amplitude of evoked EPSPs compared to those in the wild-type larvae carrying the pan neuronal Gal4 driver (C155 Gal4, n = 7; p > 0.05) (Figure 10C and 10D). However, neuronal overexpression of the wild-type syntaxin 1A in the syx3–69/null mutant background (i.e., C155 Gal4/+; UAS-Syx 1A/+; syx3–69/syxΔ229 [23]) resulted in a dramatic reduction in the frequency of constitutive secretion to 8.8 Hz (n = 14) from 24.9 Hz (n = 10) in C155 Gal4/+; syx3–69/syxΔ229 mutant larvae (p < 0.001). This rate is significantly higher than that in the C155 Gal4 larvae (5.3 Hz, n = 7) or overexpression alone (C155 Gal4/+; UAS-Syx 1A/+, 4.9 Hz, n = 9) (p < 0.05). Overexpression of the wild-type syntaxin 1A in the syx3–69/null mutant background also significantly reduced the average amplitude of EPSPs from 49.8 mV (n = 10) in C155 Gal4/+; syx3–69/syxΔ229 mutant larvae to 39.3 mV (n = 14). This EPSP amplitude is similar to that in the C155 Gal4/+; UAS-Syx 1A/+ background (38.0 mV, n = 9), but significantly higher compared to that in the C155 Gal4 larvae (32.8 mV, n = 7). These results demonstrate that neuronal expression of the wild-type syntaxin 1A rescues the mutant phenotype by specifically neutralizing the dominant positive effects on both constitutive and evoked secretion induced by the T254I mutant syntaxin 1A.
This study reports the behavioral, electrophysiological, biochemical, genetic, structural, and molecular results from a re-investigation of the syx3–69 mutant in Drosophila. These findings contradict an earlier report [21] on both the experimental evidence and conclusions concerning the effects of the T254I mutation in syntaxin 1A on synaptic transmission. Multiple lines of evidence demonstrate that the T254I mutation in the syx3–69 mutant fly blocks neither synaptic transmission nor SNARE complex assembly at restrictive temperatures. More importantly, we have gone steps further by revealing an evolutionarily conserved structural feature among syntaxin orthologs in regulating both constitutive secretion and Ca2+-regulated exocytosis.
One of the major new findings from this study is that the T254I mutant syntaxin 1A in the syx3–69 mutant dramatically stimulates vesicle fusion. At the restrictive temperature, the syx3–69 flies exhibit uncontrolled hyperactivities and enhanced neuronal firing. At the permissive temperature, SNARE complex assembly is moderately enhanced, whereas the rate of constitutive vesicle fusion is dramatically increased in the syx3–69 mutant. Importantly, this enhancement of constitutive secretion persists in Ca2+-free saline and when intracellular Ca2+ is further reduced by chelation. This implies that spontaneous vesicle fusion is less dependent upon Ca2+, a conclusion consistent with those reported in a number of synapses [16–18], including the Drosophila NMJ [19]. Although we do not suggest that vesicle fusion is absolutely independent of intracellular Ca2+, our studies support the notion that the T254I mutation makes vesicle fusion more efficient, regardless of whether it is constitutive or Ca2+-regulated fusion.
Another major finding is that despite the high rate of constitutive fusion, the vesicle pool is not depleted, implying that vesicle docking or priming is enhanced in the syx3–69 mutant via a yet unidentified mechanism. Consistent with the increase in mini frequency, evoked transmitter release is significantly increased in the syx3–69 mutant. Thus, the T254I mutation stimulates both constitutive and evoked vesicle fusion. This increase in evoked transmitter release correlates well with the enhanced assembly of SNARE complexes in the mutant fly.
The third interesting finding is that the T254I mutant syntaxin 1A exerts a dominant positive effect on vesicle fusion. In heterozygous syx3–69 mutant (syx3–69/+), the rate of spontaneous fusion is slightly higher than that in the wild-type larvae, whereas evoked release remains at the homozygote level. Two lines of genetic and electrophysiological evidence suggest that this dominant positive effect is specifically associated with the T254I mutation in the syntaxin locus. First, the dominant positive effect persists in larvae carrying only one copy of the T254I mutation in the null mutant background (i.e., syx3–69/syxΔ229). Second, neuronal overexpression of the wild-type syntaxin 1A effectively rescues the effect of the T254I mutant in the syx3–69/null mutant background. It is important to note that neuronal overexpression of the wild-type syntaxin 1A protein in the wild-type background has little effect on both spontaneous and evoked vesicle fusion. Therefore, the counterbalance exerted by the wild-type syntaxin 1A is specific to the T254I mutant syntaxin 1A. Taken together, these observations lend further support to the notion that the I254 mutant syntaxin 1A is more efficacious than the wild-type T254 syntaxin 1A in promoting vesicle fusion.
How might the T254I mutation exert such a dramatic effect on vesicle fusion? The precise mechanism is unknown; however, we believe the effect of the mutant protein can be better explained by examining the structural impact of the point mutation on the SNARE complex. The formation of the SNARE complex is generally accepted as an essential step in vesicle fusion. This conclusion is supported by considerable evidence accumulated over the last decade using a variety of experimental methods, including the use of specific neurotoxins to cleave SNARE proteins, and genetic mutations or deletion of SNARE genes [1–3]. Based on structural and functional studies of the core complex [6–8], it has been postulated that the assembly of the SNARE complex involves a “zippering” process in which complex formation starts at the N-termini of the four helices, followed by zippering of the core “layer” of the SNARE bundle towards the C-terminal bundles. The process of zippering is also believed to provide the energy necessary to bring the vesicle close to the plasma membrane [2,3,9]. To date, most of the data supporting this zipper model came from observations of “loose” and “tight” states of SNARE complexes in neuroendocrine cells [51], at crayfish neuromuscular synapses [52], and in liposome fusion [53]. It is also indirectly supported by genetic mutations of the helical region of SNAREs (see discussions in [7]) and by the ability of inhibitory peptides of the helical region of SNAREs to block both core complex assembly in vitro and transmitter secretion in PC12 cells [54,55]. At present, both the precise mode of SNARE complex formation [56] and the role of the complex in vesicle fusion [3,57] are not fully resolved. Nonetheless, the zipper model serves as a good starting point for experimental testing of SNARE structure and function.
The T254I mutation is located at a strategic location near the end of the zipper, a presumed final step before vesicle fusion takes place. We have made three interesting observations of the +7 layer by sequence and structural comparison. First, with a few exceptions, nearly all syntaxins involved predominantly in regulated vesicle exocytosis at synapses or neurosecretory cells have a common hydrophilic residue, threonine, at position 254 in the +7 layer. In contrast, most syntaxins acting in the constitutive secretory pathways have one of the highly conserved hydrophobic residues (I, L, or V). Second, there is a conserved switch in the +7 layer packing among SNARE complexes used in different secretory pathways. This layer is loosely packed in “synaptic” SNAREs, but tightly bundled together in “constitutive” SNAREs, where hydrophobic residues (I, L, or V) may enhance direct intermolecular interactions among the four α-helices. Third, our structural modeling suggests that the T254I mutant +7 layer is more tightly packed than is the wild type, and that it resembles more the tight packing found in the endosomal SNARE core complex [20].
Based on our sequence and structural analyses, we favor the idea that a structural alteration of the +7 layer induced by the T254I mutation in syntaxin 1A may best account for our experimental observations. The extraordinarily high rate of spontaneous fusion detected in the syx3–69 mutant appears to support the “zipper model” or a modified zipper model [56], suggesting that tightening the SNARE complex does promote vesicle fusion. That overexpression of T254 syntaxin 1A specifically counteracts I254 mutant syntaxin 1A in vesicle fusion implies that the relatively loose packing of the +7 layer containing the wild-type syntaxin 1A may serve as an internal brake to dampen vesicle fusion. Once this brake is removed by the T254I substitution, the mutant SNARE complex lowers the energy barrier for vesicle fusion beyond a point of no return in a manner that is relatively less dependent on intracellular Ca2+ [15–19]. This working model also explains why evoked release is enhanced in both homozygotes and heterozygotes.
We should stress that our results do not permit us to conclude whether or not SNARE complexes directly mediate vesicle fusion. Although pairing of SNARE proteins has been shown to mediate liposomal vesicle fusion in vitro [57], the rate of liposomal fusion is slow. More importantly, new evidence suggests that SNARE proteins alone may bring the membranes in close apposition, but do not drive vesicle fusion under more physiological conditions [58]. The failure to mediate fusion in vitro suggests that other factors may either assist SNARE function or directly mediate vesicle fusion in vivo. Consistent with this idea, the vesicular ATPase (Vo) and synaptotagmin I have been reported to act either downstream of, or synergistically with, the SNAREs in vesicle fusion [50,59]. It is interesting to note that a G50E mutation in the N-terminal domain of SNAP-25 has previously been found to enhance both constitutive secretion and Ca2+-evoked release in Drosophila at the permissive temperature [60]. Unlike the T254I mutation, this G50E (G43E in mammals) mutation is thought to cause a conformation change of SNAP-25 such that the mutant SNARE complex is more ready to mediate vesicle fusion. The precise mechanism by which the G50E mutation promotes vesicle fusion remains to be resolved. Nonetheless, the T254I and G50E mutations offer two alternative structural changes to promote SNARE-mediated vesicle fusion. It is evident that much is still to be discovered about SNARE structure and function. The results presented here reveal a novel intrinsic mechanism by which SNARE-mediated vesicle fusion is regulated. These findings not only advance the understanding of synaptic transmission, but also have broad implications on vesicle fusion at different cellular pathways.
The syx3–69 mutant fly [21] was obtained from the laboratories of Drs. Troy Littleton (Massachusetts Institute of Technology), Barry Ganetzky (University of Wisconsin-Madison), and Leo Pallanck (University of Washington). This mutant line was maintained on a balancer chromosome (TM6B) and out-crossed to prevent potential accumulation of modifiers. The syntaxin null allele (syxΔ229 [23]) and UAS-Syntaxin [50] flies were obtained from Dr. Hugo Bellen's laboratory. The wild-type Canton S (CS or +/+) strain, Shibirets1 (Shits1), and paralyticts1 (parats1) originally obtained from the Bloomington Drosophila Stock Center were maintained in B. Z.'s lab. Flies were cultured on standard fly medium at room temperatures (∼20–22 °C). Unless otherwise specified, 3- to 5-d-old flies of both sexes were used in the adult experiments described.
The modeling of the T251I mutation (equivalent to the Drosophila T254I mutation) in syntaxin 1A was accomplished using PyMol [61]. The appropriate residue was modified (mutated) in the SNARE complex (1SFC, chain B). Ray tracing for Figure 5C was also performed with PyMol.
All dynamics calculations were carried out with GROMACS v3.3, in which the Gromacs 96 force field was used throughout [35,36]. The X-ray structure of the SNARE complex (chains A[synaptobrevin], B[syntaxin], C[SNAP-25], and D[SNAP-25] of 1SFC) was used as the starting model. Ordered water molecules and ordered divalent ions were excluded from this calculation. Hydrogen atom positions were calculated with the pdb2gmx program provided by GROMACS, resulting in 3,001 atoms (excluding SPC water molecules) in the native structure. Sixteen sodium ions were selected by the genion program included with GROMACS to negate the net negative charge of the SNARE bundle. A rectangular box of water (the SPC water model) extending 20 Å in every direction from the boundary of the protein component was calculated by the editconf program included with GROMACS [35,36]. Initially, the protein structure was minimized until convergence. Position-restrained molecular dynamics was used to equilibrate (simulation time of 20 ps) the water molecules with the protein. After energy minimization of the entire system was completed (protein + solvent + counter-ions), 5 ns of molecular dynamics trajectories were computed at 300 K. Once the system was equilibrated, a representative model was extracted from the trajectory (at 2 ns) and examined (Figure 5).
The methods described by Tolar and Pallanck (1998) [11] and by Littleton et al. (1998) [21] were closely followed for fly treatments and for the extraction of SDS-resistant SNARE complexes. Briefly, adult syx3–69 and CS flies were exposed to 38 °C for 20 min or kept at 22 °C, and then rapidly frozen in liquid nitrogen. Heads were separated from the body by brief vortexing and approximately 20 heads were collected on a sheet of paper under a constant superfusion with liquid nitrogen. Taking care to avoid introduction of air bubbles, fly heads were ground gently in 100-μl SDS sample buffer with a plastic pestle in an Eppendorf tube followed by centrifugation. The supernatant was collected, diluted to a final concentration at 0.25–0.5 heads/10 μl in sample buffer, and loaded onto gels at 10–15 μl per well. Samples were run on a discontinuous SDS-polyacrylamide gel: a 4% stacking gel, an upper 7.5%–8% resolving gel, and a lower 18% resolving gel to minimize excessive transfer of the monomer [11]. Commercial 4%–18% gradient gels were also used in one fourth of the experiments. Proteins were transferred to nitrocellulose membranes by running at 30 V overnight in a 4 °C room, as outlined by the manufacturer's instructions (Bio-Rad, Hercules, California, United States). Membranes were probed with a monoclonal antibody to syntaxin 1A (8C3; 1:100) [21]. Bands representing monomeric syntaxin 1A and the 7S complex were detected by enhanced chemiluminescence (ECL; GE Healthcare, Amersham Biosciences, Piscataway, New Jersey, United States) and quantified with ImageJ (National Institutes of Health [NIH], http://rsb.info.nih.gov/ij/). The SNARE complex level is expressed as a 7S complex to monomer (syntaxin) ratio and normalized to that in control (CS) flies at room temperature. To obtain oligomeric or multimeric complexes of the 7S complex [39], we either prolonged the exposure time of the film or loaded up to one head equivalent volume onto the gel. The ratio of the multimeric complex to the syntaxin 1A monomer was determined similarly to that for the 7S complex [11]. In most experiments, the blot was re-probed for tubulin to ensure that the protein loading level was similar in each lane.
Twenty-one adult fly heads were collected from wild-type (CS) and syx3–69 mutant flies (4- to 5-d-old) and ground on ice in 100-μl immunoprecipitation (IP) buffer containing 150 mM NaCl and 20 mM Tris (pH 7.5), and mini complete protease inhibitors (which were added at one tablet/10-ml IP buffer; Roche, Basel, Switzerland). The fly-head extract was mixed well with 4-μl 25% Triton X-100 (BioRad, Hercules, California, United States) and incubated for 15 min on ice. After a brief and gentle spin to remove cuticle debris, 20 μl of the supernatant was saved as “input” for control loading. The remaining supernatant was incubated with 10-μl sera against the Drosophila SNAP-25 (rabbit, N-terminal [40], and mixed with 20-μl prewashed CL-4B protein A Sepharose beads (GE Healthcare, Amersham Biosciences) by gentle rotation for 1–3 h at 4 °C. After removing the supernatant, the beads were washed four times with the IP buffer. The immunoprecipitates were eluted by boiling the beads in 50-μl sample buffer. The precipitates along with “input” were resolved on standard SDS gel and subject to Western blot analysis. SDS-resistant complexes were prepared separately and included on the gel as additional controls. The blot was sequentially probed with the 8C3 syntaxin 1A monoclonal antibody, a neuronal synaptobrevin (N-Syb) polyclonal antibody (guinea pig [62]), and a different SNAP-25 antibody (which also recognizes the close homolog SNAP-24 [40]). To ensure the specificity of immunoprecipitation, the blot was also probed with an antibody to α-tubulin (Sigma, St. Louis, Missouri, United States). The ECL method was used for protein detection. The band intensity was quantified with ImageJ (NIH).
The standard method of third instar larval electrophysiology described by Jan and Jan (1976) [43] and by Stewart et al. (1994) [63] was used to record spontaneous mEPSPs and action potential–evoked EPSPs in current clamp mode. The saline [Ca2+] was 0 or 0.8 mM for mEPSP recordings and 0.4, 0.8, 1, or 1.5 mM for EPSP recordings (specified in the text). A total of 1 μM tetrodotoxin was added to the saline to block action potentials when minis were recorded [44,64]. Because of the unusually high rate of spontaneous minis in the syx3–69 mutant larvae, many minis were clustered together or on top of one another. This made it difficult to ascertain the amplitude of individual minis. In this case, minis were analyzed following an EPSP after the membrane potential had returned to pre-stimulation levels. Quantal content was determined as the ratio of the average EPSP amplitude and the average mini amplitude after correction of EPSP amplitude for nonlinear summation following the methods described by Stevens (1976) [45] and Feeney et al. (1998) [46]. Corrected EPSP amplitude = E{Ln[E/(E − recorded EPSP)]}, where E = difference between reversal potential and resting potential. The reversal potential used in this correction was 0 mV [46]. For simplicity, the average amplitude of EPSPs presented in Figures 9F, 10B, and 10D was not corrected for nonlinear summation. The room temperature was 19–20 °C.
Mini recordings were also performed in larvae treated with EGTA-AM (Invitrogen, Molecular Probes, Carlsbad, California, United States). The final concentration of EGTA-AM was 10 μm in Ca2+-free HL-3 saline, diluted from a 40 mM stock in 20% Fluronic F-127 (a low-toxicity dispersing agent) in dimethyl sulfoxide (DMSO). The preparation was incubated in the EGTA-AM saline at room temperature for 30 min and washed with Ca2+-free saline prior to recording. The effect of Ca2+ chelation was monitored at the end of the mini recording. Upon switching back to 0.8 mM Ca2+ saline, evoked release was dramatically reduced (unpublished data), suggesting most, if not all, of the large number of Ca2+ ions evoked by an action potential were chelated by the intraterminal EGTA. Control experiments with Ca2+-free saline containing the same final concentration (0.025%) of the Fluronic F-127/DMSO solvent were also conducted.
For the vesicle depletion assay, we used relatively higher concentrations of Ca2+ (1.5 mM for the wild type and 1 mM for the syx3–69 mutant) to ensure achieving a rapid decline of EPSP amplitudes. Basal release was monitored at 0.2 Hz prior to a 5-min repetitive stimulation at 10 Hz. Usually, the recovery from the 10-Hz stimulation was also monitored by 0.2-Hz stimulation immediately afterwards. The basal amplitude of EPSPs was averaged from at least three consecutive EPSPs and used to normalize the average EPSP amplitude during the 10-Hz stimulation. At the onset of repetitive stimulation, three consecutive EPSPs were used to give the average amplitude at time 0, at 30 sec, and at every minute during the remaining the 10-Hz stimulation periods. The normalized plot, as shown in Figure 8C, allows one to estimate the decline rate and vesicle pools. The first ten EPSPs were also analyzed to estimate the depletion rate of the RRP.
The method of Tanouye and Wyman (1980) [31] was adapted for stimulating giant fiber neurons and for recording synaptic potentials and action potentials in DLMs. Flies were mounted on a slide with dental wax. Sharp glass microelectrodes (25 MΩ, filled with 3 M KCl) were used to record intracellularly from DLMs, whereas the giant fiber neurons were stimulated with a sharp tungsten electrode placed either inside the compound eye or in the cervical connective (1–6 V, 120-μs duration). A homemade temperature stage was used to rapidly (within 45 s) increase the recording chamber to 38 °C. The temperature probe was placed in dental wax next to the mounted fly to ensure the accuracy of the set-point temperature. The fly chamber was then rapidly cooled to room temperature by pumping ice-cold water through the metal stage. Wild type (CS), syx3–69, and Shits1 were used for this set of experiments. The room temperature was 19–20 °C.
For ERG recordings, 2- to 3-d-old flies (CS, Shits1, or syx3–69) were mounted on a slide with modeling clay and placed on a temperature-controlled stage. A sharp tungsten electrode was inserted gently in the thorax or abdomen of the fly and served as a reference electrode. A sharp glass microelectrode was inserted just through the cuticle into the compound eye. The fly was then allowed to adapt to the dark for a few minutes. ERGs were evoked by rapidly exposing the eye to white light for a brief duration (1–2 s). The fly chamber temperature was raised to 38 °C using a homemade temperature controller. To ensure the accuracy of the temperature experienced by the fly, the monitor probe was placed as close as possible to the experimental fly. Another experimental fly was mounted beside the fly being recorded so that it could be monitored for paralysis during 38 °C and recovery after the temperature was returned to the permissive temperature.
Flies were incubated at 38 °C in water-warmed glass vials for 10 min and then transferred to a sheet of paper placed on the lab bench for recovery at room temperature (∼22 °C). The time and number of flies capable of standing were recorded and plotted. A total of 230 syx3–69 flies were tested in 17 different trials (10–15 flies per trial). Wild-type (CS) flies were used as controls (+/+) in a few trials, and they were not paralyzed at this temperature.
Flies (syx3–69 and Shits1, 1- to 2-d-old) were mounted with their ventral sides up on a slide with modeling clay and viewed with a dissection microscope at 100×. Flies were placed in a chamber whose temperature was rapidly raised to 38 °C within 45 s by a homemade temperature controller and rapidly cooled to 20 °C (within 1 min) by circulating ice-cold water around the chamber. A pair of sharp tungsten electrodes was placed into the compound eyes to electrically stimulate the giant fiber neurons in the brain (1–6 V, 100-μs duration, 5 Hz). Spontaneous and evoked leg movements of these flies were recorded using a digital camera. Videos S4 and S5 are on syx3–69 and Shits1, respectively. Still clips from these videos are presented in Figure 2C.
Results are presented as mean ± standard error of the mean (SEM). The paired Student t-test was used to analyze the level of SNARE complexes, whereas the unpaired t-test was used to treat the electrophysiological results. In all cases, differences of p < 0.05 were considered statistically significant.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.